Optimization and Evaluation of Ensemble Learning Models for Intelligent Lithology Identification Using Seismic Data

Optimization and Evaluation of Ensemble Learning Models for Intelligent Lithology Identification Using Seismic Data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Optimization and Evaluation of Ensemble Learning Models for Intelligent Lithology Identification Using Seismic Data Wang Jingyi, Jiang Li, Zhibing Feng, Huang Xiao, Yao Zhenan, Zhang Bocheng This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6795780/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 10 Dec, 2025 Read the published version in Acta Geophysica → Version 1 posted 6 You are reading this latest preprint version Abstract Lithology identification is one of the key geological interpretation tasks in oil and gas exploration, which directly affects the accuracy of reservoir modeling and resource evaluation. At present, seismic lithology prediction research has the limitation of over-focusing on the optimization of a single algorithm and lacking systematic comparison of ensemble models and hyperparameter strategies. To this end, this study employs recursive feature elimination, NM-SMOTE sampling, and four hyperparameter optimization methods. These are applied to well seismic data from the F3 exploration area in the North Sea to evaluate random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), categorical boosting (CatBoost) and stacked ensemble models (SEM).The experimental results show that in terms of hyperparameter optimization, the Optuna algorithm achieves the best balance between computational efficiency and model performance, and its optimization effect is significantly better than that of the traditional grid search method. In the context of single models, CatBoost shows the best prediction performance (AUC = 0.91), with clear boundaries for sandstone and mudstone identification and the best spatial continuity of the prediction results. The comparative analysis of different ensemble models shows that random forest has the highest stability, followed by LightGBM, while XGBoost is more sensitive to data noise, resulting in a instability in the prediction results. It is worth noting that the classification performance of the SEM is limited under complex geological conditions such as thin interbeds. This study systematically evaluates the technical characteristics of each model and proposes model selection criteria for different geological application scenarios, providing important theoretical basis and method support for the practical application of intelligent lithology identification technology. Lithology Identification Seismic Data Ensemble Learning Hyperparameter Optimization Comparative Analysis Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 Figure 18 1 Introduction With global hydrocarbon exploration targets gradually shifting towards complex and concealed reservoirs, lithology intelligent identification technology based on seismic and well logging data has become increasingly critical for reservoir characterization (Ali et al. 2025 ; Cheng et al. 2023 ). Traditional lithology identification methods, such as crossplot analysis and statistical inversion, exhibit significant limitations, primarily manifested as insufficient seismic data resolution leading to challenges in thin-bedded reservoir identification, well-seismic calibration errors affecting model accuracy, and the subjective nature of manual interpretation failing to meet objective classification demands for complex lithological associations (Hossain 2020 ). Although seismic attribute analysis and pre-stack elastic parameter inversion (Wang et al. 2009 ) establish statistical relationships through rock physics models, their inversion processes suffer from non-uniqueness and low computational efficiency. In contrast, ensemble learning methods such as RF and gradient boosting trees (GBT) demonstrate notable advantages in lithology prediction due to their superior nonlinear modeling capabilities and noise resistance. Current research on lithology intelligent identification still faces several key challenges, including: 1) limited generalization ability of single models for complex geological features, 2) inefficient hyperparameter optimization in high-dimensional feature spaces (Bergstra et al. 2012), 3) impacts of class imbalance and noisy samples on model robustness (Chawla et al. 2002 ), 4) feature-label matching deviations caused by spatiotemporal misalignment between well logging and seismic data. Various solutions have been proposed to address these issues, covering the integration of ensemble learning with automated hyperparameter optimization (Akiba et al. 2019 ; Saleh et al. 2025 ; Soulaimani et al. 2024 ; Snoek et al. 2012 ; Wolpert 1992 ), improved sample balancing algorithms (He et al. 2008 ; Wang et al. 2021 ), and data alignment techniques, et al (Yu et al. 2025 ; Sakoe et al. 1978; Jacobs et al. 1991 ; Chen et al. 2021 ; Wu et al. 2019 ; Wu et al. 2023 ). However, existing studies predominantly focus on individual algorithmic optimizations without systematically comparing performance differences among diverse ensemble models and hyperparameter optimization strategies, with relatively limited exploration of dynamic feature selection mechanisms and data balancing approaches (Li et al. 2023 ; Wang et al. 2021 ). This study systematically evaluates performance differences among four mainstream ensemble algorithms (RF, XGBoost, LightGBM, CatBoost) and stacking models based on 3D seismic and well logging data from the F3 block in the North Sea. An integrated "feature selection-hyperparameter optimization-ensemble modeling" progressive framework is established by incorporating the Optuna hyperparameter optimization framework, recursive feature elimination technique, and NM-SMOTE sample balancing strategy. The research employs five-fold cross-validation combined with multi-dimensional evaluation metrics for comprehensive model assessment, aiming to provide reliable methodological guidance and theoretical foundations for lithology intelligent prediction in complex sedimentary systems. 2 Method 2.1 Random Forest Random Forest is an ensemble learning method based on bagging, introduced by Breiman in 2001 (Breiman 2001 ). The basic unit is the decision tree and the overall model is formed by integrating multiple decision trees (Fig. 1 ). Its core mechanism includes three parts (Tang 2024): 1) Building n diversified training subsets for each tree through Bootstrap sampling; 2) Generating k independent decision trees based on random feature selection; 3) The integration stage adopts a aggregation strategy: the final category is determined by majority voting for classification tasks, and the output prediction value is calculated by the mean for regression tasks. 2.2 XGBoost XGBoost is an efficient gradient boosting decision tree (GBDT) algorithm introduced by Chen et al. in 2016 (Chen 2016 ). It iteratively constructs decision trees by optimizing an objective function composed of a loss function and regularization components. Compared with traditional gradient boosting methods, XGBoost uses a second-order Taylor expansion to accelerate the gradient optimization process, introduces L1/L2 regularization to control model complexity, and improves computational efficiency through pre-sorting and weighted quantile techniques. It has good feature selection capabilities and robustness. 2.3 LightGBM LightGBM is a fast and effective gradient boosting decision tree (GBDT) method proposed by the Microsoft team in 2017 as an open-source project (Ke et al. 2017 ; Li et al. 2025 ). It focuses on improving the efficiency of large-scale data processing, using the histogram algorithm to reduce the amount of calculation; introducing unilateral gradient sampling and mutually exclusive feature bundling to optimize the data structure; and using a leaf-wise (best-first) tree growth strategy with depth constraints strategy to select the maximum gain for splitting leaf nodes each time (Fig. 2 ), which reduces the complexity of the model while reducing the computational overhead, and improves the generalization ability and accuracy of the model. 2.4 CatBoost CatBoost is an improved gradient boosting decision tree algorithm first proposed by the Yandex research team in 2017 (Prokhorenkova et al. 2018 ). It solves the problems of prediction offset and classification feature processing in traditional gradient boosting through three technical innovations: ordered boosting strategy (dynamically adjusting sample order to avoid gradient estimation bias), symmetric tree structure (using the same splitting rule for nodes at the same level), and ordered target statistics. The ordered target statistics balances the influence of noise through historical sample statistics and weight coefficient α, reducing the target leakage risk of category features. These technical innovations enable CatBoost to significantly improve generalization ability and computational efficiency while retaining the high precision advantage of GBDT. The formula (Yang et al. 2025 ) for calculating the target statistic of a sample is: $$\:{\text{x}}_{\text{k}}^{\text{i}}=\frac{{\sum\:}_{{\text{x}}_{\text{j}}\in\:{\text{D}}_{\text{k}}}\left\{{\text{x}}_{\text{k}}^{\text{i}}={\text{x}}_{\text{j}}^{\text{i}}\right\}{\text{y}}_{\text{i}}+{\alpha\:}\text{P}}{{\sum\:}_{{\text{x}}_{\text{j}\in\:{\text{D}}_{\text{k}}}}\left\{{\text{x}}_{\text{k}}^{\text{i}}={\text{x}}_{\text{j}}^{\text{i}}\right\}+{\alpha\:}}$$ In the formula, \(\:{\text{x}}_{\text{k}}^{\text{i}}\) 、 \(\:{\text{x}}_{\text{j}}\) 、 \(\:{\text{x}}_{\text{j}}^{\text{i}}\) 、 \(\:{\text{y}}_{\text{i}}\) are all training samples ; \(\:{\text{D}}_{\text{k}}\) is the data set before the K-th sample ; α is the weight coefficient ; P is the priori value. 2.5 Stacked Integration The Stacking Ensemble Method is an ensemble learning method that trains a meta-learner by combining the prediction results of multiple base learners to improve the performance of the model (Fig. 3 ). The base learner can be different types of models (such as Decision Trees, Support Vector Machines, Neural Networks, etc.), and the meta-learner is usually a simple regression or classification model (such as linear regression, logistic regression, etc.). The core idea is to integrate the outputs of multiple base learners by introducing a meta-learner to generate a more powerful prediction model (Wolpert 1992 ). For the integrated model, the stacking method is used in the study. The previously selected hyper-parametric models (RF, XGBoost, LightGBM, CatBoost) are used as the base models, and six methods such as regression algorithm, RF, GBT, SVM, and MLP are defined as meta-models. The meta-model is evaluated by 5-fold cross-validation, and the optimal meta-model is selected to construct the final stacking model for validation set prediction evaluation. The choice of meta-learner is based on two considerations : 1) Diversity : random forest and SVM are good at dealing with nonlinear features and linear separable problems respectively, and MLP can capture deep feature interaction ; 2) Computational efficiency : logistic regression and gradient boosting trees reduce computational overhead while ensuring accuracy. 2.6 Model hyperparameter optimization In order to fully exploit the performance of different models, four mainstream hyperparameter optimization strategies (Random Search, Grid Search, Bayesian Optimization and Optuna) are used to systematically optimize each base model. The optimal combination between the hyperparameter optimization strategy and each model is found by comprehensively comparing the evaluation parameters : Random Search: Randomized search CV (RandomizedSearchCV) is a hyper-parameter optimization method based on random sampling. It efficiently approximates the global optimal solution by combining cross-validation and is suitable for preliminary screening (such as the distribution range of parameters such as learning rate and maximum depth). In the specific implementation, the learning rate is set to a uniform distribution (uniform) of 0.01–0.3, the maximum depth (3–50), the number of iterations (100–1000), the number of trees (100–500) and the number of leaf nodes (20–150) are all integer random sampling (randint), and 50 iterations are performed through 5-fold cross-validation to obtain an approximate optimal parameter combination while ensuring search efficiency. Grid Search: Grid search CV (GridSearchCV) is a systematic hyper-parameter optimization method. By exhausting all combinations of predefined parameter spaces, the search grid is refined based on random search results (such as 0.5 / 1 / 1.5 times the benchmark value of the learning rate), and the calculation cost is high but the result is accurate. Bayesian Optimization algorithm: Bayesian optimization is a hyper-parameter optimization method based on probability model, which is suitable for scenarios with high computational cost of objective function (Pelikan et al. 1999 ; Gu et al. 2023 ). This method approximates the distribution of objective function by constructing Gaussian process surrogate model. The optimization process is divided into two stages : first, 10 random explorations are carried out to initialize the surrogate model, and then 50 iterative optimizations are carried out. In each iteration, the most potential hyper-parameter combination is selected for evaluation through balanced exploration and exploitation of acquisition functions (such as EI or UCB). The optimized objective function is 5-fold cross-validation accuracy. The search scope covers key parameters such as n _ estimators, learning _ rate, max _ depth, iterations, etc. Optuna algorithm: Optuna is an automated hyperparameter optimization framework based on Bayesian optimization, which supports a variety of optimization algorithms (such as TPE, CMA-ES, etc.). The framework is based on Bayesian optimization theory and achieves efficient parameter search through TPE algorithm (Akiba et al. 2019 ). In each model optimization, the search space of key parameters such as n _ estimators (100–1000), learning _ rate (0.01–0.3), max _ depth (3–50) is set, and 50 iterations are carried out with F1-score as the optimization goal. Optuna significantly reduces the computational cost through asynchronous parallel and early termination, which is especially suitable for complex model optimization in high-dimensional parameter space. 3 Data Sources and Preprocessing The data used in this study are from the publicly available seismic logging data set of the F3 exploration area in the North Sea, covering an area of about 387 km 2 . It is a default demonstration data set provided by the Opendtect software. The block is composed of a large river delta system sedimentary system. The delta inclusions are composed of sandstone and mudstone, and the porosity is generally high (20%-33%). The Inline range of the selected work area for this test is 100 ~ 700, the Crossline range is 300 ~ 1250, the time range is 408 ~ 1136 ms, the track spacing is 25 meters, and the sampling interval is 4 ms. Wells F02-1, F03-2, F03-4 and F06-1 in the work area were selected, which contain three-dimensional post-stack seismic data, longitudinal wave impedance data of seismic inversion, natural gamma data of geostatistical inversion, and lithology data of four wells (including sandstone and mudstone). Seismic data profile of Well F06-1 in the exploration area (Fig. 4 ). The red series (+ 2000 to + 8000) corresponds to positive phase reflection waves, revealing high wave impedance interfaces such as sandstone; the blue series (-2000 to -8000) represents negative phase reflections, which mostly indicate shale, mudstone or gas-bearing strata, and here indicates mudstone; the amplitude intensity is presented through the color saturation gradient, among which the yellow highlight band (Inline 400–600, 750-950ms) shows abnormally strong reflections, which may correspond to unconformity surfaces or the top boundaries of oil and gas reservoirs; the vertical black dotted lines (Inline 244, Inline 550) are typical stratigraphic tracing lines, showing that the amplitude polarity of the T3 reflection layer is reversed at 950 ms, suggesting a sudden change in stratigraphic lithology; a red-blue transition zone appears near 1000ms, reflecting the stratigraphic contact relationship of the sedimentary sequence transitioning from sandstone to mudstone; the oblique amplitude structure (Inline 300–500, 800-1000ms) has a progradational reflection configuration, which may be the seismic response of the delta front sedimentary body; the amplitude difference rate of its top and bottom interfaces is as high as 62%, and the amplitude variation with offset (AVO) characteristics show a Class III anomaly at Inline 450, which is consistent with the theoretical response model of gas-bearing sandstone. The preprocessing steps for the test area data are as follows: Resampling: First, read the time-depth relationship between the well point lithology data and the well data, and remove the outliers in the lithology data. Then perform time-depth conversion of the lithology data and resample according to the sampling rate of the seismic data (4ms). Finally, only the data retained within the time range of the geophysical data and located in the target layer are intercepted as lithology sample labels. After processing, a total of 352 lithology samples of two types are obtained, including 193 mudstones and 159 sandstones. Class balance: In order to further improve the robustness and recognition accuracy of the model and enhance the generalization performance, this study uses the nearest neighbor removal algorithm (NM) combined