Bandgap Prediction of Binary Compounds via a Machine Learning Approach Utilizing Gradient Boosting Regression

Bandgap Prediction of Binary Compounds via a Machine Learning Approach Utilizing Gradient Boosting Regression | 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 Bandgap Prediction of Binary Compounds via a Machine Learning Approach Utilizing Gradient Boosting Regression Mauludi Ariesto Pamungkas, Annida Early Rofiah, Agus Naba This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8819164/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 27 Apr, 2026 Read the published version in Journal of Computational Electronics → Version 1 posted 7 You are reading this latest preprint version Abstract This study developed a predictive model for the bandgap of binary materials using a machine learning approach based on the Gradient Boosting Regressor (GBR). The model showed exceptional performance on the training set, achieving a Mean Absolute Error (MAE) of 0.087, a Root Mean Squared Error (RMSE) of 0.118, and an R 2 of 98%. While the model's performance on the unseen test data was lower, with an R 2 of 77%, this still indicates a strong capability to predict bandgaps, explaining over 77% of the variation. The performance drop suggests a minor degree of overfitting to the training data. Visualization of the results shows that while most predictions align closely with the ideal reference line, a few outliers lead to significant errors for certain materials. This is likely because compositional features alone are insufficient to capture the intricate physical properties governing these materials. Bandgap Binary Compounds Machine Learning Gradient Boosting Regression Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1. Introduction Bandgap is a critical parameter in materials engineering, enabling the design of materials suitable for specific applications. For instance, Solar Cells are critical photovoltaic devices engineered to convert solar energy into electrical power efficiently. This functionality necessitates the use of a semiconductor material characterized by a narrow to optimal bandgap, typically in the range of 1.1 to 1.8 eV. This specific bandgap range is crucial to maximizing the absorption of the broad solar spectrum, thereby ensuring the efficient conversion of incident light into usable electricity. A variety of materials, including binary compounds such as Gallium Arsenide (GaAs), as well as emerging thin-film technologies like copper indium gallium selenide (CIGS), various Chalcopyrite materials, and organic semiconductors are employed to achieve this goal [1], [2], [3], [4]. Light-Emitting Diodes (LEDs) are another prime example of solid-state devices intimately linked to the bandgap. Their operation relies on the strategic use of a tunable wide bandgap semiconductor to achieve electroluminescence by emitting light within specific visible or ultraviolet (UV) spectra when an electrical current is applied, which results from electron-hole recombination across the bandgap. GaN and GaP binary compounds, including their ternary combinations (GaNP), constitute effective materials for LED technology. GaP is commonly used in LEDs to produce red through green light, adding nitrogen (N) as a dopant specifically enables the emission of green light. The suitability of Gallium Nitride (GaN), which is also a binary alloy, for high efficiency of GaN-based blue LEDs stems from its wide, direct bandgap of 3.4 eV, which facilitates light emission across the blue to ultraviolet spectrum[5], [6]. Several binary semiconductor alloys, such as Gallium Nitride (GaN), Silicon Carbide (SiC), Zinc Oxide (ZnO), and Gallium Oxide, are used in photodetectors to filter out visible light interference due to their large bandgap energy. Photodetectors are essential devices in optical sensing, designed to convert incident photons into an electrical signal. This specialized function requires a tunable wide to ultra-wide bandgap semiconductor, as the bandgap energy directly dictates the range of detectable light wavelengths. [7], [8], [9]. Meanwhile, other binary semiconductors, such as Silicon Carbide (SiC), and Gallium Nitride (GaN), are also used in transistor fabrication due to their wide bandgaps and other favorable physical properties[10], [11], [12]. Experimental data are considered the ground truth values for the bandgap. These values are typically obtained using techniques such as UV-Vis absorption spectroscopy[13], diffuse reflectance spectroscopy[14], or photoluminescence spectroscopy[15]. While experimental measurements provide the most accurate bandgap values under specific measurement conditions (temperature, pressure, sample quality), the process is often expensive, time-consuming, and highly dependent on sample quality (purity, crystallinity). Consequently, for materials that are difficult to synthesize or are thermodynamically unstable, reliable experimental data are often scarce or unavailable. In contrast, Density Functional Theory (DFT) is the most common quantum mechanical computational method used to predict the electronic properties of materials, including the bandgap. DFT calculations offer advantages in terms of speed and cost-effectiveness and can be applied to hypothetical materials that have not yet been synthesized. DFT can systematically predict bandgaps for thousands of materials, effectively bridging gaps in experimental databases. However, standard DFT approximations (such as LDA or GGA) systematically underestimate (yield values that are too low) the bandgap compared to experimental values. This discrepancy, often significant (frequently 50% or more), is widely known as the bandgap problem[16], [17]. Machine Learning (ML) emerges as an efficient solution to overcome these limitations[17], [18]. Specifically, supervised learning models can be trained using available data (both experimental and computational) to make fast and accurate bandgap predictions for novel materials[19]. Among various ML algorithms, Gradient Boosting Regression (GBR) offers unique advantages that make it an ideal choice for this task. GBR is an ensemble technique that constructs models sequentially. A new model is developed to correct the errors (residuals) produced by the preceding models using gradients (mathematical slopes). This process is anticipated to yield high predictive performance. Its accuracy has been rigorously validated across a diverse range of applications, including the prediction of disease outbreaks[20], the identification of geochemical anomalies[21], individual electricity consumption forecasting[22], reservoir characterization[23], and also bandgap prediction[24]. Prior research has demonstrated the efficacy of machine learning in predicting the bandgap of diverse materials, primarily focusing on perovskites[24], [25], 2D materials[26], zincblende[27], ZnO[28], and transition metals[29]. Nevertheless, there is a notable lack of specialized studies targeting binary alloys, a gap that our current work aims to fill. This study aims to develop a predictive model for estimating the bandgap values of binary compounds based on their chemical composition using a machine learning approach based on the Gradient Boosting Regressor (GBR). The results showed that the model performed excellently on the training data, achieving an R² score of 98%, MAE of 0.087 eV, and RMSE of 0.118 eV. However, performance decreased on the test data, with an R² of 77.4%, indicating potential overfitting to the training set. Despite this, the model was still able to explain a large portion of the bandgap variance in previously unseen material. 