with the synthetic minority class oversampling algorithm (SMOTE) to form the NM-SMOTE method (Chawla et al. 2002 ; Wang et al. 2021 ) to resample the samples. The processing results are shown in Fig. 5 . On the one hand, this method increases the minority class samples (sandstone) through SMOTE, and on the other hand, it uses NM to remove noise samples to achieve the optimal balance of the data set (mudstone: sandstone = 193:193). Feature optimization: In the process of extracting lithological sample features, based on well logging curves and seismic attribute data, parameters with significant correlation with lithological classification, such as natural gamma (GR), longitudinal impedance (IP), and instantaneous frequency, were preliminarily screened out (Guyon et al. 2003). Then, recursive feature elimination (RFE) with a step size of 1 (one feature is eliminated in each iteration) and random forest were used as the base model. The feature weights were calculated using the Gini coefficient. The scoring criteria were balanced accuracy and combined with five-fold cross-validation for secondary screening. By gradually eliminating the features with the lowest contribution, the optimal number of features was determined to be 8. Finally, the lithological data of the four wells were used as lithological sample labels, and the well bypass data of the extracted lithological sample features were matched with them to form a lithological sample set. The analysis based on recursive feature elimination cross-validation (RFECV) showed that there was a significant correlation between the number of features and model performance. As shown in Fig. 6 , when the number of features was 8, the model achieved the best performance, with a balanced accuracy of 0.664 ± 0.068 (mean ± standard deviation), which had high classification performance and good stability. From the perspective of the impact of the number of features on model performance, when the number of features increased from 6 to 8, the model accuracy increased from 0.650 to 0.664, indicating that the two newly added features made a significant contribution to lithology discrimination; when the number of features exceeded 8, the model accuracy decreased by 0.012, indicating that the additional features may lead to overfitting and reduce the generalization ability of the model. In terms of stability, the model was the most unstable when there were 3 features (standard deviation 0.089), while the standard deviation dropped to 0.068 when there were 8 features, indicating that the model was the most robust to data fluctuations under this configuration. After comprehensive consideration of performance and stability, 8 features were finally determined as the optimal choice, and they were ranked in order of importance: natural gamma (GR), instantaneous amplitude, instantaneous frequency, longitudinal wave impedance (IP), root mean square energy, 24 Hz frequency division energy, 44 Hz frequency division energy, and 64 Hz frequency division energy. This feature combination effectively avoids the risk of overfitting while ensuring the accuracy of the model. 4 Modeling and evaluation 4.1 Evaluation indicators This study constructs a multi-dimensional evaluation system for systematic analysis. Firstly, the core indicators such as Accuracy, Recall, F1 score (F1-Score) and area under the receiver operating characteristic curve (AUC) were selected to comprehensively evaluate the classification ability of the model in different scenarios. In order to ensure the reliability of the results, 5-fold cross-validation was used. Each iteration was randomly divided into training set and verification set. Finally, the average value of each index was taken as the evaluation benchmark of the model performance. In order to visually display the performance of the model, three visualization methods are used : 1) box plot is used to display the distribution and stability of the accuracy of the model. The box median reflects the average level of the model, the interquartile range ( IQR ) reflects the fluctuation range, the line is used to mark the extreme value, and the outliers reveal the abnormal performance of the model. 2) The bar chart is used to compare the average accuracy between models, and the error line represents the standard deviation ; 3) The line chart is used to track the trend of accuracy with iteration in the process of hyperparameter optimization. In addition, in order to quantify the performance difference between models, the relative improvement rate index is introduced, and the calculation formula is as follows : The improvement rate = ( optimization model accuracy-base model accuracy ) / base model accuracy × 100%. This evaluation scheme combines quantitative indicators and multi-dimensional visualization methods, which can effectively avoid the misleading caused by the contingency of single indicator or data division, and enhance the credibility and explanatory power of the results. 4.2 Comparative analysis of hyperparameter optimization Through five-fold cross-validation (30% of the test set, three independent repeated experiments), this section compares the performance of four mainstream hyperparameter optimization methods (Grid Search, Random Search, Bayesian Optimization, Optuna) in different ensemble models. The main evaluation dimensions are : cross-validation accuracy, test set accuracy, training time, and relative improvement rate. In the Random Forest Model, Bayesian Search showed the best performance in the cross-validation stage. Its accuracy on the test set (Test Accuracy = 0.812) was significantly better than other methods (p < 0.05), and its cross-validation accuracy (CV Accuracy = 0.797) was also the highest value (Fig. 7 a). The analysis of Fig. 7 b further shows that the prediction results of Bayesian Search are the most stable (interquartile range IQR = 0.020), and the median and mean are highly consistent, indicating that its parameter search path has a low variance characteristic. In contrast, the IQR of random search and grid search are both 0.040, and the distribution is obviously skewed, reflecting the inefficiency of their search strategies and the instability of their results. This result verifies the effectiveness of Bayesian Search in dynamically balancing exploration and utilization capabilities through surrogate models (such as Gaussian processes). Its mechanism of iterative optimization of the search space based on historical feedback can significantly reduce the attempts of invalid parameter combinations. Table 1 summarizes the training time and improvement rate of each method. Bayesian search achieved an improvement rate of 6.56% with a training time of 86.32 seconds, with a significant efficiency advantage ( p < 0.01), while Optuna ranked second with an accuracy of 0.805. Its ratio of improvement rate (5.64%) to training time (101.13 seconds) (0.056%·s⁻¹) was significantly better than Bayesian search's 0.076%·s⁻¹, indicating the advantage of dynamic resource allocation strategy in the efficiency-accuracy trade-off. In contrast, although the grid search improved the improvement rate by 2.10%, it took as long as 150.40 seconds (an increase of 74%), verifying the inefficiency of the exhaustive strategy in high-dimensional parameter space. It is worth noting that the Bayesian range refinement and feature optimization variant has a training time of more than 130 seconds due to the shrinking of parameter range and the ranking of feature importance, but the improvement rate is only 4.72%~5.00%, which has not surpassed the basic version. This phenomenon shows that excessively restricting the search space or adding feature screening processes may weaken the global exploration ability of the Bayesian method. Based on the above results, Bayesian search is the preferred solution for parameter tuning because it has a balanced performance in accuracy, stability and efficiency, especially in high-dimensional non-convex optimization problems. Table 1 Comprehensive Performance of Random Forest Hyperparameter Optimization Method CV mean cv_std Test accuracy Training time(s) Improvement rate(%) Random Search 0.785 0.033 0.762 40.81s 0.00% Grid Search 0.790 0.033 0.778 150.40s 2.10% Bayesian Search 0.797 0.027 0.812 86.32s 6.56% Bayesian Search Range Refinement 0.793 0.039 0.798 130.34s 4.72% Bayesian Search Feature Optimization 0.793 0.039 0.800 131.52s 5.00% Optuna 0.797 0.031 0.805 101.13s 5.64% In the XGBoost model, the experimental results (Table 2 ) show that Optuna is significantly superior to other methods in terms of test accuracy (0.778) and improvement rate (4.29%), and the training time (10.32 seconds) is only 1.5% of Grid search (697.13 seconds), reflecting its best balance between model performance and computational efficiency. Although the cross-validation mean of Optuna (0.770) is slightly lower than that of Grid search (0.775) (Fig. 8 a), its test accuracy is higher and the standard deviation (cv _ std = 0.036) is the smallest, indicating that its optimization strategy can effectively alleviate the risk of overfitting. In contrast, Grid search is easy to fall into local optimal solution due to exhaustive parameter traversal, resulting in performance degradation in the test phase (test accuracy 0.762), and the standard deviation is large (cv _ std = 0.042). This result is consistent with the research of Bergstra et al. (Bergstra et al. 2012), that is, systematic search may not necessarily improve generalization ability, but may reduce stability due to high variance parameter combinations. Optuna 's training time is significantly shorter than other methods because of the synergistic effect of its TPE algorithm and early stopping strategy ( Pruning ). It prioritizes the exploration of high-potential parameter regions through probability models, and eliminates inefficient candidate schemes, thereby reducing redundant calculations. In contrast, the time cost of Grid search increases exponentially with the parameter dimension (such as the full combination traversal of hyperparameters such as learning rate and tree depth), resulting in a time of 11.6 minutes in this experiment, which is difficult to meet the real-time requirements in large-scale data scenarios. From the stability of the results (Fig. 8 b), the stability of Grid search is the best (IQR = 0.024), and the data distribution is concentrated. The Optuna model has the largest IQR value (0.074), reflecting that the optimization process is sensitive to the initial parameters. From the perspective of improvement rate, Optuna has a performance improvement of 4.29% compared with Random search (benchmark method), which is significantly higher than Bayesian search (1.21%) and Grid search (2.15%). This result verifies the superiority of Optuna in the efficiency of parameter space exploration. By dynamically adjusting the search direction through the reinforcement learning framework, Optuna avoids the dependence of Bayesian methods on prior distribution (such as the risk of suboptimal solutions when the Gaussian process hypothesis does not hold), and overcomes the blindness of Random search. Table 2 Comprehensive performance of XGBoost hyperparameter optimization Method CV mean cv_std Test accuracy Training time(s) Improvement rate(%) Random Search 0.757 0.058 0.746 4.87s 0.00% Grid Search 0.775 0.042 0.762 697.13s 2.15% Bayesian Search 0.765 0.045 0.755 21.75s 1.21% Optuna 0.770 0.036 0.778 10.32s 4.29% In the LightGBM model, as shown in Fig. 9 a, the Optuna framework is significantly superior to other methods in terms of test accuracy (0.778) and cross-validation stability (cv _ std = 0.031), with an increase of 8.96% (p < 0.01), and the boxplot (Fig. 9 b) shows that the distribution of test results is the most concentrated (IQR = 0.025), indicating that the model generalization ability is statistically robust. In contrast, although Grid Search achieves 6.44% performance improvement, the training time is 82.67 seconds, and the computational cost is more than 50 times that of Optuna (1.64 seconds), which verifies the inefficiency of exhaustive search strategy in high-dimensional parameter space. The Random Search and the original model (Ori _ best _ lgbm) did not achieve significant improvement (p > 0.05). The test accuracy was stable at the level of 0.714, and the cross-validation fluctuation range was 0.073 ( cv _ std ), which further explained the limitations of no guiding parameter exploration. The training time of Optuna is only 1.64 seconds (Table 3 ), the efficiency is 9.9 times (16.20 seconds) of the original model, and it is 98% faster than the Grid Search (82.67 seconds). It is recommended as the preferred solution for high-dimensional parameter optimization tasks. Table 3 Comprehensive performance of LightGBM hyperparameter optimization Method CV mean cv_std Test accuracy Training time(s) Improvement rate(%) Ori_best_lgbm 0.757 0.073 0.714 16.20s 0.0% Random Search 0.785 0.051 0.714 1.93s 0.0% Grid Search 0.770 0.048 0.760 82.67s 6.44% Bayesian Search 0.785 0.034 0.750 6.76s 5.04% Optuna 0.780 0.031 0.778 1.64s 8.96% It can be seen from Fig. 10 and Table 4 that in the CatBoost model, Optuna is superior to other methods in terms of test accuracy (0.810), improvement rate (6.3%) and IQR value. At the same time, the training time is the shortest (65.47s), and the CV standard deviation is the lowest (0.034), indicating that it has high efficiency and stability. Bayesian optimization further improves the convergence speed by predicting the potential benefits of parameter combinations through probability density functions, which is consistent with the conclusion of Snoek et al. on the efficiency of Bayesian optimization in black box function optimization(Snoek et al. 2012 ). Bayesian search is second only to Optuna in test accuracy (0.785) and improvement rate (3.02%), but its training time (158.28s) increases significantly. This is because Bayesian optimization needs to repeatedly update the Gaussian process model to predict parameter performance, and the computational complexity increases with the number of iterations. In contrast, Optuna 's TPE algorithm reduces the model update cost through a hierarchical sampling strategy, which is more efficient. The grid search series and random search perform poorly, the test accuracy is not more than 0.77, and the training time is generally more than 200 s. Although the exhaustive strategy of grid search can ensure the traversal of the local optimal solution, its time cost increases exponentially with the parameter dimension, and the search method of fixed step size is easy to miss the optimal interval. Dense sampling enhances the sensitivity of the model to training data noise. Similar conclusions have also been mentioned in the study of Hastie et al. (Hastie et al. 2004 ), that is, over-reliance on mesh accuracy may lead to the sacrifice of generalization of the model in exchange for local accuracy improvement. The non-directional sampling of random search leads to a large number of invalid attempts. Although its theoretical convergence is better than that of grid search, it is difficult to approach the global optimum under the finite number of iterations. This finding is consistent with the conclusion of Hutter and Bergstra et al. (Hutter et al. 2011 ; Bergstra et al. 2012) on hyperparameter optimization methods, that is, random search lacks guidance in high-dimensional space and is inefficient. For engineering scenarios that require fast iteration (such as online parameter tuning), Optuna 's high efficiency makes it the first choice. The training time is only 65.47 s, which is more than 3 times faster than Grid Search, and there is no need to sacrifice accuracy. This advantage is particularly critical in industrial applications with high real-time requirements. If the absolute accuracy is the only goal and the computing resources are sufficient, Bayesian search can be considered. Although its training takes a long time, the fineness of its parameter optimization may bring a weak but critical performance improvement in specific tasks. Table 4 Comprehensive performance of CatBoost hyperparameter optimization Method CV mean cv_std Test accuracy Training time(s) Improvement rate(%) Random Search 0.785 0.053 0.762 244.34s 0.00% Grid Search Best Parameters 0.765 0.038 0.778 200.58s 2.36% Grid Search 0.789 0.054 0.770 213.05s 1.05% Grid Search Small Range 0.790 0.054 0.762 215.22s 0.00% Bayesian Search 0.80 0.041 0.785 158.28s 3.02% Optuna 0.805 0.034 0.810 65.47s 6.3% 4.3 Comparison and evaluation of model performance Based on the accuracy, F1 score and AUC index, combined with the systematic comparison of ROC curve and precision-recall curve, this study evaluated the performance differences of five mainstream machine learning models (Random Forest, XGBoost, LightGBM, CatBoost, Stacking Model) in classification tasks. The results show that CatBoost performs best on the AUC value (0.91) and the precision under the condition of low recall rate (Recall0.95) (Fig. 12 ). It has a small IQR (Fig. 11 a). All the verification results are located in the range of 0.785–0.825, indicating that it has good generalization ability for data and can be stably played on multiple different training sets. The model incorporates a built-in categorical feature optimization mechanism (e.g., dynamic One-Hot encoding and the ordered lifting algorithm), which significantly enhances overfitting resistance and accuracy in small-sample scenarios. This makes it particularly suitable for applications requiring high sensitivity to missed detections and stable output performance. By integrating the advantages of multi-base models, the Stacking Model approaches CatBoost at the AUC (0.90) and