2. Computational Methods The stages of this research are shown in Fig. 1 . The data for this study were obtained from the Materials Project database using Pymatgen, a Python library that facilitates direct data retrieval. Pymatgen was employed to download binary material data. The data acquisition process was conducted by applying specific filtering criteria to include only materials with bandgap values ranging from 0.5 to 4 eV, resulting in a dataset of 3026 samples. This specific range was selected due to its high relevance to semiconductor applications commonly utilized in electronic devices. The collected material properties comprise the material ID, chemical formula, elemental composition, bandgap energy, density, formation energy per atom, volume, the number of atomic sites per unit cell, and total energy per atom.. Subsequently, this data was processed using Matminer for the extraction of additional features. Once the data were acquired, the subsequent stage involved feature extraction from the composition data using Matminer. This Python library is designed to streamline the automated process of extracting features from materials. The chemical formula strings were converted into composition objects using StrToComposition from Matminer. In this study, the extracted features encompass various material composition properties that can influence the band gap value. Feature extraction is the process of retrieving essential information from raw data to be used as input for a Machine Learning (ML) model. The features were specifically extracted based on the material's chemical composition. The primary features utilized in the modeling are the composition features generated using ElementProperty.from_preset('magpie') from the Matminer library with the 'magpie' preset. This process transforms the chemical composition data into numerical features that describe the elemental properties of the constituent elements, such as electronegativity, atomic radius, ionization energy, ionic radius, melting point, boiling point, heat capacity, elastic modulus, magnetic moment, and cohesive energy. Magpie does not merely take the value of a single element but also calculates aggregate statistics of the material's constituent elements, including Mean (the average of all elements in the compound), Range (the difference between the maximum and minimum values), and Mean absolute deviation (the average deviation from the mean value). This process yielded 142 features from Matminer. These features are highly instrumental in numerically representing the material characteristics for use in machine learning algorithms. However, not all these features may have a significant contribution to the bandgap prediction. Therefore, the results of this feature extraction were combined with existing manual features, such as density, formation energy, and crystallographic information. The combined dataset, comprising the extracted and manual features, was subsequently cleaned and preprocessed in preparation for the model training phase. The data cleaning process included the following steps: Imputation or removal of samples with missing values (i.e., handling of null or NAN entries). Removal of highly correlated features (with a Pearson correlation coefficient greater than 0.95) to reduce data redundancy, prevent overfitting, and mitigate multicollinearity, which can otherwise degrade model performance. Elimination of features exhibiting high sparsity (i.e., those containing a large number of zero values) to ensure data quality and maintain feature informativeness. The dataset was subsequently split into the input features (X) and the output target variable (y), with the latter representing the material's band gap that the model is designed to predict. The primary model selected was the Gradient Boosting Regressor (GBR) due to its robust capability in handling non-linear relationships between the features (input variables) and the target. The boosting process is iterative, with each new model influenced by the performance of those built previously. It integrates multiple weak learners to create a single, highly accurate predictive model. Through an iterative process, the performance of preceding models is evaluated so that subsequent models can focus on correcting prior errors, utilizing either voting or averaging mechanisms. This iteration process is represented in the following equation: $$\:{F}_{m}\left(x\right)={F}_{m-1}\left(x\right)+{\gamma\:}_{m}{h}_{m}\left(x\right)$$ \(\:{F}_{m}\left(x\right)\:\) represents the prediction model at the m-th iteration, where the prediction accuracy improves as iterations progress. \(\:{F}_{m-1}\left(x\right)\:\) is the model in the previous iteration \(\:{h}_{m}\left(x\right)\:\) is the new model (usually a small decision tree) drilled to predict the error (residual) of the previous model. \(\:\:\:\:\:\:\:\:{\gamma\:}_{m}\) is the learning rate, which is how much the new model contributes to the total model The technique for evaluating the ML model's performance was executed by partitioning the dataset into two subsets (training and testing), and then training and testing the model accordingly. In this research, a direct data split into training and testing sets was performed to achieve more stable results. Subsequently, feature selection was implemented using SelectFromModel with the Gradient Boosting Regressor as the estimator to select the most important features based on the median threshold value. 3. Results and Discussion The data extraction process yielded a total of 3026 binary compound data points. The distribution of the bandgap values is presented in Table 1 .a, the descriptive statistics of the bandgap are shown in Table 1 .b, and Fig. 2 as follows: Table 1 a. Distribution of band gap values Bandgap Energy (eV) Number of materials 0.50–0.85 544 0.85–1.20 506 1.20–1.55 394 1.55–1.90 343 1.90–2.25 348 2.25–2.60 336 2.60–2.95 154 2.95–3.30 153 3.30–3.65 129 3.65–4.00 129 Table 1 b. Descriptive Statistics of the Bandgap Energy Statistical Parameters Values count (number of data ) 3026 mean 1.763 eV std (Standard Deviation) 0.916 eV min (minimum value) 0.501 eV 25% quartil (Q1) 0.977 eV 50% quartil (median) 1.618 eV 75% quartil (Q3) 2.364 eV max (nmaximum value) 3.999 eV Modus 0.613 eV Tables 1 .a, 1.b, and Fig. 2 reveal a relatively left-skewed distribution of band gap values, with the majority concentrated below 2 eV. This dominance of narrow-bandgap semiconductors highlights their significant potential in photovoltaic and electronic applications. However, this high variation in the target variable, especially in the low band gap regime, poses a considerable challenge for predictive modeling. Consequently, this statistical insight guided the data separation process using quantile binning to guarantee a balanced label distribution during the train-test split. The initial dataset is defined by the target variable (band gap) and a set of preliminary input features. These predictors were directly extracted from the Materials Project and are fundamentally rooted in the atomic and structural characteristics of the compounds. Figure 3 details the dataset structure, which includes physical properties like the formula_pretty, composition, density, formation_energy_per_atom, volume, nsites, energy_per_atom, and the spacegroup number. After acquiring the material data from the Materials Project, the subsequent step involved extracting numerical features based on the chemical composition of each material using matminer, a Python library that provides various materials science-based featurizers. The chemical composition-based features were extracted utilizing the ElementProperty featurizer with the 'magpie' preset. Employing the 'magpie' preset yielded a total of 135 numerical features derived from the 'composition' column. This specific preset is one of the most commonly used in materials property prediction as it encompasses various statistical measures (mean, maximum, minimum, range, standard deviation) of fundamental elemental properties, such as atomic number, atomic weight, Atomic radius, density, and melting point, Electron affinity, electronegativity, band gap, and specific heat. In addition to the composition-based features, supplementary numerical data from the Materials Project were utilized to complete the material information. These features were manually integrated and merged into the dataset using the material_id as the key. Following this integration, the total number of numerical features reached 142 columns (135 from magpie + 7 manual). Subsequently, data cleaning was performed to remove irrelevant features. Specifically, columns containing over 100 zero values or exhibiting zero variance were deemed uninformative and discarded. Furthermore, features showing very high correlation (redundant) were addressed. Features with a correlation coefficient exceeding were considered redundant, and one of the pair was removed. This step was crucial to prevent multicollinearity in the predictive model. Following the data cleaning process, the total number of features was reduced from 142 to approximately 62. To optimize performance and prevent overfitting, feature selection was subsequently performed using the SelectFromModel method based on the GradientBoostingRegressor. The initial model was trained using all cleaned features. Then, SelectFromModel was employed with a threshold='median' parameter to select only the features whose feature importance was above the median importance value. This selection process yielded twenty-nine features, which were deemed the most crucial in influencing the band gap value. This entire procedure implicitly balances the number of features and the complexity of the predictive model. From a total dataset of 3,026 entries, the data were partitioned into two subsets—the training set and the test set—using a ratio of 20:80. As shown in the Fig. 4 , this allocation resulted in 2,420 samples for training and 606 samples for testing, respectively. The dataset was partitioned using a stratified sampling technique, specifically the StratifiedShuffleSplit algorithm. This method ensures that the distribution of the target variable (band gap) remains consistent and balanced across both subsets. Given that band gap is a continuous variable, a binning process was applied to discretize the values into six distinct quantiles using the pd.qcut function. To ensure statistical stability, only bins containing more than ten data points were retained. This stratification strategy was implemented to ensure that the model receives a representative distribution for both training and testing, thereby mitigating bias toward specific bandgap ranges. This is particularly crucial for materials data, where bandgap values can vary extensively depending on the underlying elemental composition The top eight features that were most influential in predicting the band gap values of binary compounds were identified using the Gradient Boosting Regressor model, as illustrated in Fig. 5 . The values displayed represent the relative importance of each feature in contributing to the model's performance in predicting the target variable bandgap. A higher importance value indicates that the feature plays a more significant role in shaping the model's decision-making process during bandgap prediction. Figure 5 shows that the formation energy per atom provides the highest contribution (approximately 22%) to the model's predictive performance. The observed trend aligns with the principle that thermodynamic stability, indicated by formation energy, is intrinsically linked to bond strength and the resulting energy band gap. In materials with high ionic character, large electronegativity differences lead to strong electron localization at the anion sites, resulting in a wide band gap. In contrast, materials with lower formation energies tend toward covalency and electron delocalization, which narrows the gap. Stronger bonds yield greater orbital splitting between bonding and antibonding states, inherently increasing the band gap. This strong interdependence between formation energy and the band gap has been demonstrated by Al-Qahtani et al. in their DFT study of highly stable phases of inorganic ABX 3 halides[30]. A robust correlation between atomic density and the energy band gap has been extensively documented in prior literature. The comprehensive compilation of semiconductor parameters for III-V alloys by Vurgaftman et al. provides systematic evidence of how the lattice parameter—and by extension, the structural density—governs electronic properties[31]. High atomic density implies a reduced interatomic spacing, which facilitates a more pronounced overlap of neighboring atomic orbitals. This intensified orbital overlap causes discrete atomic energy levels to broaden into energy bands. Consequently, as the valence and conduction bands widen due to increased overlap, the forbidden energy region between them, known as the band gap, concurrently diminishes. The Mendeleev Number (MN) serves as a one-dimensional index assigned to chemical elements to sequentially categorize them based on similarities in electronegativity, atomic radius, and valence electrons[32], [33]. A heightened electronegativity gradient between constituent atoms, such as in Ga-As bonds, strengthens the ionic character of the bond. This disparity enhances the periodic potential, which effectively widens the separation between the valence and conduction bands. Conversely, a smaller atomic radius facilitates reduced bond lengths, leading to a robust overlap of neighboring atomic orbitals. Furthermore, the configuration of valence electrons (e.g., s, p, d orbitals) dictates the band filling, symmetry, and hybridization strength. For instance, the formidable covalent bonding resulting from sp 3 hybridization, as observed in diamond, necessitates higher excitation energy, thereby leading to an increased energy band gap[34]. The space group number represents the crystalline symmetry of a material. This symmetry dictates the arrangement and interaction of energy bands across the Brillouin zone, playing a crucial role in the formation of band gaps. Consequently, even in composition-based predictive models, space group information remains a critical factor in determining the material's band gap value[35]. The covalent radius of constituent elements also affects electron and hole localization. Larger covalent radii favor hole localization, while smaller radii facilitate electron localization, both of which drive changes in the energy band structure and the resulting band gap[36]. Following the initial training phase, feature selection was performed using the SelectFromModel method with a 'median' threshold. Under this approach, only features with importance scores exceeding the median value of all features were retained. This process effectively mitigates overfitting and enhances model interpretability. Consequently, twenty-nine optimal features were identified for the predictive model. After feature selection, the model was retrained to achieve superior accuracy. Figure 6 illustrates the model's performance by comparing the ground truth bandgap values against the predicted values for both the training and testing sets. Following back-transformation via np.expm1, the model's predictions were evaluated against actual values. Performance on the training set yielded an MAE of 0.087 eV, an RMSE of 0.118 eV, and an R 2 of 0.983. This high R 2 value signifies that the model explains 98.3% of the training data variance, reflecting a robust fit during the learning phase. Furthermore, the model was evaluated using a testing dataset—comprising data unseen during the training phase—to assess its generalizability. The evaluation yielded a Mean Absolute Error (MAE) of 0.293 eV, a Root Mean Squared Error (RMSE) of 0.433 eV, and an R-squared (R 2 ) value of 0.774. In evaluating regression performance, the coefficient of determination (R 2 ) serves as a critical metric, representing the proportion of variance in the target variable (the band gap) that can be explained by the input features. R 2 values range from 0 to 1, where a value approaching 1 signifies superior predictive capability. In the fields of physics and chemistry, R 2 values between 0.70 and 0.99 are generally regarded as indicative of high-quality results. This expectation arises because physical data are typically more deterministic and exhibit stronger physical correlations between variables[37]. The distribution of the band gap values predicted by the machine learning model, alongside the corresponding ground truth values, is presented in Fig. 7 . The model developed in this study achieved an $ R^2 $ value of 0.774 on the testing set, indicating that approximately 77.4% of the variance in the band gap values can be explained by the input features. The observed decline in performance on the testing data suggests a degree of slight overfitting, where the model may have overfitted to the training set, thereby limiting its generalizability to unseen data. This is likely since composition-based features alone may not fully capture the intricate electronic structures of materials, leaving more complex physical and structural properties underrepresented. Nevertheless, the GBR model developed here serves as a viable initial approach for predicting the band gap of binary materials based on their chemical composition. Figure 8 illustrates the correlation between predicted and actual band gap values, labeled with chemical formulas and absolute errors. The results are displayed via scatter plots for the training and testing phases, where the y = x dashed line represents ideal prediction. Most testing points lie in close proximity to the reference line, facilitating the identification of significant outliers. While the overall distribution confirms the model's accuracy, the presence of specific outliers suggests the existence of unique electronic properties or complexities that transcend the limitations of a purely compositional approach. Table 2 Comparison of the GBR-predicted band gaps and the actual values Material ID Formula Actual Bandgap (eV) Bandgap Predicted by this model (eV) Absolute Error mp-1205297 Tb 2 S 3 0.7331 0.7329 0.0002 mp-556784 ZnS 1.9920 1.9910 0.0010 mp-680145 CdI 2 2.3186 2.3175 0.0011 mp-1038759 Mg 2 Bi 0.8771 0.8755 0.0016 mp-27194 SnI 2 1.7061 1.7080 0.0019 mp-624418 CdI 2 2.3651 2.3677 0.0026 mp-422 BeS 3.1453 3.1491 0.0038 mp-567296 CdI 2 2.3218 2.3258 0.0040 mp-776402 B 10 H 13 3.2507 3.2548 0.0041 mp-650112 Fe 3 O 4 0.9248 0.9206 0.0042 mp-624399 CdI 2 2.3414 2.3461 0.0047 mp-1245096 V 2 O 5 0.7321 0.7273 0.0048 mp-680205 PbI 2 2.3541 2.3485 0.0054 mp-27267 ReP 4 0.8553 0.8488 0.0065 mp-10009 GaTe 0.7853 0.7919 0.0066 mp-1080367 CeSe 2 0.9720 0.9787 0.0067 mp-759097 VF 3 1.6507 1.6440 0.0067 mp-1244954 Al 2 O 3 3.5224 3.5297 0.0073 mp-1080356 CeSe 2 0.9568 0.9648 0.0080 mp-22009 PbSe 1.2995 1.2910 0.0085 mp-1196206 C 9 Cl 1.3653 1.3739 0.0085 mp-1279995 Mn 3 O 4 0.8312 0.8310 0.0087 mp-759504 Fe 10 O 11 1.3086 1.3179 0.0093 mp-22398 FeF 3 3.1061 3.1156 0.0095 mp-2371 Ga 2 Te 5 0.7231 0.7135 0.0094 mp-1714 FeSi 2 0.6966 0.7067 0.0101 mp-680163 CdI 2 2.3597 2.3486 0.0111 mp-2667 CsAu 1.0233 1.0344 0.0111 mp-680093 CdI 2 2.3352 2.3468 0.0116 mp-561153 XeF 3 2.6269 2.6148 0.0121 A comparative analysis between GBR-predicted band gaps and measured values is summarized in Table 2 . The minimal error observed indicates the high precision of the GBR machine learning model in predicting band gaps. While the current application is limited to binary compounds, we believe the model possesses the robustness to yield accurate predictions for diverse material classes. 