accuracy (0.77) levels, showing a balanced bias-variance control ability, but its training time (1200 seconds) is significantly higher than other models (Fig. 11 b). In contrast, although LightGBM and XGBoost are superior in computational efficiency (training time is about 800 seconds and 100 seconds, respectively), their AUC values are 0.82 and 0.79, respectively, and the accuracy drops sharply under high recall conditions (Fig. 13 b), which requires targeted parameter adjustment to improve generalization ability. Random forest achieved a balance between stability (IQR = 0.06) and AUC (0.88) (Fig. 11 a, Fig. 12 ), but did not break through the performance boundary of single ensemble learning. In summary, if the accuracy and robustness are the core requirements, CatBoost is preferred ; if you need to balance accuracy and resource constraints, you can choose the Stacking Model or the optimized LightGBM. The research results provide a quantitative basis for model selection in complex scenarios. The follow-up work needs to further explore the joint optimization strategy of model lightweight and adversarial example defense. From Fig. 13 a (ROC curve analysis), we can see that the performance ranking of each model is CatBoost (AUC = 0.91) with the best performance, and the curve is closest to the upper left corner. When FPR = 0.2, TPR has reached 0.8, and when FPR = 0.3, TPR = 0.85, showing strong early recognition ability, and significantly better than other models when FPR 0.5, it performs better than CatBoost, and when FPR = 0.2, TPR is about 0.8, which is suitable for medium risk tolerance scenarios; Random Forest (AUC = 0.88) performs stably but slightly worse than CatBoost and the stacking model, LightGBM (AUC = 0.82) and XGBoost (AUC = 0.79) are relatively weak, and the XGBoost curve is closest to the diagonal, indicating that its discrimination ability is relatively limited. From Fig. 13 b (P-R curve analysis), it can be seen that in the performance characteristics, CatBoost maintains the highest curve position, Precision remains above 0.9 when Recall = 0.5, and Precision drops rapidly to about 0.78 after Recall > 0.7, which is suitable for the accuracy scenario with Recall requirement < 0.7. The precision of the stacking model drops rapidly from its peak value to 0.5, indicating that under extremely low recall conditions(Recall), the model can only identify the most easily classifiable positive samples (such as sandstone with simple lithology and obvious logging response). The accuracy increases from 0.5 to 0.95, indicating that as the recall rate increases, the model gradually covers more ' medium-difficulty ' positive samples (such as pebbly sandstone or argillaceous sandstone). The feature discrimination of these samples is good, and the model can maintain high accuracy stably, reaching a peak accuracy (0.95) at Recall = 0.57. It shows that the model has a strong ability to identify most common lithologies (such as sandstone and mudstone) and meets the actual needs. 4.4 Lithology prediction results F02-1, F06-1, F03-2 and F03-4 are used as verification wells in turn, and the lithology samples of the other three wells are used for model training. In order to facilitate the comparison of the prediction results of each method, the prediction results of each method are listed here when F06-1 is used as the verification well. Figures 14 , 15 , 16 , 17 and 18 are the formation lithology and sandstone probability of the horizontal survey line of F06-1 well predicted by the five models. Figures 14 a, 15 a, 16 a, 17 a and 18 a are the model prediction results of the constructed random forest method, XGBoost method, LightGBM method, CatBoost method and stacked ensemble method, respectively. It can be seen that the prediction results of lithology by each method have a certain correspondence with the lithology of the verification wells used, but the prediction accuracy and resolution of different methods are slightly different. Among them, CatBoost is more precise in the identification of sand and mudstone. The probability of sandstone gradually changes along the Inline, which is in line with the geological law, and the lithology continuity is better (Fig. 17 a). The probability profile of sandstone is stable (Fig. 17 b). The probability value shows a more gentle gradient along the Inline, and there is no mutation in the sand-mud transition zone, indicating that the model is more robust to noise or complex features. The Stacking model demonstrates good performance on easily classifiable samples (e.g., thick sandstone layers or mudstone), while its effectiveness in classifying complex samples (e.g., thin interbedded layers) is limited (Fig. 18 ). The random forest prediction boundary is blurred, the sandstone probability profile jumps obviously (as shown in Fig. 14 b, the shallow lateral color jumps obviously), and the shallow sandstone recognition ability is weak. The LightGBM thick sandstone prediction is continuous (Time 900-1100ms), but the shallow layer (Time < 900ms) is not continuous (the classification near the fault in Fig. 16 a is blurred). XGBoost has a strong ability to capture details, but as shown in Fig. 15 , it is sensitive to data noise, the continuity of lithology prediction is not strong, and the overall accuracy is the lowest. The classification algorithm of each model is feasible in seismic reservoir classification. At the well location, the actual data is consistent with the classification prediction results. On the other hand, the classification results far away from the well position outline the corresponding response of the stratigraphic morphology to the seismic events on the seismic profile (Wang et al. 2023 ). It can be seen that the prediction results have a good correspondence with the lithology of the verification well, and the sandstone probability profile of CatBoost has the highest degree of agreement with the actual lithology (Fig. 17 ), especially in the target layer (depth 700–1136 meters). The accuracy of sandstone identification is more than 75%. Figures 14 b, 15 b, 16 b, 17 b and 18 b are the sandstone probability profiles of each method, which quantitatively reflect the possibility that the formation lithology in the target interval belongs to sandstone. 5 Conclusion This study compares and analyzes the performance of four hyperparameter optimization methods and five ensemble models in lithology identification, and draws the following conclusions: (1) In terms of optimization methods, the Optuna framework achieves the best balance between efficiency and accuracy, and CatBoost achieves an accuracy of 0.810 in only 65.47 seconds after optimization. Random search demonstrates the fastest execution speed, and XGBoost optimization takes only 4.87 seconds. Bayesian optimization shows the best stability. (2) In terms of model performance, CatBoost achieves the highest AUC (0.91), excelling in complex lithology identification.The stacked ensemble model exhibits superior robustness (5-fold CV standard deviation: 0.041). Random forest training is the most efficient (average 53 seconds). LightGBM and XGBoost are suitable for fast response and standard accuracy scenarios, respectively. (3) Furthermore, our findings confirm that combining resampling techniques with feature selection significantly improves the robustness and adaptability of machine learning models in seismic reservoir lithology identification. This is consistent with recent studies highlighting the synergy of these two strategies for handling imbalanced geological data. The research results provide an important model selection basis for intelligent lithology identification in oil and gas exploration. Declarations Conflict of interest All authors acknowledged that there is no conflict of interest on record. Funding China National Nuclear Corporation-State Key Laboratory of Nuclear Resources and Environment (East China University of Technology) Joint Innovation Fund Project (No. 2023NRE-LH-08); National Natural Science Foundation of China (Nos. U2244205, U2067202). Author contributions All authors listed made a substantial, direct, and intellectual contribution to the work and approved it for publication Data Availability Statement The seismic data supporting this study are publicly available in the Terranubis open data repository at https://terranubis.com/datainfo/F3-Demo-2023 . Processed datasets and Python scripts generated during this research are available from the corresponding author upon reasonable request. References Akiba T, Sano S, Yanase T et al (2019) Optuna: A next-generation hyperparameter optimization framework. Knowl Discovery Data Min 2623–2631. https://doi.org/10.1145/3292500.3330701 Ali M, Changxingyue H, Wei N et al (2025) Optimizing seismic-based reservoir property prediction: a synthetic data-driven approach using convolutional neural networks and transfer learning with real data integration. Artif Intell Rev 58:31. https://doi.org/10.1007/s10462-024-11030-8 Amiri-Ramsheh B, Larestani A, Atashrouz S et al (2025) Toward accurate prediction of carbon dioxide (CO2) compressibility factor using tree-based intelligent schemes (XGBoost and LightGBM) and equations of state. Results Eng Volume 25 104035, ISSN 2590 – 1230. https://doi.org/10.1016/j.rineng.2025.104035 Bergstra J, Bengio Y (2012) Random search for hyper-parameter optimization. J Mach Learn Res 13(1):281–305. https://doi.org/10.1016/j.chemolab.2011.12.002 Breiman L (2001) Random forests. Mach Learn 45:5–32. https://doi.org/10.1023/A:1010933404324 Chawla NV, Bowyer KW, Hall LO et al (2002) SMOTE: Synthetic Minority Over-sampling Technique. J Artif Intell Res 16(1):321–357. https://doi.org/10.1613/jair.953 Cheng X, Zhou J, Fu H et al (2023) Applicability and Application of Machine Learning Algorithm in Logging Interpretation. Northwest Geol 56(4):336–348. https://doi.org/10.12401/j.nwg.2023029 Chen Y, Zhao L, Pan J et al (2021) Deep carbonate reservoir characterization using multi-seismic attributes via machine learning with physical constraints. J Geophys Eng 18(5):761–775. https://doi.org/10.1093/jge/gxab049 Chen T, Guestrin C (2016) XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '16). Association for Computing Machinery, New York, NY, USA, 785–794. https://doi.org/10.1145/2939672.2939785 Guyon I, Elisseeff A (2003) An introduction to variable and feature selection. J Mach Learn Res 3(7–8):1157–1182. https://doi.org/10.1162/153244303322753616 Gu Y, Zhang S, Wang W et al (2023) Lithofacies prediction driven by logging-based Bayesian-optimized ensemble learning: a case study of lacustrine carbonate reservoirs. Geophys Prospect 71:1835–1872. https://doi.org/10.1111/1365-2478.13258 Gao S, Xu M, Zhao L et al (2023) Seismic predictions of fluids via supervised deep learning: Incorporating various class-rebalance strategies. Geophysics 88(4):M185–M200. https://doi.org/10.1190/geo2022-0363.1 He H, Bai Y, Garcia EA et al (2008) ADASYN: Adaptive Synthetic Sampling Approach for Imbalanced Learning. Proceedings of the International Joint Conference on Neural Networks 1322–1328. https://doi.org/10.1109/IJCNN.2008.4633969 Hastie T, Tibshirani R, Friedman JH et al (2004) The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Math Intelligencer 27:83–85. https://doi.org/10.1007/BF02985802 Hutter F, Hoos HH, Leyton-Brown K (2011) Sequential model-based optimization for general algorithm configuration. In Proceedings of the 5th international conference on Learning and Intelligent Optimization (LION'05). Springer-Verlag, Berlin, Heidelberg 507–523. https://doi.org/10.1007/978-3-642-25566-3_40 Hossain S (2020) Application of seismic attribute analysis in fluvial seismic geomorphology. J Petroleum Explor Prod Technol 10:1009–1019. https://doi.org/10.1007/s13202-019-00809-z Jacobs RA, Jordan MI, Nowlan SJ et al (1991) Adaptive mixtures of local experts. Neural Comput 3(1):79–87. https://doi.org/10.1162/neco.1991.3.1.79 Kouassi A, Pan L, Wang X et al (2023) Identification of karst cavities from 2D seismic wave impedance images based on gradient-boosting decision trees algorithms (GBDT): Case of Ordovician fracture-vuggy carbonate reservoir. Energies 16(2):643. https://doi.org/10.3390/en16020643 Ke G, Meng Q, Finley T et al (2017) LightGBM: A highly efficient gradient boosting decision tree. Neural Inform Process Syst 30:3146–3154 Li L, Sharma A, Cheng S (2025) Explainable AI based LightGBM prediction model to predict default borrower in social lending platform. Intell Syst Appl Volume 26:2667–3053. https://doi.org/10.1016/j.iswa.2025.200514 Li C, Wang L, Chen Y et al (2023) Prediction Model of Soils' Preconsolidation Pressure Based on Bayesian Ensemble Learning Algorithm. Earth Sci 48(5):1780–1792. https://doi.org/10.3799/dqkx.2022.450 (in Chinese) Naveed S, Husnain M (2025) A drug recommendation system based on response prediction: Integrating gene expression and K-mer fragmentation of drug SMILES using LightGBM. Intelligence-Based Med Volume 11:2666–5212. https://doi.org/10.1016/j.ibmed.2025.100206 Prokhorenkova L, Gusev G, Vorobev A et al (2018) CatBoost: Unbiased boosting with categorical features. Neural Inform Process Syst 31:6638–6648. https://doi.org/10.48550/arXiv.1706.09516 Pelikan M, Goldberg DE, Cantú-Paz E (1999) BOA: the Bayesian optimization algorithm. Genetic and Evolutionary Computation Conference 525–532. https://ci.nii.ac.jp/naid/10012961429 Sakoe H, Chiba S (1978) Dynamic programming algorithm optimization for spoken word recognition. IEEE Trans Acoust Speech Signal Process 26(1):43–49. https://doi.org/10.1109/TASSP.1978.1163055 Snoek J, Larochelle H, Adams RP (2012) Practical bayesian optimization of machine learning algorithms. Neural Inform Process Syst 4. https://doi.org/10.48550/arXiv.1206.2944 Saleh K, Mabrouk WM, Metwally A (2025) Machine learning model optimization for compressional sonic log prediction using well logs in Shahd SE field, Western Desert, Egypt. Sci Rep 15:14957. https://doi.org/10.1038/s41598-025-97938-9 Soulaimani S, Soulaimani A, Abdelrahman K et al (2024) Advanced machine learning artificial neural network classifier for lithology identification using Bayesian optimization. Front Earth Sci 12:1473325. https://doi.org/10.3389/feart.2024.1473325 Topór T, Sowiżdżał K (2022) Application of machine learning tools for seismic reservoir characterization study of porosity and saturation type. Nafta-Gaz 3:165–175. https://doi.org/10.18668/NG.2022.03.01 Tang Y, Yang J, You Y et al (2024) Multi-output prediction for TBM operation parameters based on stacking ensemble algorithm. Tunn Undergr Space Technol Volume 152:0886–7798. https://doi.org/10.1016/j.tust.2024.105960 Wang M, Yang J, Wang X et al (2023) Identification of Shale Lithofacies by Well Logs Based on Random Forest Algorithm. Earth Sci 48(1):130–142. https://doi.org/10.3799/dqkx.2022.181 (in Chinese) Wang W, Lin C, Wang J et al (2009) Pre-stack elastic parameters inversion and its application. Geophys Prospect Petroleum 48(5):483–487. https://doi.org/10.3969/j.issn.1000-1441.2009.05.010 (in Chinese) Wolpert DH (1992) Stacked generalization. Neural networks 5(2):241–259. https://doi.org/10.1016/S0893-6080(05)80023-1 Wu X, Liang L, Shi Y et al (2019) FaultSeg3D: Using synthetic data sets to train an end-to-end convolutional neural network for 3D seismic fault segmentation. Geophysics 84(3):IM35–IM45. https://doi.org/10.1190/geo2018-0646.1 Wu J, Shi Y, Wang K et al (2023) Automatic seismic fault identification based on an improved U-Net network. Acta Geophys. https://doi.org/10.1007/s11600-023-01200-7 Wang G, Song J, Xu F et al (2021) Random Forests lithology prediction method for imbalanced data sets. Oil Geophys Prospect 56(4):679–687. https://doi.org/10.13810/j.cnki.issn.1000-7210.2021.04.001 (in Chinese) Yu H, Wang H, Wu L et al (2025) A precise time–depth conversion method for coal seam based on machine learning and seismic velocity inversion. Acta Geophys. https://doi.org/10.1007/s11600-025-01592-8 Yang F, Xiao F, Shi B et al (2025) Deterioration trend assessment of hydropower units based on Optuna-CatBoost and CRITIC evaluation method. Hydropower Energy Sci 43(02):206–210. https://doi.org/10.20040/j.cnki.1000-7709.2025.20241751 (in Chinese) Zheng Y, Li X, Chen Y et al (2024) Short-term wind power forecasting method for extreme weather based on stacking multi-model fusion. High Voltage Eng 50(09):3871–3882. https://doi.org/10.13336/j.1003-6520.hve.20240490 Zou Y, Li Q, Que W et al (2024) Comparison of lithology identification of in situ leaching logging data based on different machine learning algorithms. Uranium Min Metall 1–10. https://doi.org/10.13426/j.cnki.yky.2024.10.09 (in Chinese) Cite Share Download PDF Status: Published Journal Publication published 10 Dec, 2025 Read the published version in Acta Geophysica → Version 1 posted Editorial decision: Major revisions 15 Aug, 2025 Reviewers agreed at journal 08 Jul, 2025 Reviewers invited by journal 08 Jul, 2025 Editor invited by journal 19 Jun, 2025 Editor assigned by journal 06 Jun, 2025 First submitted to journal 04 Jun, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6795780","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":482319723,"identity":"ff3a9bfa-cf97-4def-98c2-6d0f9b01cfb0","order_by":0,"name":"Wang Jingyi","email":"","orcid":"","institution":"East China University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Wang","middleName":"","lastName":"Jingyi","suffix":""},{"id":482319724,"identity":"c26fbc91-ec80-42c5-8d4f-1fe2cd03a8b9","order_by":1,"name":"Jiang Li","email":"","orcid":"","institution":"East China University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Jiang","middleName":"","lastName":"Li","suffix":""},{"id":482319725,"identity":"97567ce8-a90d-49af-80fc-c1ef729a154b","order_by":2,"name":"Zhibing