4. Conclusion This study successfully developed an efficient machine learning model to predict the bandgap values of binary compounds based on chemical composition using the Gradient Boosting Regressor (GBR) algorithm. Evaluation results on the training dataset demonstrated exceptional performance, achieving an MAE of 0.087 eV, an RMSE of 0.118 eV, and an R 2 of 98.3%, indicating the model’s robust ability to capture the relationship between composition and bandgap with high precision. However, when evaluated on unseen data, the performance decreased, yielding an R 2 of 0.774. This decline suggests the presence of overfitting, likely due to the limitations of composition-based features in fully representing the complexity of electronic structures. Furthermore, the presence of outliers—specific materials with unique electronic characteristics—also influenced the predictive accuracy. Despite these challenges, the model’s performance remains well within the acceptable range for materials science applications. Visualization of the predicted versus actual values shows that most data points align closely with the reference line (y = x) with minimal error, notwithstanding the outliers that reflect the distinct properties of certain materials. Declarations Conflict of interest The authors declare no conflict of interest Funding The authors gratefully acknowledge the financial support provided by Universitas Brawijaya through research grant contract number : 02115.52/UN10.F0/901/B/PG/2025 Author Contribution Mauludi Ariesto Pamungkas established the conceptual framework, performed the formal analysis, and developed the research methodology. He was also responsible for securing resources, providing overall supervision, and leading the manuscript's review and editing process.Annida Early Rofiah developed the software and machine learning models, generated all data visualizations, and prepared the original draft of the manuscript.Agus Naba contributed through technical supervision and assisted in the final editing of the paper. Data Availability The data underlying this article will be shared on reasonable request to the corresponding author. The data that support the findings of this study are openly available in Materials Project References O. Guesmi, M. Ben Arbia, F. Saidi, M. Ben Rabeh, and H. Maaref, “Experimental and computational studies of CCTS material: Novel design of solar cell for high efficiency,” Solar Energy , vol. 262, p. 111806, Sep. 2023, doi: 10.1016/j.solener.2023.111806. B. R. Sutherland, “Solar Materials Find Their Band Gap,” Joule , vol. 4, no. 5, pp. 984–985, May 2020, doi: 10.1016/j.joule.2020.05.001. N. Papež, R. Dallaev, Ş. Ţălu, and J. Kaštyl, “Overview of the Current State of Gallium Arsenide-Based Solar Cells,” Materials , vol. 14, no. 11, p. 3075, Jun. 2021, doi: 10.3390/ma14113075. C. Battaglia, A. Cuevas, and S. De Wolf, “High-efficiency crystalline silicon solar cells: status and perspectives,” Energy Environ. Sci. , vol. 9, no. 5, pp. 1552–1576, 2016, doi: 10.1039/C5EE03380B. Y. Wang et al. , “A review of gallium phosphide nanophotonics towards omnipotent nonlinear devices,” Nanophotonics , vol. 13, no. 18, pp. 3207–3252, Aug. 2024, doi: 10.1515/nanoph-2024-0172. S. Jiang et al. , “Study on Light Extraction from GaN-based Green Light-Emitting Diodes Using Anodic Aluminum Oxide Pattern and Nanoimprint Lithography,” Sci Rep , vol. 6, no. 1, p. 21573, Feb. 2016, doi: 10.1038/srep21573. H. D. Jabbar, M. A. Fakhri, and M. Jalal AbdulRazzaq, “Gallium Nitride –Based Photodiode: A review,” Materials Today: Proceedings , vol. 42, pp. 2829–2834, 2021, doi: 10.1016/j.matpr.2020.12.729. A. Aldalbahi et al. , “A new approach for fabrications of SiC based photodetectors,” Sci Rep , vol. 6, no. 1, p. 23457, Mar. 2016, doi: 10.1038/srep23457. M. Banari, N. Memarian, I. Concina, and A. Vomiero, “UV photodetector study based on Ce: ZnO nanostructures with different concentration of Ce dopant,” Optical Materials , vol. 146, p. 114576, Dec. 2023, doi: 10.1016/j.optmat.2023.114576. S. M. Abd El-Azeem and S. M. El-Ghanam, “Comparative study of gallium nitride and silicon carbide MOSFETs as power switching applications under cryogenic conditions,” Cryogenics , vol. 107, p. 103071, Apr. 2020, doi: 10.1016/j.cryogenics.2020.103071. Q. Yu, “Comparative Analysis of Sic and Gan: Third-Generation Semiconductor Materials,” HSET , vol. 81, pp. 484–490, Jan. 2024, doi: 10.54097/2q3qyj85. M. I. Idris and A. B. Horsfall, “3D structures for silicon carbide transistors utilising Al 2 O 3 as a gate dielectric,” Materials Science in Semiconductor Processing , vol. 128, p. 105727, Jun. 2021, doi: 10.1016/j.mssp.2021.105727. P. H. M. Andrade, C. Volkringer, T. Loiseau, A. Tejeda, M. Hureau, and A. Moissette, “Band gap analysis in MOF materials: Distinguishing direct and indirect transitions using UV–vis spectroscopy,” Applied Materials Today , vol. 37, p. 102094, Apr. 2024, doi: 10.1016/j.apmt.2024.102094. V. Mishra et al. , “Diffuse reflectance spectroscopy: An effective tool to probe the defect states in wide band gap semiconducting materials,” Materials Science in Semiconductor Processing , vol. 86, pp. 151–156, Nov. 2018, doi: 10.1016/j.mssp.2018.06.025. N. Roisin, M.-S. Colla, R. Scaffidi, T. Pardoen, D. Flandre, and J.-P. Raskin, “Band gap reduction in highly-strained silicon beams predicted by first-principles theory and validated using photoluminescence spectroscopy,” Optical Materials , vol. 144, p. 114347, Oct. 2023, doi: 10.1016/j.optmat.2023.114347. X. Zheng, A. J. Cohen, P. Mori-Sánchez, X. Hu, and W. Yang, “Improving Band Gap Prediction in Density Functional Theory from Molecules to Solids,” Phys. Rev. Lett. , vol. 107, no. 2, p. 026403, Jul. 2011, doi: 10.1103/PhysRevLett.107.026403. I. Jihad, M. H. S. Anfa, S. M. Alqahtani, and F. H. Alharbi, “DFT-PBE band gap correction using machine learning with a reduced set of features,” Computational Materials Science , vol. 244, p. 113153, Sep. 2024, doi: 10.1016/j.commatsci.2024.113153. T. Ko, T. Park, M. Kim, and K. Min, “Enhancing predictions of experimental band gap using machine learning and knowledge transfer,” Materials Today Communications , vol. 41, p. 110717, Dec. 2024, doi: 10.1016/j.mtcomm.2024.110717. P. R. Regonia, C. M. Pelicano, R. Tani, A. Ishizumi, H. Yanagi, and K. Ikeda, “Predicting the band gap of ZnO quantum dots via supervised machine learning models,” Optik , vol. 207, p. 164469, Apr. 2020, doi: 10.1016/j.ijleo.2020.164469. R. Huang, C. McMahan, B. Herrin, A. McLain, B. Cai, and S. Self, “Gradient boosting: A computationally efficient alternative to Markov chain Monte Carlo sampling for fitting large Bayesian spatio-temporal binomial regression models,” Infectious Disease Modelling , vol. 10, no. 1, pp. 189–200, Mar. 2025, doi: 10.1016/j.idm.2024.09.008. M. Seyedrahimi-Niaraq, H. Mahdiyanfar, and M. H. Olyaee, “Comparison of the performance of gradient boost, linear regression, decision tree, and voting algorithms to separate geochemical anomalies areas in the fractal environment,” Artificial Intelligence in Geosciences , vol. 6, no. 2, p. 100156, Dec. 2025, doi: 10.1016/j.aiig.2025.100156. L. F. M. Sepulveda et al. , “Forecasting of individual electricity consumption using Optimized Gradient Boosting Regression with Modified Particle Swarm Optimization,” Engineering Applications of Artificial Intelligence , vol. 105, p. 104440, Oct. 2021, doi: 10.1016/j.engappai.2021.104440. D. A. Otchere, T. O. A. Ganat, J. O. Ojero, B. N. Tackie-Otoo, and M. Y. Taki, “Application of gradient boosting regression model for the evaluation of feature selection techniques in improving reservoir characterisation predictions,” Journal of Petroleum Science and Engineering , vol. 208, p. 109244, Jan. 2022, doi: 10.1016/j.petrol.2021.109244. D. Yang et al. , “Bandgap prediction