Feng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyElEQVRIiWNgGAWjYDACCcYHQJJNhp+Z+fADIrUwG4C08Ei2s6UZkKKFgcfgPI+CBFE65Gc3Mz4u+MXHY3yYh8GAocYmmqAWxjmHmY1n9rHxmB3mPfCA4VhabgMhLcwS+cekeXtAWvgSDBgbDhPWwiaRzP4bpMW4mcdAgigtPBLJbMw8P9h4DJiJ1SIhc5hZmreBjUfiMDCQE4jxCyjEPvP8OSbH33/48IMPNTaEtYABY9sxCCOBKOVg8KeGeLWjYBSMglEw8gAAJMM1o7dHL+gAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0002-5063-1674","institution":"East China University of Technology","correspondingAuthor":true,"prefix":"","firstName":"Zhibing","middleName":"","lastName":"Feng","suffix":""},{"id":482319726,"identity":"28968b0c-7913-420c-9aad-d610068a1b98","order_by":3,"name":"Huang Xiao","email":"","orcid":"","institution":"Geologic party No. 243, China National Nuclear Corporation","correspondingAuthor":false,"prefix":"","firstName":"Huang","middleName":"","lastName":"Xiao","suffix":""},{"id":482319727,"identity":"da021283-973a-45bd-a418-fb3f90b6c879","order_by":4,"name":"Yao Zhenan","email":"","orcid":"","institution":"East China University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Yao","middleName":"","lastName":"Zhenan","suffix":""},{"id":482319728,"identity":"f9bb2ade-9d31-4263-aab7-09283aaeba26","order_by":5,"name":"Zhang Bocheng","email":"","orcid":"","institution":"East China University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Zhang","middleName":"","lastName":"Bocheng","suffix":""}],"badges":[],"createdAt":"2025-06-01 13:04:53","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6795780/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6795780/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s11600-025-01723-1","type":"published","date":"2025-12-10T15:56:58+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":86392001,"identity":"23e0ee6a-ea1e-4ee2-b02a-fc61dfa75d46","added_by":"auto","created_at":"2025-07-10 07:06:47","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":28989,"visible":true,"origin":"","legend":"\u003cp\u003eRandom forest model structure (Zou et al. 2024)\u003c/p\u003e","description":"","filename":"image1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/869143972d590d208e1fb30d.jpeg"},{"id":86393561,"identity":"f49907ca-7500-4469-a475-59de7f99611c","added_by":"auto","created_at":"2025-07-10 07:22:47","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":9978,"visible":true,"origin":"","legend":"\u003cp\u003eLightGBM architecture (Amiri-Ramsheh et al. 2025; Naveed et al. 2025)\u003c/p\u003e","description":"","filename":"image2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/2c88e166c5fd19322c6fdd87.jpeg"},{"id":86391997,"identity":"b2b42cf2-c5a3-4e1e-981e-98507ebb5bb5","added_by":"auto","created_at":"2025-07-10 07:06:47","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":25353,"visible":true,"origin":"","legend":"\u003cp\u003eBasic structure diagram of stacked integration, where TS represents the training set (Zheng et al. 2024)\u003c/p\u003e","description":"","filename":"image3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/1224b8c62e5e5cbbe3a235e5.jpeg"},{"id":86391999,"identity":"67480781-5480-42b0-8c97-b757232c0294","added_by":"auto","created_at":"2025-07-10 07:06:47","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":118592,"visible":true,"origin":"","legend":"\u003cp\u003eSeismic data profile of F06-1 well in the F3 exploration area of the North Sea\u003c/p\u003e","description":"","filename":"image4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/300d58ba6a93fbc0ec31955a.jpeg"},{"id":86392009,"identity":"fd1bddb3-e543-4a85-b711-52ea9a889760","added_by":"auto","created_at":"2025-07-10 07:06:48","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":54580,"visible":true,"origin":"","legend":"\u003cp\u003eSample distribution before and after NM-SMOTE balancing. \u003cstrong\u003ea\u003c/strong\u003e Original data distribution, \u003cstrong\u003eb\u003c/strong\u003eBalanced sample distribution.\u003c/p\u003e","description":"","filename":"image5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/2390854b532ec6890e57ba4a.jpeg"},{"id":86392012,"identity":"818dc154-80ee-4b97-aea8-290cc07045ac","added_by":"auto","created_at":"2025-07-10 07:06:48","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":67295,"visible":true,"origin":"","legend":"\u003cp\u003eResult of automatic feature optimization. \u003cstrong\u003ea \u003c/strong\u003eCross-validation balances the accuracy distribution\u003cstrong\u003e, b \u003c/strong\u003eStandard Deviation Distribution\u003c/p\u003e","description":"","filename":"image6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/96de099fd983e0dc395b5bd1.jpeg"},{"id":86393047,"identity":"423fd934-026a-4aca-9a0c-6e2d982b65b5","added_by":"auto","created_at":"2025-07-10 07:14:47","extension":"jpeg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":66763,"visible":true,"origin":"","legend":"\u003cp\u003eRandom Forest hyperparameter optimization results. \u003cstrong\u003ea\u003c/strong\u003e Cross-validation accuracy. \u003cstrong\u003eb\u003c/strong\u003eCross-validation score distribution\u003c/p\u003e","description":"","filename":"image7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/ee3cbbfadd9064a87cff0858.jpeg"},{"id":86392019,"identity":"285ed68a-abe6-44c6-b6e1-f6134287e323","added_by":"auto","created_at":"2025-07-10 07:06:48","extension":"jpeg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":60903,"visible":true,"origin":"","legend":"\u003cp\u003eXGBoost hyperparameter optimization results.\u003cstrong\u003e a\u003c/strong\u003e Cross-validation accuracy. \u003cstrong\u003eb\u003c/strong\u003e Cross-validation score distribution\u003c/p\u003e","description":"","filename":"image8.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/951781bea5358c80fdae1b83.jpeg"},{"id":86393057,"identity":"16f7b48b-1c3e-49c9-9884-8736582e6d04","added_by":"auto","created_at":"2025-07-10 07:14:48","extension":"jpeg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":61601,"visible":true,"origin":"","legend":"\u003cp\u003eLightGBM hyperparameter optimization results. \u003cstrong\u003ea\u003c/strong\u003e Cross-validation accuracy. \u003cstrong\u003eb\u003c/strong\u003eCross-validation score distribution\u003c/p\u003e","description":"","filename":"image9.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/25902668d98757afbda9e208.jpeg"},{"id":86392002,"identity":"ad5cae72-0783-4fc2-b286-55121be9171b","added_by":"auto","created_at":"2025-07-10 07:06:47","extension":"jpeg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":57075,"visible":true,"origin":"","legend":"\u003cp\u003eCatBoost hyperparameter optimization results.\u003cstrong\u003e a\u003c/strong\u003e Cross-validation accuracy. \u003cstrong\u003eb\u003c/strong\u003eCross-validation score distribution\u003c/p\u003e","description":"","filename":"image10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/7fb898f0e42df6de7c00dfdf.jpeg"},{"id":86393055,"identity":"745c00a1-6162-48e8-9bc5-244c54b7b174","added_by":"auto","created_at":"2025-07-10 07:14:48","extension":"jpeg","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":80257,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ea\u003c/strong\u003e Cross-validation accuracy score distribution. \u003cstrong\u003eb\u003c/strong\u003e Model average accuracy and training time\u003c/p\u003e","description":"","filename":"image11.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/13ba7ac8fd9725e8e62c4368.jpeg"},{"id":86393056,"identity":"2c7f0c9f-6551-4d17-ae2a-3b15ec329e0c","added_by":"auto","created_at":"2025-07-10 07:14:48","extension":"jpeg","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":49233,"visible":true,"origin":"","legend":"\u003cp\u003eModel performance comparison\u003c/p\u003e","description":"","filename":"image12.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/558abe4e332ec2723bf8d712.jpeg"},{"id":86392008,"identity":"9379b083-4fb6-404e-bb6e-a0ca0cfb10e7","added_by":"auto","created_at":"2025-07-10 07:06:48","extension":"jpeg","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":71176,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ea\u003c/strong\u003e ROC Curve Comparison. \u003cstrong\u003eb\u003c/strong\u003e Precision-Recall Curve Comparison\u003c/p\u003e","description":"","filename":"image13.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/aa6d37d5ac5b38e5e6673c56.jpeg"},{"id":86392017,"identity":"15be2f66-5e90-4801-9eba-bfa1f0a04433","added_by":"auto","created_at":"2025-07-10 07:06:48","extension":"jpeg","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":265824,"visible":true,"origin":"","legend":"\u003cp\u003eRandom forest lithology prediction results, \u003cstrong\u003ea\u003c/strong\u003e predicted stratum lithology, \u003cstrong\u003eb\u003c/strong\u003esandstone probability\u003c/p\u003e","description":"","filename":"image14.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/d5b9e202d5a7a6d82feaad74.jpeg"},{"id":86392035,"identity":"87047e41-daaf-4cbf-a6ef-44347ed57205","added_by":"auto","created_at":"2025-07-10 07:06:49","extension":"jpeg","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":308678,"visible":true,"origin":"","legend":"\u003cp\u003eXGBoost lithology prediction results, \u003cstrong\u003ea\u003c/strong\u003e predicted stratum lithology, \u003cstrong\u003eb\u003c/strong\u003esandstone probability\u003c/p\u003e","description":"","filename":"image15.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/42c5f8d6ef48e4533eb7bdcc.jpeg"},{"id":86392053,"identity":"66a872c4-6943-4e37-82d0-c6f5600c11f5","added_by":"auto","created_at":"2025-07-10 07:06:50","extension":"jpeg","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":292664,"visible":true,"origin":"","legend":"\u003cp\u003eLightGBM lithology prediction results, \u003cstrong\u003ea\u003c/strong\u003e predicted stratum lithology, \u003cstrong\u003eb\u003c/strong\u003esandstone probability\u003c/p\u003e","description":"","filename":"image16.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/cabfccb95293921e7e176882.jpeg"},{"id":86393071,"identity":"442d35e9-a638-4b73-a7a7-814ea58b7184","added_by":"auto","created_at":"2025-07-10 07:14:49","extension":"jpeg","order_by":17,"title":"Figure 17","display":"","copyAsset":false,"role":"figure","size":291814,"visible":true,"origin":"","legend":"\u003cp\u003eCatBoost lithology prediction results, \u003cstrong\u003ea\u003c/strong\u003e predicted stratum lithology, \u003cstrong\u003eb\u003c/strong\u003esandstone probability\u003c/p\u003e","description":"","filename":"image17.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/ccb96373d54d28cda6d6e573.jpeg"},{"id":86393064,"identity":"18150e7a-5bca-4105-9680-8c9ac69403a1","added_by":"auto","created_at":"2025-07-10 07:14:49","extension":"jpeg","order_by":18,"title":"Figure 18","display":"","copyAsset":false,"role":"figure","size":290395,"visible":true,"origin":"","legend":"\u003cp\u003eLithology prediction results of stacking model, \u003cstrong\u003ea\u003c/strong\u003e predicted stratum lithology, \u003cstrong\u003eb\u003c/strong\u003esandstone probability\u003c/p\u003e","description":"","filename":"image18.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/d55c6f5f7928f4dcbffc53fb.jpeg"},{"id":98243489,"identity":"7092e718-87d9-4ded-bb2e-98967addd385","added_by":"auto","created_at":"2025-12-15 16:06:38","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3107225,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6795780/v1/41dd2802-fac0-42c5-b3e3-956d6ed51921.pdf"}],"financialInterests":"","formattedTitle":"Optimization and Evaluation of Ensemble Learning Models for Intelligent Lithology Identification Using Seismic Data","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eWith global hydrocarbon exploration targets gradually shifting towards complex and concealed reservoirs, lithology intelligent identification technology based on seismic and well logging data has become increasingly critical for reservoir characterization (Ali et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Cheng et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Traditional lithology identification methods, such as crossplot analysis and statistical inversion, exhibit significant limitations, primarily manifested as insufficient seismic data resolution leading to challenges in thin-bedded reservoir identification, well-seismic calibration errors affecting model accuracy, and the subjective nature of manual interpretation failing to meet objective classification demands for complex lithological associations (Hossain \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Although seismic attribute analysis and pre-stack elastic parameter inversion (Wang et al. \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2009\u003c/span\u003e) establish statistical relationships through rock physics models, their inversion processes suffer from non-uniqueness and low computational efficiency. In contrast, ensemble learning methods such as RF and gradient boosting trees (GBT) demonstrate notable advantages in lithology prediction due to their superior nonlinear modeling capabilities and noise resistance.\u003c/p\u003e\u003cp\u003eCurrent research on lithology intelligent identification still faces several key challenges, including: 1) limited generalization ability of single models for complex geological features, 2) inefficient hyperparameter optimization in high-dimensional feature spaces (Bergstra et al. 2012), 3) impacts of class imbalance and noisy samples on model robustness (Chawla et al. \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2002\u003c/span\u003e), 4) feature-label matching deviations caused by spatiotemporal misalignment between well logging and seismic data. Various solutions have been proposed to address these issues, covering the integration of ensemble learning with automated hyperparameter optimization (Akiba et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Saleh et al. \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Soulaimani et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Snoek et al. \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2012\u003c/span\u003e; Wolpert \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e1992\u003c/span\u003e), improved sample balancing algorithms (He et al. \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2008\u003c/span\u003e; Wang et al. \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), and data alignment techniques, et al (Yu et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Sakoe et al. 1978; Jacobs et al. \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e1991\u003c/span\u003e; Chen et al. \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Wu et al. \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Wu et al. \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). However, existing studies predominantly focus on individual algorithmic optimizations without systematically comparing performance differences among diverse ensemble models and hyperparameter optimization strategies, with relatively limited exploration of dynamic feature selection mechanisms and data balancing approaches (Li et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Wang et al. \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThis study systematically evaluates performance differences among four mainstream ensemble algorithms (RF, XGBoost, LightGBM, CatBoost) and stacking models based on 3D seismic and well logging data from the F3 block in the North Sea. An integrated \"feature selection-hyperparameter optimization-ensemble modeling\" progressive framework is established by incorporating the Optuna hyperparameter optimization framework, recursive feature elimination technique, and NM-SMOTE sample balancing strategy. The research employs five-fold cross-validation combined with multi-dimensional evaluation metrics for comprehensive model assessment, aiming to provide reliable methodological guidance and theoretical foundations for lithology intelligent prediction in complex sedimentary systems.\u003c/p\u003e"},{"header":"2 Method","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n\u003ch2\u003e2.1 Random Forest\u003c/h2\u003e\n\u003cp\u003eRandom Forest is an ensemble learning method based on bagging, introduced by Breiman in 2001 (Breiman \u003cspan class=\"CitationRef\"\u003e2001\u003c/span\u003e). The basic unit is the decision tree and the overall model is formed by integrating multiple decision trees (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e). Its core mechanism includes three parts (Tang 2024):\u003c/p\u003e\n\u003c/div\u003e\n\u003cp\u003e1) Building n diversified training subsets for each tree through Bootstrap sampling;\u003c/p\u003e\n\u003cp\u003e2) Generating k independent decision trees based on random feature selection;\u003c/p\u003e\n\u003cp\u003e3) The integration stage adopts a aggregation strategy: the final category is determined by majority voting for classification tasks, and the output prediction value is calculated by the mean for regression tasks.\u003c/p\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\n\u003ch2\u003e2.2 XGBoost\u003c/h2\u003e\n\u003cp\u003eXGBoost is an efficient gradient boosting decision tree (GBDT) algorithm introduced by Chen et al. in 2016 (Chen \u003cspan class=\"CitationRef\"\u003e2016\u003c/span\u003e). It iteratively constructs decision trees by optimizing an objective function composed of a loss function and regularization components. Compared with traditional gradient boosting methods, XGBoost uses a second-order Taylor expansion to accelerate the gradient optimization process, introduces L1/L2 regularization to control model complexity, and improves computational efficiency through pre-sorting and weighted quantile techniques. It has good feature selection capabilities and robustness.