and design of halide double perovskites via ensemble machine learning,” Chemical Engineering Journal , vol. 525, p. 169979, Dec. 2025, doi: 10.1016/j.cej.2025.169979. F. J. Kusuma et al. , “Direct band gap prediction of single and double perovskite using cost-sensitive ensemble learning,” Journal of Alloys and Compounds , vol. 1037, p. 182102, Aug. 2025, doi: 10.1016/j.jallcom.2025.182102. H. He, Y. Wang, Y. Qi, Z. Xu, Y. Li, and Y. Wang, “From prediction to design: Recent advances in machine learning for the study of 2D materials,” Nano Energy , vol. 118, p. 108965, Dec. 2023, doi: 10.1016/j.nanoen.2023.108965. M. Alsalman, S. M. Alqahtani, and F. H. Alharbi, “Bandgap energy prediction of senary zincblende III–V semiconductor compounds using machine learning,” Materials Science in Semiconductor Processing , vol. 161, p. 107461, Jul. 2023, doi: 10.1016/j.mssp.2023.107461. H. Lamouadene, M. El Kassaoui, M. El Yadari, A. El Kenz, and A. Benyoussef, “Exploring modeling techniques for predicting band gaps of Doped-ZnO: A Machine learning approach,” Chemical Physics , vol. 591, p. 112603, Mar. 2025, doi: 10.1016/j.chemphys.2025.112603. M. H. Saeed et al. , “Determination of bandgap of period 3, 4, and 5 transition metal dopants on zinc oxide using an artificial neural network based approach,” Chemometrics and Intelligent Laboratory Systems , vol. 242, p. 104983, Nov. 2023, doi: 10.1016/j.chemolab.2023.104983. S. M. Alqahtani, A. Q. Alsayoud, and F. H. Alharbi, “Structures, band gaps, and formation energies of highly stable phases of inorganic ABX 3 halides: A = Li, Na, K, Rb, Cs, Tl; B = Be, Mg, Ca, Ge, Sr, Sn, Pb; and X = F, Cl, Br, I,” RSC Adv. , vol. 13, no. 13, pp. 9026–9032, 2023, doi: 10.1039/D3RA00185G. I. Vurgaftman, J. R. Meyer, and L. R. Ram-Mohan, “Band parameters for III–V compound semiconductors and their alloys,” Journal of Applied Physics , vol. 89, no. 11, pp. 5815–5875, Jun. 2001, doi: 10.1063/1.1368156. F. R. Boer and D. G. Pettifor, The Structures of binary compounds . in Cohesion and structure, no. v. 2. Amsterdam New York: North-Holland Sole distributors for the USA and Canada, Elsevier Science Pub. Co, 1989. P. Villars, K. Cenzual, J. Daams, Y. Chen, and S. Iwata, “Data‐Driven Atomic Environment Prediction for Binaries Using the Mendeleev Number. Part 1. Composition AB.,” ChemInform , vol. 35, no. 23, p. chin.200423004, Jun. 2004, doi: 10.1002/chin.200423004. J. C. Phillips, “Ionicity of the Chemical Bond in Crystals,” Rev. Mod. Phys. , vol. 42, no. 3, pp. 317–356, Jul. 1970, doi: 10.1103/RevModPhys.42.317. H. Watanabe and L. Lu, “Space-group theory of photonic bands,” Phys. Rev. Lett. , vol. 121, no. 26, p. 263903, Dec. 2018, doi: 10.1103/PhysRevLett.121.263903. M. P. Polak, P. Scharoch, and R. Kudrawiec, “The effect of isovalent doping on the electronic band structure of group IV semiconductors,” J. Phys. D: Appl. Phys. , vol. 54, no. 8, p. 085102, Feb. 2021, doi: 10.1088/1361-6463/abc503. A. Gupta, T. S. Stead, and L. Ganti, “Determining a Meaningful R-squared Value in Clinical Medicine,” Academic Medicine & Surgery , Oct. 2024, doi: 10.62186/001c.125154. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 27 Apr, 2026 Read the published version in Journal of Computational Electronics → Version 1 posted Editorial decision: Revision requested 16 Mar, 2026 Reviews received at journal 23 Feb, 2026 Reviewers agreed at journal 12 Feb, 2026 Reviewers invited by journal 10 Feb, 2026 Editor assigned by journal 09 Feb, 2026 Submission checks completed at journal 09 Feb, 2026 First submitted to journal 07 Feb, 2026 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-8819164","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":590470252,"identity":"f2defb9b-0f86-4902-8714-0bbdebf4fa6f","order_by":0,"name":"Mauludi Ariesto Pamungkas","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyklEQVRIiWNgGAWjYLCCBBDB3sAG5rBBuIS1SDDwHCBFCwNIi0QCG7KluIFuA/OzDw8q6ur4Jd+YPWCosWPgYyegxewAm/GMhDOHJSRn55gbMBxLZmDjeUBIC4MxQ2LbAQmD2zlmEgxsBxjYJAjawv6ZIfFfnYTBzTNALf+I0sIDtKWBWcLgBo+ZBGMbcVqKGRKOHZac2ZNWbpDYl8xDhF/YNzP+qKnj52c/vO3Bh292cvLtBGxhkEc2E6iYh4D6UTAKRsEoGAXEAACQzjn3H1RR8QAAAABJRU5ErkJggg==","orcid":"","institution":"Universitas Brawijaya","correspondingAuthor":true,"prefix":"","firstName":"Mauludi","middleName":"Ariesto","lastName":"Pamungkas","suffix":""},{"id":590470253,"identity":"894fc932-c4f1-4774-ad58-257a92730ddb","order_by":1,"name":"Annida Early Rofiah","email":"","orcid":"","institution":"Universitas Brawijaya","correspondingAuthor":false,"prefix":"","firstName":"Annida","middleName":"Early","lastName":"Rofiah","suffix":""},{"id":590470254,"identity":"93214468-1c71-4262-8d9e-475d9a354253","order_by":2,"name":"Agus Naba","email":"","orcid":"","institution":"Universitas Brawijaya","correspondingAuthor":false,"prefix":"","firstName":"Agus","middleName":"","lastName":"Naba","suffix":""}],"badges":[],"createdAt":"2026-02-08 04:53:25","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8819164/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8819164/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s10825-026-02547-y","type":"published","date":"2026-04-27T15:57:47+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":102588749,"identity":"9903608e-d734-4c3e-a6c0-52cdfc65ac63","added_by":"auto","created_at":"2026-02-13 10:49:02","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":141918,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic illustration of the overall research process\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8819164/v1/fb3bbf3927b11d4fadb88ef7.png"},{"id":102747257,"identity":"356b93d4-63ad-4d99-b6ac-59a348a88327","added_by":"auto","created_at":"2026-02-16 09:04:18","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":55765,"visible":true,"origin":"","legend":"\u003cp\u003eThe distribution of band gaps within a dataset\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8819164/v1/58caa00d5ebb0c8e37db9380.png"},{"id":102747123,"identity":"0795b1e9-804f-4181-8c9f-2029d62126ce","added_by":"auto","created_at":"2026-02-16 09:03:53","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":177075,"visible":true,"origin":"","legend":"\u003cp\u003eExample of initial dataset atributes\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8819164/v1/3a66cc8ddb2f49f934f7109c.png"},{"id":102588746,"identity":"451ad01c-f7af-4897-a67a-2a7945c52ac7","added_by":"auto","created_at":"2026-02-13 10:49:01","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":175717,"visible":true,"origin":"","legend":"\u003cp\u003eThe histograms for the train-test split datasets\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8819164/v1/f45d440f4db9b0de27509624.png"},{"id":102588744,"identity":"2332b709-89f6-4ee9-9b0b-e899b3e6b037","added_by":"auto","created_at":"2026-02-13 10:49:01","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":33893,"visible":true,"origin":"","legend":"\u003cp\u003eThe eight most significant parameters governing the energy bandgap.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8819164/v1/3c19780cca4a37c389e95caf.png"},{"id":102588750,"identity":"c83138f4-a415-4228-9a00-06f39bb2bd7c","added_by":"auto","created_at":"2026-02-13 10:49:02","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":43060,"visible":true,"origin":"","legend":"\u003cp\u003eTraining outcomes evaluating GBR model accuracy through the comparison of predicted and actual band gaps\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8819164/v1/779a13388599e4a3fc12036e.png"},{"id":102588747,"identity":"8cd7e4fd-b9d4-4306-beae-92b45051eee6","added_by":"auto","created_at":"2026-02-13 10:49:01","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":53984,"visible":true,"origin":"","legend":"\u003cp\u003eTesting outcomes evaluating GBR model accuracy through the comparison of predicted and actual band gaps\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8819164/v1/837738226ef030512aaf3423.png"},{"id":102588751,"identity":"ae354c7f-0b9b-4fe6-997e-68d6697ef884","added_by":"auto","created_at":"2026-02-13 10:49:03","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":84516,"visible":true,"origin":"","legend":"\u003cp\u003eBand gap prediction test and its absolute errors\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8819164/v1/2182649d512e588fa5401462.png"},{"id":108437899,"identity":"ee70213d-b94f-4c9c-ba8a-c0da3582246e","added_by":"auto","created_at":"2026-05-04 16:04:09","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1047008,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8819164/v1/bf9008ff-2e3f-41ee-b3d7-139d6b78680a.