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n\u003ch2\u003e2.3 LightGBM\u003c/h2\u003e\n\u003cp\u003eLightGBM is a fast and effective gradient boosting decision tree (GBDT) method proposed by the Microsoft team in 2017 as an open-source project (Ke et al. \u003cspan class=\"CitationRef\"\u003e2017\u003c/span\u003e; Li et al. \u003cspan class=\"CitationRef\"\u003e2025\u003c/span\u003e). It focuses on improving the efficiency of large-scale data processing, using the histogram algorithm to reduce the amount of calculation; introducing unilateral gradient sampling and mutually exclusive feature bundling to optimize the data structure; and using a leaf-wise (best-first) tree growth strategy with depth constraints strategy to select the maximum gain for splitting leaf nodes each time (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e), which reduces the complexity of the model while reducing the computational overhead, and improves the generalization ability and accuracy of the model.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n\u003ch2\u003e2.4 CatBoost\u003c/h2\u003e\n\u003cp\u003eCatBoost is an improved gradient boosting decision tree algorithm first proposed by the Yandex research team in 2017 (Prokhorenkova et al. \u003cspan class=\"CitationRef\"\u003e2018\u003c/span\u003e). It solves the problems of prediction offset and classification feature processing in traditional gradient boosting through three technical innovations: ordered boosting strategy (dynamically adjusting sample order to avoid gradient estimation bias), symmetric tree structure (using the same splitting rule for nodes at the same level), and ordered target statistics. The ordered target statistics balances the influence of noise through historical sample statistics and weight coefficient \u0026alpha;, reducing the target leakage risk of category features. These technical innovations enable CatBoost to significantly improve generalization ability and computational efficiency while retaining the high precision advantage of GBDT. The formula (Yang et al. \u003cspan class=\"CitationRef\"\u003e2025\u003c/span\u003e) for calculating the target statistic of a sample is:\u003c/p\u003e\n\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\n\u003cdiv id=\"FileID_Equa\" class=\"mathdisplay\"\u003e$$\\:{\\text{x}}_{\\text{k}}^{\\text{i}}=\\frac{{\\sum\\:}_{{\\text{x}}_{\\text{j}}\\in\\:{\\text{D}}_{\\text{k}}}\\left\\{{\\text{x}}_{\\text{k}}^{\\text{i}}={\\text{x}}_{\\text{j}}^{\\text{i}}\\right\\}{\\text{y}}_{\\text{i}}+{\\alpha\\:}\\text{P}}{{\\sum\\:}_{{\\text{x}}_{\\text{j}\\in\\:{\\text{D}}_{\\text{k}}}}\\left\\{{\\text{x}}_{\\text{k}}^{\\text{i}}={\\text{x}}_{\\text{j}}^{\\text{i}}\\right\\}+{\\alpha\\:}}$$\u003c/div\u003e\n\u003c/div\u003e\n\u003cp\u003eIn the formula, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{x}}_{\\text{k}}^{\\text{i}}\\)\u003c/span\u003e\u003c/span\u003e、\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{x}}_{\\text{j}}\\)\u003c/span\u003e\u003c/span\u003e、\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{x}}_{\\text{j}}^{\\text{i}}\\)\u003c/span\u003e\u003c/span\u003e、\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{y}}_{\\text{i}}\\)\u003c/span\u003e\u003c/span\u003e are all training samples ;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{D}}_{\\text{k}}\\)\u003c/span\u003e\u003c/span\u003e is the data set before the K-th sample ; \u0026alpha; is the weight coefficient ; P is the priori value.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n\u003ch2\u003e2.5 Stacked Integration\u003c/h2\u003e\n\u003cp\u003eThe Stacking Ensemble Method is an ensemble learning method that trains a meta-learner by combining the prediction results of multiple base learners to improve the performance of the model (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e). The base learner can be different types of models (such as Decision Trees, Support Vector Machines, Neural Networks, etc.), and the meta-learner is usually a simple regression or classification model (such as linear regression, logistic regression, etc.). The core idea is to integrate the outputs of multiple base learners by introducing a meta-learner to generate a more powerful prediction model (Wolpert \u003cspan class=\"CitationRef\"\u003e1992\u003c/span\u003e). For the integrated model, the stacking method is used in the study. The previously selected hyper-parametric models (RF, XGBoost, LightGBM, CatBoost) are used as the base models, and six methods such as regression algorithm, RF, GBT, SVM, and MLP are defined as meta-models. The meta-model is evaluated by 5-fold cross-validation, and the optimal meta-model is selected to construct the final stacking model for validation set prediction evaluation. The choice of meta-learner is based on two considerations : 1) Diversity : random forest and SVM are good at dealing with nonlinear features and linear separable problems respectively, and MLP can capture deep feature interaction ; 2) Computational efficiency : logistic regression and gradient boosting trees reduce computational overhead while ensuring accuracy.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n\u003ch2\u003e2.6 Model hyperparameter optimization\u003c/h2\u003e\n\u003cp\u003eIn order to fully exploit the performance of different models, four mainstream hyperparameter optimization strategies (Random Search, Grid Search, Bayesian Optimization and Optuna) are used to systematically optimize each base model. The optimal combination between the hyperparameter optimization strategy and each model is found by comprehensively comparing the evaluation parameters :\u003c/p\u003e\n\u003cp\u003eRandom Search: Randomized search CV (RandomizedSearchCV) is a hyper-parameter optimization method based on random sampling. It efficiently approximates the global optimal solution by combining cross-validation and is suitable for preliminary screening (such as the distribution range of parameters such as learning rate and maximum depth). In the specific implementation, the learning rate is set to a uniform distribution (uniform) of 0.01\u0026ndash;0.3, the maximum depth (3\u0026ndash;50), the number of iterations (100\u0026ndash;1000), the number of trees (100\u0026ndash;500) and the number of leaf nodes (20\u0026ndash;150) are all integer random sampling (randint), and 50 iterations are performed through 5-fold cross-validation to obtain an approximate optimal parameter combination while ensuring search efficiency.\u003c/p\u003e\n\u003cp\u003eGrid Search: Grid search CV (GridSearchCV) is a systematic hyper-parameter optimization method. By exhausting all combinations of predefined parameter spaces, the search grid is refined based on random search results (such as 0.5 / 1 / 1.5 times the benchmark value of the learning rate), and the calculation cost is high but the result is accurate.\u003c/p\u003e\n\u003cp\u003eBayesian Optimization algorithm: Bayesian optimization is a hyper-parameter optimization method based on probability model, which is suitable for scenarios with high computational cost of objective function (Pelikan et al. \u003cspan class=\"CitationRef\"\u003e1999\u003c/span\u003e; Gu et al. \u003cspan class=\"CitationRef\"\u003e2023\u003c/span\u003e). This method approximates the distribution of objective function by constructing Gaussian process surrogate model. The optimization process is divided into two stages : first, 10 random explorations are carried out to initialize the surrogate model, and then 50 iterative optimizations are carried out. In each iteration, the most potential hyper-parameter combination is selected for evaluation through balanced exploration and exploitation of acquisition functions (such as EI or UCB). The optimized objective function is 5-fold cross-validation accuracy. The search scope covers key parameters such as n _ estimators, learning _ rate, max _ depth, iterations, etc.\u003c/p\u003e\n\u003cp\u003eOptuna algorithm: Optuna is an automated hyperparameter optimization framework based on Bayesian optimization, which supports a variety of optimization algorithms (such as TPE, CMA-ES, etc.). The framework is based on Bayesian optimization theory and achieves efficient parameter search through TPE algorithm (Akiba et al. \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e). In each model optimization, the search space of key parameters such as n _ estimators (100\u0026ndash;1000), learning _ rate (0.01\u0026ndash;0.3), max _ depth (3\u0026ndash;50) is set, and 50 iterations are carried out with F1-score as the optimization goal. Optuna significantly reduces the computational cost through asynchronous parallel and early termination, which is especially suitable for complex model optimization in high-dimensional parameter space.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"3 Data Sources and Preprocessing","content":"\u003cp\u003eThe data used in this study are from the publicly available seismic logging data set of the F3 exploration area in the North Sea, covering an area of about 387 km\u003csup\u003e2\u003c/sup\u003e. It is a default demonstration data set provided by the Opendtect software. The block is composed of a large river delta system sedimentary system. The delta inclusions are composed of sandstone and mudstone, and the porosity is generally high (20%-33%). The Inline range of the selected work area for this test is 100\u0026thinsp;~\u0026thinsp;700, the Crossline range is 300\u0026thinsp;~\u0026thinsp;1250, the time range is 408\u0026thinsp;~\u0026thinsp;1136 ms, the track spacing is 25 meters, and the sampling interval is 4 ms. Wells F02-1, F03-2, F03-4 and F06-1 in the work area were selected, which contain three-dimensional post-stack seismic data, longitudinal wave impedance data of seismic inversion, natural gamma data of geostatistical inversion, and lithology data of four wells (including sandstone and mudstone).\u003c/p\u003e\n\u003cp\u003eSeismic data profile of Well F06-1 in the exploration area (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e). The red series (+\u0026thinsp;2000 to +\u0026thinsp;8000) corresponds to positive phase reflection waves, revealing high wave impedance interfaces such as sandstone; the blue series (-2000 to -8000) represents negative phase reflections, which mostly indicate shale, mudstone or gas-bearing strata, and here indicates mudstone; the amplitude intensity is presented through the color saturation gradient, among which the yellow highlight band (Inline 400\u0026ndash;600, 750-950ms) shows abnormally strong reflections, which may correspond to unconformity surfaces or the top boundaries of oil and gas reservoirs; the vertical black dotted lines (Inline 244, Inline 550) are typical stratigraphic tracing lines, showing that the amplitude polarity of the T3 reflection layer is reversed at 950 ms, suggesting a sudden change in stratigraphic lithology; a red-blue transition zone appears near 1000ms, reflecting the stratigraphic contact relationship of the sedimentary sequence transitioning from sandstone to mudstone; the oblique amplitude structure (Inline 300\u0026ndash;500, 800-1000ms) has a progradational reflection configuration, which may be the seismic response of the delta front sedimentary body; the amplitude difference rate of its top and bottom interfaces is as high as 62%, and the amplitude variation with offset (AVO) characteristics show a Class III anomaly at Inline 450, which is consistent with the theoretical response model of gas-bearing sandstone.\u003c/p\u003e\n\u003cp\u003eThe preprocessing steps for the test area data are as follows:\u003c/p\u003e\n\u003col\u003e\n\u003cli\u003e\n\u003cp\u003eResampling: First, read the time-depth relationship between the well point lithology data and the well data, and remove the outliers in the lithology data. Then perform time-depth conversion of the lithology data and resample according to the sampling rate of the seismic data (4ms). Finally, only the data retained within the time range of the geophysical data and located in the target layer are intercepted as lithology sample labels. After processing, a total of 352 lithology samples of two types are obtained, including 193 mudstones and 159 sandstones.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eClass balance: In order to further improve the robustness and recognition accuracy of the model and enhance the generalization performance, this study uses the nearest neighbor removal algorithm (NM) combined with the synthetic minority class oversampling algorithm (SMOTE) to form the NM-SMOTE method (Chawla et al. \u003cspan class=\"CitationRef\"\u003e2002\u003c/span\u003e; Wang et al. \u003cspan class=\"CitationRef\"\u003e2021\u003c/span\u003e) to resample the samples. The processing results are shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e. On the one hand, this method increases the minority class samples (sandstone) through SMOTE, and on the other hand, it uses NM to remove noise samples to achieve the optimal balance of the data set (mudstone: sandstone\u0026thinsp;=\u0026thinsp;193:193).\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eFeature optimization: In the process of extracting lithological sample features, based on well logging curves and seismic attribute data, parameters with significant correlation with lithological classification, such as natural gamma (GR), longitudinal impedance (IP), and instantaneous frequency, were preliminarily screened out (Guyon et al. 2003). Then, recursive feature elimination (RFE) with a step size of 1 (one feature is eliminated in each iteration) and random forest were used as the base model. The feature weights were calculated using the Gini coefficient. The scoring criteria were balanced accuracy and combined with five-fold cross-validation for secondary screening. By gradually eliminating the features with the lowest contribution, the optimal number of features was determined to be 8. Finally, the lithological data of the four wells were used as lithological sample labels, and the well bypass data of the extracted lithological sample features were matched with them to form a lithological sample set.\u003c/p\u003e\n\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eThe analysis based on recursive feature elimination cross-validation (RFECV) showed that there was a significant correlation between the number of features and model performance. As shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e, when the number of features was 8, the model achieved the best performance, with a balanced accuracy of 0.664\u0026thinsp;\u0026plusmn;\u0026thinsp;0.068 (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation), which had high classification performance and good stability. From the perspective of the impact of the number of features on model performance, when the number of features increased from 6 to 8, the model accuracy increased from 0.650 to 0.664, indicating that the two newly added features made a significant contribution to lithology discrimination; when the number of features exceeded 8, the model accuracy decreased by 0.012, indicating that the additional features may lead to overfitting and reduce the generalization ability of the model. In terms of stability, the model was the most unstable when there were 3 features (standard deviation 0.089), while the standard deviation dropped to 0.068 when there were 8 features, indicating that the model was the most robust to data fluctuations under this configuration. After comprehensive consideration of performance and stability, 8 features were finally determined as the optimal choice, and they were ranked in order of importance: natural gamma (GR), instantaneous amplitude, instantaneous frequency, longitudinal wave impedance (IP), root mean square energy, 24 Hz frequency division energy, 44 Hz frequency division energy, and 64 Hz frequency division energy. This feature combination effectively avoids the risk of overfitting while ensuring the accuracy of the model.\u003c/p\u003e"},{"header":"4 Modeling and evaluation","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Evaluation indicators\u003c/h2\u003e\u003cp\u003eThis study constructs a multi-dimensional evaluation system for systematic analysis. Firstly, the core indicators such as Accuracy, Recall, F1 score (F1-Score) and area under the receiver operating characteristic curve (AUC) were selected to comprehensively evaluate the classification ability of the model in different scenarios. In order to ensure the reliability of the results, 5-fold cross-validation was used. Each iteration was randomly divided into training set and verification set. Finally, the average value of each index was taken as the evaluation benchmark of the model performance.\u003c/p\u003e\u003cp\u003eIn order to visually display the performance of the model, three visualization methods are used : 1) box plot is used to display the distribution and stability of the accuracy of the model. The box median reflects the average level of the model, the interquartile range ( IQR ) reflects the fluctuation range, the line is used to mark the extreme value, and the outliers reveal the abnormal performance of the model. 2) The bar chart is used to compare the average accuracy between models, and the error line represents the standard deviation ; 3) The line chart is used to track the trend of accuracy with iteration in the process of hyperparameter optimization. In addition, in order to quantify the performance difference between models, the relative improvement rate index is introduced, and the calculation formula is as follows :\u003c/p\u003e\u003cp\u003eThe improvement rate = ( optimization model accuracy-base model accuracy ) / base model accuracy \u0026times; 100%.