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Bandgap Prediction of Binary Compounds via a Machine Learning Approach Utilizing Gradient Boosting Regression","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eBandgap is a critical parameter in materials engineering, enabling the design of materials suitable for specific applications. For instance, Solar Cells are critical photovoltaic devices engineered to convert solar energy into electrical power efficiently. This functionality necessitates the use of a semiconductor material characterized by a narrow to optimal bandgap, typically in the range of 1.1 to 1.8 eV. This specific bandgap range is crucial to maximizing the absorption of the broad solar spectrum, thereby ensuring the efficient conversion of incident light into usable electricity. A variety of materials, including binary compounds such as Gallium Arsenide (GaAs), as well as emerging thin-film technologies like copper indium gallium selenide (CIGS), various Chalcopyrite materials, and organic semiconductors are employed to achieve this goal [1], [2], [3], [4].\u003c/p\u003e \u003cp\u003eLight-Emitting Diodes (LEDs) are another prime example of solid-state devices intimately linked to the bandgap. Their operation relies on the strategic use of a tunable wide bandgap semiconductor to achieve electroluminescence by emitting light within specific visible or ultraviolet (UV) spectra when an electrical current is applied, which results from electron-hole recombination across the bandgap. GaN and GaP binary compounds, including their ternary combinations (GaNP), constitute effective materials for LED technology. GaP is commonly used in LEDs to produce red through green light, adding nitrogen (N) as a dopant specifically enables the emission of green light. The suitability of Gallium Nitride (GaN), which is also a binary alloy, for high efficiency of GaN-based blue LEDs stems from its wide, direct bandgap of 3.4 eV, which facilitates light emission across the blue to ultraviolet spectrum[5], [6].\u003c/p\u003e \u003cp\u003eSeveral binary semiconductor alloys, such as Gallium Nitride (GaN), Silicon Carbide (SiC), Zinc Oxide (ZnO), and Gallium Oxide, are used in photodetectors to filter out visible light interference due to their large bandgap energy. Photodetectors are essential devices in optical sensing, designed to convert incident photons into an electrical signal. This specialized function requires a tunable wide to ultra-wide bandgap semiconductor, as the bandgap energy directly dictates the range of detectable light wavelengths. [7], [8], [9]. Meanwhile, other binary semiconductors, such as Silicon Carbide (SiC), and Gallium Nitride (GaN), are also used in transistor fabrication due to their wide bandgaps and other favorable physical properties[10], [11], [12].\u003c/p\u003e \u003cp\u003eExperimental data are considered the ground truth values for the bandgap. These values are typically obtained using techniques such as UV-Vis absorption spectroscopy[13], diffuse reflectance spectroscopy[14], or photoluminescence spectroscopy[15]. While experimental measurements provide the most accurate bandgap values under specific measurement conditions (temperature, pressure, sample quality), the process is often expensive, time-consuming, and highly dependent on sample quality (purity, crystallinity). Consequently, for materials that are difficult to synthesize or are thermodynamically unstable, reliable experimental data are often scarce or unavailable.\u003c/p\u003e \u003cp\u003eIn contrast, Density Functional Theory (DFT) is the most common quantum mechanical computational method used to predict the electronic properties of materials, including the bandgap. DFT calculations offer advantages in terms of speed and cost-effectiveness and can be applied to hypothetical materials that have not yet been synthesized. DFT can systematically predict bandgaps for thousands of materials, effectively bridging gaps in experimental databases. However, standard DFT approximations (such as LDA or GGA) systematically underestimate (yield values that are too low) the bandgap compared to experimental values. This discrepancy, often significant (frequently 50% or more), is widely known as the bandgap problem[16], [17].\u003c/p\u003e \u003cp\u003eMachine Learning (ML) emerges as an efficient solution to overcome these limitations[17], [18]. Specifically, supervised learning models can be trained using available data (both experimental and computational) to make fast and accurate bandgap predictions for novel materials[19]. Among various ML algorithms, Gradient Boosting Regression (GBR) offers unique advantages that make it an ideal choice for this task. GBR is an ensemble technique that constructs models sequentially. A new model is developed to correct the errors (residuals) produced by the preceding models using gradients (mathematical slopes). This process is anticipated to yield high predictive performance. Its accuracy has been rigorously validated across a diverse range of applications, including the prediction of disease outbreaks[20], the identification of geochemical anomalies[21], individual electricity consumption forecasting[22], reservoir characterization[23], and also bandgap prediction[24]. Prior research has demonstrated the efficacy of machine learning in predicting the bandgap of diverse materials, primarily focusing on perovskites[24], [25], 2D materials[26], zincblende[27], ZnO[28], and transition metals[29]. Nevertheless, there is a notable lack of specialized studies targeting binary alloys, a gap that our current work aims to fill.\u003c/p\u003e \u003cp\u003eThis study aims to develop a predictive model for estimating the bandgap values of binary compounds based on their chemical composition using a machine learning approach based on the Gradient Boosting Regressor (GBR). The results showed that the model performed excellently on the training data, achieving an R\u0026sup2; score of 98%, MAE of 0.087 eV, and RMSE of 0.118 eV. However, performance decreased on the test data, with an R\u0026sup2; of 77.4%, indicating potential overfitting to the training set. Despite this, the model was still able to explain a large portion of the bandgap variance in previously unseen material.\u003c/p\u003e"},{"header":"2. Computational Methods","content":"\u003cp\u003eThe stages of this research are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The data for this study were obtained from the Materials Project database using Pymatgen, a Python library that facilitates direct data retrieval. Pymatgen was employed to download binary material data. The data acquisition process was conducted by applying specific filtering criteria to include only materials with bandgap values ranging from 0.5 to 4 eV, resulting in a dataset of 3026 samples. This specific range was selected due to its high relevance to semiconductor applications commonly utilized in electronic devices. The collected material properties comprise the material ID, chemical formula, elemental composition, bandgap energy, density, formation energy per atom, volume, the number of atomic sites per unit cell, and total energy per atom.. Subsequently, this data was processed using Matminer for the extraction of additional features. Once the data were acquired, the subsequent stage involved feature extraction from the composition data using Matminer. This Python library is designed to streamline the automated process of extracting features from materials. The chemical formula strings were converted into composition objects using StrToComposition from Matminer. In this study, the extracted features encompass various material composition properties that can influence the band gap value. Feature extraction is the process of retrieving essential information from raw data to be used as input for a Machine Learning (ML) model. The features were specifically extracted based on the material's chemical composition.\u003c/p\u003e \u003cp\u003eThe primary features utilized in the modeling are the composition features generated using ElementProperty.from_preset('magpie') from the Matminer library with the 'magpie' preset. This process transforms the chemical composition data into numerical features that describe the elemental properties of the constituent elements, such as electronegativity, atomic radius, ionization energy, ionic radius, melting point, boiling point, heat capacity, elastic modulus, magnetic moment, and cohesive energy.\u003c/p\u003e \u003cp\u003eMagpie does not merely take the value of a single element but also calculates aggregate statistics of the material's constituent elements, including Mean (the average of all elements in the compound), Range (the difference between the maximum and minimum values), and Mean absolute deviation (the average deviation from the mean value). This process yielded 142 features from Matminer. These features are highly instrumental in numerically representing the material characteristics for use in machine learning algorithms. However, not all these features may have a significant contribution to the bandgap prediction. Therefore, the results of this feature extraction were combined with existing manual features, such as density, formation energy, and crystallographic information.