\u003c/p\u003e\u003cp\u003eThis evaluation scheme combines quantitative indicators and multi-dimensional visualization methods, which can effectively avoid the misleading caused by the contingency of single indicator or data division, and enhance the credibility and explanatory power of the results.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e4.2 Comparative analysis of hyperparameter optimization\u003c/h2\u003e\u003cp\u003eThrough five-fold cross-validation (30% of the test set, three independent repeated experiments), this section compares the performance of four mainstream hyperparameter optimization methods (Grid Search, Random Search, Bayesian Optimization, Optuna) in different ensemble models. The main evaluation dimensions are : cross-validation accuracy, test set accuracy, training time, and relative improvement rate.\u003c/p\u003e\u003cp\u003eIn the Random Forest Model, Bayesian Search showed the best performance in the cross-validation stage. Its accuracy on the test set (Test Accuracy\u0026thinsp;=\u0026thinsp;0.812) was significantly better than other methods (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05), and its cross-validation accuracy (CV Accuracy\u0026thinsp;=\u0026thinsp;0.797) was also the highest value (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003ea). The analysis of Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eb further shows that the prediction results of Bayesian Search are the most stable (interquartile range IQR\u0026thinsp;=\u0026thinsp;0.020), and the median and mean are highly consistent, indicating that its parameter search path has a low variance characteristic. In contrast, the IQR of random search and grid search are both 0.040, and the distribution is obviously skewed, reflecting the inefficiency of their search strategies and the instability of their results. This result verifies the effectiveness of Bayesian Search in dynamically balancing exploration and utilization capabilities through surrogate models (such as Gaussian processes). Its mechanism of iterative optimization of the search space based on historical feedback can significantly reduce the attempts of invalid parameter combinations. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes the training time and improvement rate of each method. Bayesian search achieved an improvement rate of 6.56% with a training time of 86.32 seconds, with a significant efficiency advantage (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.01), while Optuna ranked second with an accuracy of 0.805. Its ratio of improvement rate (5.64%) to training time (101.13 seconds) (0.056%\u0026middot;s⁻\u0026sup1;) was significantly better than Bayesian search's 0.076%\u0026middot;s⁻\u0026sup1;, indicating the advantage of dynamic resource allocation strategy in the efficiency-accuracy trade-off. In contrast, although the grid search improved the improvement rate by 2.10%, it took as long as 150.40 seconds (an increase of 74%), verifying the inefficiency of the exhaustive strategy in high-dimensional parameter space. It is worth noting that the Bayesian range refinement and feature optimization variant has a training time of more than 130 seconds due to the shrinking of parameter range and the ranking of feature importance, but the improvement rate is only 4.72%~5.00%, which has not surpassed the basic version. This phenomenon shows that excessively restricting the search space or adding feature screening processes may weaken the global exploration ability of the Bayesian method. Based on the above results, Bayesian search is the preferred solution for parameter tuning because it has a balanced performance in accuracy, stability and efficiency, especially in high-dimensional non-convex optimization problems.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComprehensive Performance of Random Forest Hyperparameter Optimization\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMethod\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCV mean\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003ecv_std\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest accuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eTraining time(s)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eImprovement rate(%)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRandom Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.785\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.033\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.762\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e40.81s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.00%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrid Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.790\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.033\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.778\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e150.40s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e2.10%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBayesian Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.797\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.027\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.812\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e86.32s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e6.56%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBayesian Search Range Refinement\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.793\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.039\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.798\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e130.34s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e4.72%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBayesian Search Feature Optimization\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.793\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.039\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.800\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e131.52s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e5.00%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOptuna\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.797\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.031\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.805\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e101.13s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e5.64%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn the XGBoost model, the experimental results (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) show that Optuna is significantly superior to other methods in terms of test accuracy (0.778) and improvement rate (4.29%), and the training time (10.32 seconds) is only 1.5% of Grid search (697.13 seconds), reflecting its best balance between model performance and computational efficiency. Although the cross-validation mean of Optuna (0.770) is slightly lower than that of Grid search (0.775) (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003ea), its test accuracy is higher and the standard deviation (cv _ std\u0026thinsp;=\u0026thinsp;0.036) is the smallest, indicating that its optimization strategy can effectively alleviate the risk of overfitting. In contrast, Grid search is easy to fall into local optimal solution due to exhaustive parameter traversal, resulting in performance degradation in the test phase (test accuracy 0.762), and the standard deviation is large (cv _ std\u0026thinsp;=\u0026thinsp;0.042). This result is consistent with the research of Bergstra et al. (Bergstra et al. 2012), that is, systematic search may not necessarily improve generalization ability, but may reduce stability due to high variance parameter combinations. Optuna 's training time is significantly shorter than other methods because of the synergistic effect of its TPE algorithm and early stopping strategy ( Pruning ). It prioritizes the exploration of high-potential parameter regions through probability models, and eliminates inefficient candidate schemes, thereby reducing redundant calculations. In contrast, the time cost of Grid search increases exponentially with the parameter dimension (such as the full combination traversal of hyperparameters such as learning rate and tree depth), resulting in a time of 11.6 minutes in this experiment, which is difficult to meet the real-time requirements in large-scale data scenarios. From the stability of the results (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003eb), the stability of Grid search is the best (IQR\u0026thinsp;=\u0026thinsp;0.024), and the data distribution is concentrated. The Optuna model has the largest IQR value (0.074), reflecting that the optimization process is sensitive to the initial parameters. From the perspective of improvement rate, Optuna has a performance improvement of 4.29% compared with Random search (benchmark method), which is significantly higher than Bayesian search (1.21%) and Grid search (2.15%). This result verifies the superiority of Optuna in the efficiency of parameter space exploration. By dynamically adjusting the search direction through the reinforcement learning framework, Optuna avoids the dependence of Bayesian methods on prior distribution (such as the risk of suboptimal solutions when the Gaussian process hypothesis does not hold), and overcomes the blindness of Random search.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComprehensive performance of XGBoost hyperparameter optimization\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMethod\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCV mean\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003ecv_std\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest accuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eTraining time(s)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eImprovement rate(%)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRandom Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.757\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.058\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.746\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e4.87s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.00%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrid Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.775\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.042\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.762\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e697.13s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e2.15%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBayesian Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.765\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.045\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.755\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e21.75s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1.21%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOptuna\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.770\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.036\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.778\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e10.32s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e4.29%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn the LightGBM model, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003ea, the Optuna framework is significantly superior to other methods in terms of test accuracy (0.778) and cross-validation stability (cv _ std\u0026thinsp;=\u0026thinsp;0.031), with an increase of 8.96% (p\u0026thinsp;\u0026lt;\u0026thinsp;0.01), and the boxplot (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eb) shows that the distribution of test results is the most concentrated (IQR\u0026thinsp;=\u0026thinsp;0.025), indicating that the model generalization ability is statistically robust. In contrast, although Grid Search achieves 6.44% performance improvement, the training time is 82.67 seconds, and the computational cost is more than 50 times that of Optuna (1.64 seconds), which verifies the inefficiency of exhaustive search strategy in high-dimensional parameter space. The Random Search and the original model (Ori _ best _ lgbm) did not achieve significant improvement (p\u0026thinsp;\u0026gt;\u0026thinsp;0.05). The test accuracy was stable at the level of 0.714, and the cross-validation fluctuation range was 0.073 ( cv _ std ), which further explained the limitations of no guiding parameter exploration. The training time of Optuna is only 1.64 seconds (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e), the efficiency is 9.9 times (16.20 seconds) of the original model, and it is 98% faster than the Grid Search (82.67 seconds). It is recommended as the preferred solution for high-dimensional parameter optimization tasks.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComprehensive performance of LightGBM hyperparameter optimization\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMethod\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCV mean\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003ecv_std\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest accuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eTraining time(s)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eImprovement rate(%)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOri_best_lgbm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.757\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.073\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.714\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e16.20s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.0%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRandom Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.785\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.051\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.714\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1.93s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.0%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrid Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.770\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.048\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.760\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e82.67s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e6.44%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBayesian Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.785\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.034\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.750\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e6.76s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e5.04%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOptuna\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.780\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.031\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.778\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1.64s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e8.96%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIt can be seen from Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e and Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e that in the CatBoost model, Optuna is superior to other methods in terms of test accuracy (0.810), improvement rate (6.3%) and IQR value. At the same time, the training time is the shortest (65.47s), and the CV standard deviation is the lowest (0.034), indicating that it has high efficiency and stability. Bayesian optimization further improves the convergence speed by predicting the potential benefits of parameter combinations through probability density functions, which is consistent with the conclusion of Snoek et al. on the efficiency of Bayesian optimization in black box function optimization(Snoek et al. \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2012\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eBayesian search is second only to Optuna in test accuracy (0.785) and improvement rate (3.02%), but its training time (158.28s) increases significantly. This is because Bayesian optimization needs to repeatedly update the Gaussian process model to predict parameter performance, and the computational complexity increases with the number of iterations. In contrast, Optuna 's TPE algorithm reduces the model update cost through a hierarchical sampling strategy, which is more efficient. The grid search series and random search perform poorly, the test accuracy is not more than 0.77, and the training time is generally more than 200 s.\u003c/p\u003e\u003cp\u003eAlthough the exhaustive strategy of grid search can ensure the traversal of the local optimal solution, its time cost increases exponentially with the parameter dimension, and the search method of fixed step size is easy to miss the optimal interval. Dense sampling enhances the sensitivity of the model to training data noise. Similar conclusions have also been mentioned in the study of Hastie et al. (Hastie et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2004\u003c/span\u003e), that is, over-reliance on mesh accuracy may lead to the sacrifice of generalization of the model in exchange for local accuracy improvement.\u003c/p\u003e\u003cp\u003eThe non-directional sampling of random search leads to a large number of invalid attempts. Although its theoretical convergence is better than that of grid search, it is difficult to approach the global optimum under the finite number of iterations. This finding is consistent with the conclusion of Hutter and Bergstra et al. (Hutter et al. \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2011\u003c/span\u003e; Bergstra et al. 2012) on hyperparameter optimization methods, that is, random search lacks guidance in high-dimensional space and is inefficient.