\u003c/p\u003e \u003cp\u003eThe combined dataset, comprising the extracted and manual features, was subsequently cleaned and preprocessed in preparation for the model training phase. The data cleaning process included the following steps:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eImputation or removal of samples with missing values (i.e., handling of null or NAN entries).\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eRemoval of highly correlated features (with a Pearson correlation coefficient greater than 0.95) to reduce data redundancy, prevent overfitting, and mitigate multicollinearity, which can otherwise degrade model performance.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eElimination of features exhibiting high sparsity (i.e., those containing a large number of zero values) to ensure data quality and maintain feature informativeness.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThe dataset was subsequently split into the input features (X) and the output target variable (y), with the latter representing the material's band gap that the model is designed to predict.\u003c/p\u003e \u003cp\u003eThe primary model selected was the Gradient Boosting Regressor (GBR) due to its robust capability in handling non-linear relationships between the features (input variables) and the target.\u003c/p\u003e \u003cp\u003eThe boosting process is iterative, with each new model influenced by the performance of those built previously. It integrates multiple weak learners to create a single, highly accurate predictive model. Through an iterative process, the performance of preceding models is evaluated so that subsequent models can focus on correcting prior errors, utilizing either voting or averaging mechanisms. This iteration process is represented in the following equation:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:{F}_{m}\\left(x\\right)={F}_{m-1}\\left(x\\right)+{\\gamma\\:}_{m}{h}_{m}\\left(x\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{F}_{m}\\left(x\\right)\\:\\)\u003c/span\u003e \u003c/span\u003erepresents the prediction model at the m-th iteration, where the prediction accuracy\u003c/p\u003e \u003cp\u003eimproves as iterations progress.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{F}_{m-1}\\left(x\\right)\\:\\)\u003c/span\u003e \u003c/span\u003eis the model in the previous iteration\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{h}_{m}\\left(x\\right)\\:\\)\u003c/span\u003e \u003c/span\u003eis the new model (usually a small decision tree) drilled to predict the error (residual)\u003c/p\u003e \u003cp\u003eof the previous model.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:\\:\\:\\:\\:\\:\\:\\:{\\gamma\\:}_{m}\\)\u003c/span\u003e \u003c/span\u003eis the learning rate, which is how much the new model contributes to the total model\u003c/p\u003e \u003cp\u003eThe technique for evaluating the ML model's performance was executed by partitioning the dataset into two subsets (training and testing), and then training and testing the model accordingly. In this research, a direct data split into training and testing sets was performed to achieve more stable results. Subsequently, feature selection was implemented using SelectFromModel with the Gradient Boosting Regressor as the estimator to select the most important features based on the median threshold value.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"3. Results and Discussion","content":"\u003cp\u003eThe data extraction process yielded a total of 3026 binary compound data points. The distribution of the bandgap values is presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e1\u003c/span\u003e.a, the descriptive statistics of the bandgap are shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e1\u003c/span\u003e.b, and Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e as follows:\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\u003ea. Distribution of band gap values\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBandgap Energy (eV)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of materials\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.50\u0026ndash;0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e544\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.85\u0026ndash;1.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e506\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1.20\u0026ndash;1.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e394\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1.55\u0026ndash;1.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e343\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1.90\u0026ndash;2.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e348\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2.25\u0026ndash;2.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e336\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2.60\u0026ndash;2.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e154\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2.95\u0026ndash;3.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e153\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3.30\u0026ndash;3.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e129\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3.65\u0026ndash;4.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e129\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\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eb. Descriptive Statistics of the Bandgap Energy\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStatistical Parameters\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eValues\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ecount (number of data )\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3026\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emean\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.763 eV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003estd (Standard Deviation)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.916 eV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emin (minimum value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.501 eV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e25% quartil (Q1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.977 eV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e50% quartil (median)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.618 eV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e75% quartil (Q3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.364 eV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emax (nmaximum value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.999 eV\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModus\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.613 eV\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\u003e \u003c/p\u003e \u003cp\u003eTables\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e1\u003c/span\u003e.a, 1.b, and Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e reveal a relatively left-skewed distribution of band gap values, with the majority concentrated below 2 eV. This dominance of narrow-bandgap semiconductors highlights their significant potential in photovoltaic and electronic applications. However, this high variation in the target variable, especially in the low band gap regime, poses a considerable challenge for predictive modeling. Consequently, this statistical insight guided the data separation process using quantile binning to guarantee a balanced label distribution during the train-test split.\u003c/p\u003e \u003cp\u003eThe initial dataset is defined by the target variable (band gap) and a set of preliminary input features. These predictors were directly extracted from the Materials Project and are fundamentally rooted in the atomic and structural characteristics of the compounds. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e details the dataset structure, which includes physical properties like the formula_pretty, composition, density, formation_energy_per_atom, volume, nsites, energy_per_atom, and the spacegroup number.