\u003c/p\u003e\u003cp\u003eFor engineering scenarios that require fast iteration (such as online parameter tuning), Optuna 's high efficiency makes it the first choice. The training time is only 65.47 s, which is more than 3 times faster than Grid Search, and there is no need to sacrifice accuracy. This advantage is particularly critical in industrial applications with high real-time requirements. If the absolute accuracy is the only goal and the computing resources are sufficient, Bayesian search can be considered. Although its training takes a long time, the fineness of its parameter optimization may bring a weak but critical performance improvement in specific tasks.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComprehensive performance of CatBoost hyperparameter optimization\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMethod\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCV mean\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003ecv_std\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest accuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eTraining time(s)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eImprovement rate(%)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRandom Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.785\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.053\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.762\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e244.34s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.00%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrid Search Best Parameters\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.765\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.038\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.778\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e200.58s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e2.36%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrid Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.789\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.054\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.770\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e213.05s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1.05%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrid Search Small Range\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.790\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.054\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.762\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e215.22s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.00%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBayesian Search\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.041\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.785\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e158.28s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e3.02%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOptuna\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.805\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.034\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.810\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e65.47s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e6.3%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e4.3 Comparison and evaluation of model performance\u003c/h2\u003e\u003cp\u003eBased on the accuracy, F1 score and AUC index, combined with the systematic comparison of ROC curve and precision-recall curve, this study evaluated the performance differences of five mainstream machine learning models (Random Forest, XGBoost, LightGBM, CatBoost, Stacking Model) in classification tasks. The results show that CatBoost performs best on the AUC value (0.91) and the precision under the condition of low recall rate (Recall0.95) (Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e). It has a small IQR (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003ea). All the verification results are located in the range of 0.785\u0026ndash;0.825, indicating that it has good generalization ability for data and can be stably played on multiple different training sets. The model incorporates a built-in categorical feature optimization mechanism (e.g., dynamic One-Hot encoding and the ordered lifting algorithm), which significantly enhances overfitting resistance and accuracy in small-sample scenarios. This makes it particularly suitable for applications requiring high sensitivity to missed detections and stable output performance.\u003c/p\u003e\u003cp\u003eBy integrating the advantages of multi-base models, the Stacking Model approaches CatBoost at the AUC (0.90) and accuracy (0.77) levels, showing a balanced bias-variance control ability, but its training time (1200 seconds) is significantly higher than other models (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003eb). In contrast, although LightGBM and XGBoost are superior in computational efficiency (training time is about 800 seconds and 100 seconds, respectively), their AUC values are 0.82 and 0.79, respectively, and the accuracy drops sharply under high recall conditions (Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003eb), which requires targeted parameter adjustment to improve generalization ability.\u003c/p\u003e\u003cp\u003eRandom forest achieved a balance between stability (IQR\u0026thinsp;=\u0026thinsp;0.06) and AUC (0.88) (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003ea, Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e), but did not break through the performance boundary of single ensemble learning. In summary, if the accuracy and robustness are the core requirements, CatBoost is preferred ; if you need to balance accuracy and resource constraints, you can choose the Stacking Model or the optimized LightGBM. The research results provide a quantitative basis for model selection in complex scenarios. The follow-up work needs to further explore the joint optimization strategy of model lightweight and adversarial example defense.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFrom Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003ea (ROC curve analysis), we can see that the performance ranking of each model is CatBoost (AUC\u0026thinsp;=\u0026thinsp;0.91) with the best performance, and the curve is closest to the upper left corner. When FPR\u0026thinsp;=\u0026thinsp;0.2, TPR has reached 0.8, and when FPR\u0026thinsp;=\u0026thinsp;0.3, TPR\u0026thinsp;=\u0026thinsp;0.85, showing strong early recognition ability, and significantly better than other models when FPR\u0026thinsp;\u0026lt;\u0026thinsp;0.4; the stacking model (AUC\u0026thinsp;=\u0026thinsp;0.90) follows closely, and when FPR\u0026thinsp;\u0026gt;\u0026thinsp;0.5, it performs better than CatBoost, and when FPR\u0026thinsp;=\u0026thinsp;0.2, TPR is about 0.8, which is suitable for medium risk tolerance scenarios; Random Forest (AUC\u0026thinsp;=\u0026thinsp;0.88) performs stably but slightly worse than CatBoost and the stacking model, LightGBM (AUC\u0026thinsp;=\u0026thinsp;0.82) and XGBoost (AUC\u0026thinsp;=\u0026thinsp;0.79) are relatively weak, and the XGBoost curve is closest to the diagonal, indicating that its discrimination ability is relatively limited.\u003c/p\u003e\u003cp\u003eFrom Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003eb (P-R curve analysis), it can be seen that in the performance characteristics, CatBoost maintains the highest curve position, Precision remains above 0.9 when Recall\u0026thinsp;=\u0026thinsp;0.5, and Precision drops rapidly to about 0.78 after Recall\u0026thinsp;\u0026gt;\u0026thinsp;0.7, which is suitable for the accuracy scenario with Recall requirement\u0026thinsp;\u0026lt;\u0026thinsp;0.7. The precision of the stacking model drops rapidly from its peak value to 0.5, indicating that under extremely low recall conditions(Recall), the model can only identify the most easily classifiable positive samples (such as sandstone with simple lithology and obvious logging response). The accuracy increases from 0.5 to 0.95, indicating that as the recall rate increases, the model gradually covers more ' medium-difficulty ' positive samples (such as pebbly sandstone or argillaceous sandstone). The feature discrimination of these samples is good, and the model can maintain high accuracy stably, reaching a peak accuracy (0.95) at Recall\u0026thinsp;=\u0026thinsp;0.57. It shows that the model has a strong ability to identify most common lithologies (such as sandstone and mudstone) and meets the actual needs.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e4.4 Lithology prediction results\u003c/h2\u003e\u003cp\u003eF02-1, F06-1, F03-2 and F03-4 are used as verification wells in turn, and the lithology samples of the other three wells are used for model training. In order to facilitate the comparison of the prediction results of each method, the prediction results of each method are listed here when F06-1 is used as the verification well. Figures\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e14\u003c/span\u003e,\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003e,\u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e16\u003c/span\u003e,\u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003e and \u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003e are the formation lithology and sandstone probability of the horizontal survey line of F06-1 well predicted by the five models. Figures\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e14\u003c/span\u003ea, \u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003ea, \u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e16\u003c/span\u003ea, \u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003ea and \u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003ea are the model prediction results of the constructed random forest method, XGBoost method, LightGBM method, CatBoost method and stacked ensemble method, respectively. It can be seen that the prediction results of lithology by each method have a certain correspondence with the lithology of the verification wells used, but the prediction accuracy and resolution of different methods are slightly different.\u003c/p\u003e\u003cp\u003eAmong them, CatBoost is more precise in the identification of sand and mudstone. The probability of sandstone gradually changes along the Inline, which is in line with the geological law, and the lithology continuity is better (Fig.\u0026nbsp;\u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003ea). The probability profile of sandstone is stable (Fig.\u0026nbsp;\u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003eb). The probability value shows a more gentle gradient along the Inline, and there is no mutation in the sand-mud transition zone, indicating that the model is more robust to noise or complex features. The Stacking model demonstrates good performance on easily classifiable samples (e.g., thick sandstone layers or mudstone), while its effectiveness in classifying complex samples (e.g., thin interbedded layers) is limited (Fig.\u0026nbsp;\u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003e). The random forest prediction boundary is blurred, the sandstone probability profile jumps obviously (as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e14\u003c/span\u003eb, the shallow lateral color jumps obviously), and the shallow sandstone recognition ability is weak. The LightGBM thick sandstone prediction is continuous (Time 900-1100ms), but the shallow layer (Time\u0026thinsp;\u0026lt;\u0026thinsp;900ms) is not continuous (the classification near the fault in Fig.\u0026nbsp;\u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e16\u003c/span\u003ea is blurred). XGBoost has a strong ability to capture details, but as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003e, it is sensitive to data noise, the continuity of lithology prediction is not strong, and the overall accuracy is the lowest.\u003c/p\u003e\u003cp\u003eThe classification algorithm of each model is feasible in seismic reservoir classification. At the well location, the actual data is consistent with the classification prediction results. On the other hand, the classification results far away from the well position outline the corresponding response of the stratigraphic morphology to the seismic events on the seismic profile (Wang et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). It can be seen that the prediction results have a good correspondence with the lithology of the verification well, and the sandstone probability profile of CatBoost has the highest degree of agreement with the actual lithology (Fig.\u0026nbsp;\u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003e), especially in the target layer (depth 700\u0026ndash;1136 meters). The accuracy of sandstone identification is more than 75%. Figures\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e14\u003c/span\u003eb, \u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003eb, \u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e16\u003c/span\u003eb, \u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003eb and \u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003eb are the sandstone probability profiles of each method, which quantitatively reflect the possibility that the formation lithology in the target interval belongs to sandstone.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eThis study compares and analyzes the performance of four hyperparameter optimization methods and five ensemble models in lithology identification, and draws the following conclusions:\u003c/p\u003e\u003cp\u003e(1) In terms of optimization methods, the Optuna framework achieves the best balance between efficiency and accuracy, and CatBoost achieves an accuracy of 0.810 in only 65.47 seconds after optimization. Random search demonstrates the fastest execution speed, and XGBoost optimization takes only 4.87 seconds. Bayesian optimization shows the best stability.\u003c/p\u003e\u003cp\u003e(2) In terms of model performance, CatBoost achieves the highest AUC (0.91), excelling in complex lithology identification.The stacked ensemble model exhibits superior robustness (5-fold CV standard deviation: 0.041). Random forest training is the most efficient (average 53 seconds). LightGBM and XGBoost are suitable for fast response and standard accuracy scenarios, respectively.\u003c/p\u003e\u003cp\u003e(3) Furthermore, our findings confirm that combining resampling techniques with feature selection significantly improves the robustness and adaptability of machine learning models in seismic reservoir lithology identification. This is consistent with recent studies highlighting the synergy of these two strategies for handling imbalanced geological data.\u003c/p\u003e\u003cp\u003eThe research results provide an important model selection basis for intelligent lithology identification in oil and gas exploration.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConflict of interest\u003c/strong\u003e\u003cp\u003eAll authors acknowledged that there is no conflict of interest on record.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e\u003cp\u003eChina National Nuclear Corporation-State Key Laboratory of Nuclear Resources and Environment (East China University of Technology) Joint Innovation Fund Project (No. 2023NRE-LH-08); National Natural Science Foundation of China (Nos. U2244205, U2067202).\u003c/p\u003e\u003ch2\u003eAuthor contributions\u003c/h2\u003e\u003cp\u003eAll authors listed made a substantial, direct, and intellectual contribution to the work and approved it for publication\u003c/p\u003e\u003ch2\u003eData Availability Statement\u003c/h2\u003e\u003cp\u003eThe seismic data supporting this study are publicly available in the Terranubis open data repository at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://terranubis.com/datainfo/F3-Demo-2023\u003c/span\u003e\u003cspan address=\"https://terranubis.com/datainfo/F3-Demo-2023\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Processed datasets and Python scripts generated during this research are available from the corresponding author upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAkiba T, Sano S, Yanase T et al (2019) Optuna: A next-generation hyperparameter optimization framework. Knowl Discovery Data Min 2623\u0026ndash;2631. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1145/3292500.3330701\u003c/span\u003e\u003cspan address=\"10.1145/3292500.3330701\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAli M, Changxingyue H, Wei N et al (2025) Optimizing seismic-based reservoir property prediction: a synthetic data-driven approach using convolutional neural networks and transfer learning with real data integration. Artif Intell Rev 58:31. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10462-024-11030-8\u003c/span\u003e\u003cspan address=\"10.1007/s10462-024-11030-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAmiri-Ramsheh B, Larestani A, Atashrouz S et al (2025) Toward accurate prediction of carbon dioxide (CO2) compressibility factor using tree-based intelligent schemes (XGBoost and LightGBM) and equations of state. Results Eng Volume 25 104035, ISSN 2590\u0026thinsp;\u0026ndash;\u0026thinsp;1230. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.rineng.2025.104035\u003c/span\u003e\u003cspan address=\"10.1016/j.rineng.2025.104035\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBergstra J, Bengio Y (2012) Random search for hyper-parameter optimization. J Mach Learn Res 13(1):281\u0026ndash;305. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.chemolab.2011.12.002\u003c/span\u003e\u003cspan address=\"10.1016/j.chemolab.2011.12.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBreiman L (2001) Random forests. Mach Learn 45:5\u0026ndash;32. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1023/A:1010933404324\u003c/span\u003e\u003cspan address=\"10.1023/A:1010933404324\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChawla NV, Bowyer KW, Hall LO et al (2002) SMOTE: Synthetic Minority Over-sampling Technique. J Artif Intell Res 16(1):321\u0026ndash;357. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1613/jair.953\u003c/span\u003e\u003cspan address=\"10.1613/jair.953\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCheng X, Zhou J, Fu H et al (2023) Applicability and Application of Machine Learning Algorithm in Logging Interpretation. Northwest Geol 56(4):336\u0026ndash;348. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.12401/j.nwg.2023029\u003c/span\u003e\u003cspan address=\"10.12401/j.nwg.2023029\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen Y, Zhao L, Pan J et al (2021) Deep carbonate reservoir characterization using multi-seismic attributes via machine learning with physical constraints. J Geophys Eng 18(5):761\u0026ndash;775. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/jge/gxab049\u003c/span\u003e\u003cspan address=\"10.1093/jge/gxab049\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen T, Guestrin C (2016) XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '16). Association for Computing Machinery, New York, NY, USA, 785\u0026ndash;794. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1145/2939672.2939785\u003c/span\u003e\u003cspan address=\"10.1145/2939672.2939785\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGuyon I, Elisseeff A (2003) An introduction to variable and feature selection. J Mach Learn Res 3(7\u0026ndash;8):1157\u0026ndash;1182. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1162/153244303322753616\u003c/span\u003e\u003cspan address=\"10.1162/153244303322753616\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGu Y, Zhang S, Wang W et al (2023) Lithofacies prediction driven by logging-based Bayesian-optimized ensemble learning: a case study of lacustrine carbonate reservoirs. Geophys Prospect 71:1835\u0026ndash;1872. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/1365-2478.13258\u003c/span\u003e\u003cspan address=\"10.1111/1365-2478.13258\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGao S, Xu M, Zhao L et al (2023) Seismic predictions of fluids via supervised deep learning: Incorporating various class-rebalance strategies. Geophysics 88(4):M185\u0026ndash;M200. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1190/geo2022-0363.1\u003c/span\u003e\u003cspan address=\"10.1190/geo2022-0363.1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHe H, Bai Y, Garcia EA et al (2008) ADASYN: Adaptive Synthetic Sampling Approach for Imbalanced Learning. Proceedings of the International Joint Conference on Neural Networks 1322\u0026ndash;1328. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/IJCNN.2008.4633969\u003c/span\u003e\u003cspan address=\"10.1109/IJCNN.2008.4633969\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHastie T, Tibshirani R, Friedman JH et al (2004) The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Math Intelligencer 27:83\u0026ndash;85. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/BF02985802\u003c/span\u003e\u003cspan address=\"10.1007/BF02985802\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHutter F, Hoos HH, Leyton-Brown K (2011) Sequential model-based optimization for general algorithm configuration. In Proceedings of the 5th international conference on Learning and Intelligent Optimization (LION'05). Springer-Verlag, Berlin, Heidelberg 507\u0026ndash;523. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/978-3-642-25566-3_40\u003c/span\u003e\u003cspan address=\"10.1007/978-3-642-25566-3_40\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHossain S (2020) Application of seismic attribute analysis in fluvial seismic geomorphology. J Petroleum Explor Prod Technol 10:1009\u0026ndash;1019. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s13202-019-00809-z\u003c/span\u003e\u003cspan address=\"10.1007/s13202-019-00809-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJacobs RA, Jordan MI, Nowlan SJ et al (1991) Adaptive mixtures of local experts. Neural Comput 3(1):79\u0026ndash;87. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1162/neco.1991.3.1.79\u003c/span\u003e\u003cspan address=\"10.1162/neco.1991.3.1.79\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKouassi A, Pan L, Wang X et al (2023) Identification of karst cavities from 2D seismic wave impedance images based on gradient-boosting decision trees algorithms (GBDT): Case of Ordovician fracture-vuggy carbonate reservoir. Energies 16(2):643. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/en16020643\u003c/span\u003e\u003cspan address=\"10.3390/en16020643\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKe G, Meng Q, Finley T et al (2017) LightGBM: A highly efficient gradient boosting decision tree. Neural Inform Process Syst 30:3146\u0026ndash;3154\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi L, Sharma A, Cheng S (2025) Explainable AI based LightGBM prediction model to predict default borrower in social lending platform. Intell Syst Appl Volume 26:2667\u0026ndash;3053. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.iswa.2025.200514\u003c/span\u003e\u003cspan address=\"10.1016/j.iswa.2025.200514\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi C, Wang L, Chen Y et al (2023) Prediction Model of Soils' Preconsolidation Pressure Based on Bayesian Ensemble Learning Algorithm. Earth Sci 48(5):1780\u0026ndash;1792. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3799/dqkx.2022.450\u003c/span\u003e\u003cspan address=\"10.3799/dqkx.2022.450\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(in Chinese)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNaveed S, Husnain M (2025) A drug recommendation system based on response prediction: Integrating gene expression and K-mer fragmentation of drug SMILES using LightGBM. Intelligence-Based Med Volume 11:2666\u0026ndash;5212. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ibmed.2025.100206\u003c/span\u003e\u003cspan address=\"10.1016/j.ibmed.2025.100206\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eProkhorenkova L, Gusev G, Vorobev A et al (2018) CatBoost: Unbiased boosting with categorical features. Neural Inform Process Syst 31:6638\u0026ndash;6648. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.1706.09516\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.1706.09516\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePelikan M, Goldberg DE, Cant\u0026uacute;-Paz E (1999) BOA: the Bayesian optimization algorithm. Genetic and Evolutionary Computation Conference 525\u0026ndash;532. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ci.nii.ac.jp/naid/10012961429\u003c/span\u003e\u003cspan address=\"https://ci.nii.ac.jp/naid/10012961429\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSakoe H, Chiba S (1978) Dynamic programming algorithm optimization for spoken word recognition. IEEE Trans Acoust Speech Signal Process 26(1):43\u0026ndash;49. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/TASSP.1978.1163055\u003c/span\u003e\u003cspan address=\"10.1109/TASSP.1978.1163055\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSnoek J, Larochelle H, Adams RP (2012) Practical bayesian optimization of machine learning algorithms. Neural Inform Process Syst 4. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.1206.2944\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.1206.2944\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSaleh K, Mabrouk WM, Metwally A (2025) Machine learning model optimization for compressional sonic log prediction using well logs in Shahd SE field, Western Desert, Egypt. Sci Rep 15:14957. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-025-97938-9\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-97938-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSoulaimani S, Soulaimani A, Abdelrahman K et al (2024) Advanced machine learning artificial neural network classifier for lithology identification using Bayesian optimization. Front Earth Sci 12:1473325. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/feart.2024.1473325\u003c/span\u003e\u003cspan address=\"10.3389/feart.2024.1473325\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTop\u0026oacute;r T, Sowiżdżał K (2022) Application of machine learning tools for seismic reservoir characterization study of porosity and saturation type. Nafta-Gaz 3:165\u0026ndash;175. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.18668/NG.2022.03.01\u003c/span\u003e\u003cspan address=\"10.18668/NG.2022.03.01\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTang Y, Yang J, You Y et al (2024) Multi-output prediction for TBM operation parameters based on stacking ensemble algorithm. Tunn Undergr Space Technol Volume 152:0886\u0026ndash;7798. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.tust.2024.105960\u003c/span\u003e\u003cspan address=\"10.1016/j.tust.2024.105960\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang M, Yang J, Wang X et al (2023) Identification of Shale Lithofacies by Well Logs Based on Random Forest Algorithm. Earth Sci 48(1):130\u0026ndash;142. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3799/dqkx.2022.181\u003c/span\u003e\u003cspan address=\"10.3799/dqkx.2022.181\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(in Chinese)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang W, Lin C, Wang J et al (2009) Pre-stack elastic parameters inversion and its application. Geophys Prospect Petroleum 48(5):483\u0026ndash;487. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3969/j.issn.1000-1441.2009.05.010\u003c/span\u003e\u003cspan address=\"10.3969/j.issn.1000-1441.2009.05.010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(in Chinese)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWolpert DH (1992) Stacked generalization. Neural networks 5(2):241\u0026ndash;259. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/S0893-6080(05)80023-1\u003c/span\u003e\u003cspan address=\"10.1016/S0893-6080(05)80023-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWu X, Liang L, Shi Y et al (2019) FaultSeg3D: Using synthetic data sets to train an end-to-end convolutional neural network for 3D seismic fault segmentation. Geophysics 84(3):IM35\u0026ndash;IM45. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1190/geo2018-0646.1\u003c/span\u003e\u003cspan address=\"10.1190/geo2018-0646.1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWu J, Shi Y, Wang K et al (2023) Automatic seismic fault identification based on an improved U-Net network. Acta Geophys. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11600-023-01200-7\u003c/span\u003e\u003cspan address=\"10.1007/s11600-023-01200-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang G, Song J, Xu F et al (2021) Random Forests lithology prediction method for imbalanced data sets. Oil Geophys Prospect 56(4):679\u0026ndash;687. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.13810/j.cnki.issn.1000-7210.2021.04.001\u003c/span\u003e\u003cspan address=\"10.13810/j.cnki.issn.1000-7210.2021.04.001\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(in Chinese)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYu H, Wang H, Wu L et al (2025) A precise time\u0026ndash;depth conversion method for coal seam based on machine learning and seismic velocity inversion. Acta Geophys. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11600-025-01592-8\u003c/span\u003e\u003cspan address=\"10.1007/s11600-025-01592-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang F, Xiao F, Shi B et al (2025) Deterioration trend assessment of hydropower units based on Optuna-CatBoost and CRITIC evaluation method. Hydropower Energy Sci 43(02):206\u0026ndash;210. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.20040/j.cnki.1000-7709.2025.20241751\u003c/span\u003e\u003cspan address=\"10.20040/j.cnki.1000-7709.2025.20241751\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(in Chinese)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZheng Y, Li X, Chen Y et al (2024) Short-term wind power forecasting method for extreme weather based on stacking multi-model fusion. High Voltage Eng 50(09):3871\u0026ndash;3882. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.13336/j.1003-6520.hve.20240490\u003c/span\u003e\u003cspan address=\"10.13336/j.1003-6520.hve.20240490\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZou Y, Li Q, Que W et al (2024) Comparison of lithology identification of in situ leaching logging data based on different machine learning algorithms. Uranium Min Metall 1\u0026ndash;10. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.13426/j.cnki.yky.2024.10.09\u003c/span\u003e\u003cspan address=\"10.13426/j.cnki.yky.2024.10.09\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e(in Chinese)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"acta-geophysica","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"agph","sideBox":"Learn more about [Acta Geophysica](http://link.springer.com/journal/11600)","snPcode":"11600","submissionUrl":"https://www.editorialmanager.com/agph/default2.aspx","title":"Acta Geophysica","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Lithology Identification, Seismic Data, Ensemble Learning, Hyperparameter Optimization, Comparative Analysis","lastPublishedDoi":"10.21203/rs.3.rs-6795780/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6795780/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLithology identification is one of the key geological interpretation tasks in oil and gas exploration, which directly affects the accuracy of reservoir modeling and resource evaluation. At present, seismic lithology prediction research has the limitation of over-focusing on the optimization of a single algorithm and lacking systematic comparison of ensemble models and hyperparameter strategies. To this end, this study employs recursive feature elimination, NM-SMOTE sampling, and four hyperparameter optimization methods. These are applied to well seismic data from the F3 exploration area in the North Sea to evaluate random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), categorical boosting (CatBoost) and stacked ensemble models (SEM).The experimental results show that in terms of hyperparameter optimization, the Optuna algorithm achieves the best balance between computational efficiency and model performance, and its optimization effect is significantly better than that of the traditional grid search method. In the context of single models, CatBoost shows the best prediction performance (AUC = 0.91), with clear boundaries for sandstone and mudstone identification and the best spatial continuity of the prediction results. The comparative analysis of different ensemble models shows that random forest has the highest stability, followed by LightGBM, while XGBoost is more sensitive to data noise, resulting in a instability in the prediction results. It is worth noting that the classification performance of the SEM is limited under complex geological conditions such as thin interbeds. This study systematically evaluates the technical characteristics of each model and proposes model selection criteria for different geological application scenarios, providing important theoretical basis and method support for the practical application of intelligent lithology identification technology.\u003c/p\u003e","manuscriptTitle":"Optimization and Evaluation of Ensemble Learning Models for Intelligent Lithology Identification Using Seismic Data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-10 07:06:43","doi":"10.21203/rs.3.rs-6795780/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Major revisions","date":"2025-08-15T10:48:41+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"","date":"2025-07-08T11:36:52+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-07-08T10:45:51+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"Acta Geophysica","date":"2025-06-19T14:17:44+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-06-06T13:03:55+00:00","index":"","fulltext":""},{"type":"submitted","content":"Acta Geophysica","date":"2025-06-04T10:30:05+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"acta-geophysica","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"agph","sideBox":"Learn more about [Acta Geophysica](http://link.springer.com/journal/11600)","snPcode":"11600","submissionUrl":"https://www.editorialmanager.com/agph/default2.aspx","title":"Acta Geophysica","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"6204a58d-d32c-4610-92eb-fdd0747448a1","owner":[],"postedDate":"July 10th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-12-15T16:00:28+00:00","versionOfRecord":{"articleIdentity":"rs-6795780","link":"https://doi.org/10.1007/s11600-025-01723-1","journal":{"identity":"acta-geophysica","isVorOnly":false,"title":"Acta Geophysica"},"publishedOn":"2025-12-10 15:56:58","publishedOnDateReadable":"December 10th, 2025"},"versionCreatedAt":"2025-07-10 07:06:43","video":"","vorDoi":"10.1007/s11600-025-01723-1","vorDoiUrl":"https://doi.org/10.1007/s11600-025-01723-1","workflowStages":[]},"version":"v1","identity":"rs-6795780","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6795780","identity":"rs-6795780","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}