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAfter acquiring the material data from the Materials Project, the subsequent step involved extracting numerical features based on the chemical composition of each material using matminer, a Python library that provides various materials science-based featurizers. The chemical composition-based features were extracted utilizing the ElementProperty featurizer with the 'magpie' preset. Employing the 'magpie' preset yielded a total of 135 numerical features derived from the 'composition' column. This specific preset is one of the most commonly used in materials property prediction as it encompasses various statistical measures (mean, maximum, minimum, range, standard deviation) of fundamental elemental properties, such as atomic number, atomic weight, Atomic radius, density, and melting point, Electron affinity, electronegativity, band gap, and specific heat.\u003c/p\u003e \u003cp\u003eIn addition to the composition-based features, supplementary numerical data from the Materials Project were utilized to complete the material information. These features were manually integrated and merged into the dataset using the material_id as the key. Following this integration, the total number of numerical features reached 142 columns (135 from magpie\u0026thinsp;+\u0026thinsp;7 manual). Subsequently, data cleaning was performed to remove irrelevant features. Specifically, columns containing over 100 zero values or exhibiting zero variance were deemed uninformative and discarded. Furthermore, features showing very high correlation (redundant) were addressed. Features with a correlation coefficient exceeding were considered redundant, and one of the pair was removed. This step was crucial to prevent multicollinearity in the predictive model.\u003c/p\u003e \u003cp\u003eFollowing the data cleaning process, the total number of features was reduced from 142 to approximately 62. To optimize performance and prevent overfitting, feature selection was subsequently performed using the SelectFromModel method based on the GradientBoostingRegressor. The initial model was trained using all cleaned features. Then, SelectFromModel was employed with a threshold='median' parameter to select only the features whose feature importance was above the median importance value. This selection process yielded twenty-nine features, which were deemed the most crucial in influencing the band gap value. This entire procedure implicitly balances the number of features and the complexity of the predictive model. From a total dataset of 3,026 entries, the data were partitioned into two subsets\u0026mdash;the training set and the test set\u0026mdash;using a ratio of 20:80. As shown in the Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, this allocation resulted in 2,420 samples for training and 606 samples for testing, respectively.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe dataset was partitioned using a stratified sampling technique, specifically the StratifiedShuffleSplit algorithm. This method ensures that the distribution of the target variable (band gap) remains consistent and balanced across both subsets. Given that band gap is a continuous variable, a binning process was applied to discretize the values into six distinct quantiles using the pd.qcut function. To ensure statistical stability, only bins containing more than ten data points were retained. This stratification strategy was implemented to ensure that the model receives a representative distribution for both training and testing, thereby mitigating bias toward specific bandgap ranges. This is particularly crucial for materials data, where bandgap values can vary extensively depending on the underlying elemental composition\u003c/p\u003e \u003cp\u003eThe top eight features that were most influential in predicting the band gap values of binary compounds were identified using the Gradient Boosting Regressor model, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. The values displayed represent the relative importance of each feature in contributing to the model's performance in predicting the target variable bandgap. A higher importance value indicates that the feature plays a more significant role in shaping the model's decision-making process during bandgap prediction.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows that the formation energy per atom provides the highest contribution (approximately 22%) to the model's predictive performance. The observed trend aligns with the principle that thermodynamic stability, indicated by formation energy, is intrinsically linked to bond strength and the resulting energy band gap. In materials with high ionic character, large electronegativity differences lead to strong electron localization at the anion sites, resulting in a wide band gap. In contrast, materials with lower formation energies tend toward covalency and electron delocalization, which narrows the gap. Stronger bonds yield greater orbital splitting between bonding and antibonding states, inherently increasing the band gap. This strong interdependence between formation energy and the band gap has been demonstrated by Al-Qahtani et al. in their DFT study of highly stable phases of inorganic ABX\u003csub\u003e3\u003c/sub\u003e halides[30].\u003c/p\u003e \u003cp\u003eA robust correlation between atomic density and the energy band gap has been extensively documented in prior literature. The comprehensive compilation of semiconductor parameters for III-V alloys by Vurgaftman et al. provides systematic evidence of how the lattice parameter\u0026mdash;and by extension, the structural density\u0026mdash;governs electronic properties[31]. High atomic density implies a reduced interatomic spacing, which facilitates a more pronounced overlap of neighboring atomic orbitals. This intensified orbital overlap causes discrete atomic energy levels to broaden into energy bands. Consequently, as the valence and conduction bands widen due to increased overlap, the forbidden energy region between them, known as the band gap, concurrently diminishes.\u003c/p\u003e \u003cp\u003eThe Mendeleev Number (MN) serves as a one-dimensional index assigned to chemical elements to sequentially categorize them based on similarities in electronegativity, atomic radius, and valence electrons[32], [33]. A heightened electronegativity gradient between constituent atoms, such as in Ga-As bonds, strengthens the ionic character of the bond. This disparity enhances the periodic potential, which effectively widens the separation between the valence and conduction bands. Conversely, a smaller atomic radius facilitates reduced bond lengths, leading to a robust overlap of neighboring atomic orbitals. Furthermore, the configuration of valence electrons (e.g., s, p, d orbitals) dictates the band filling, symmetry, and hybridization strength. For instance, the formidable covalent bonding resulting from sp\u003csup\u003e3\u003c/sup\u003e hybridization, as observed in diamond, necessitates higher excitation energy, thereby leading to an increased energy band gap[34].\u003c/p\u003e \u003cp\u003eThe space group number represents the crystalline symmetry of a material. This symmetry dictates the arrangement and interaction of energy bands across the Brillouin zone, playing a crucial role in the formation of band gaps. Consequently, even in composition-based predictive models, space group information remains a critical factor in determining the material's band gap value[35]. The covalent radius of constituent elements also affects electron and hole localization. Larger covalent radii favor hole localization, while smaller radii facilitate electron localization, both of which drive changes in the energy band structure and the resulting band gap[36].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFollowing the initial training phase, feature selection was performed using the SelectFromModel method with a 'median' threshold. Under this approach, only features with importance scores exceeding the median value of all features were retained. This process effectively mitigates overfitting and enhances model interpretability. Consequently, twenty-nine optimal features were identified for the predictive model. After feature selection, the model was retrained to achieve superior accuracy. Figure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e illustrates the model's performance by comparing the ground truth bandgap values against the predicted values for both the training and testing sets. Following back-transformation via np.expm1, the model's predictions were evaluated against actual values. Performance on the training set yielded an MAE of 0.087 eV, an RMSE of 0.118 eV, and an R\u003csup\u003e2\u003c/sup\u003e of 0.983. This high R\u003csup\u003e2\u003c/sup\u003e value signifies that the model explains 98.3% of the training data variance, reflecting a robust fit during the learning phase.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFurthermore, the model was evaluated using a testing dataset\u0026mdash;comprising data unseen during the training phase\u0026mdash;to assess its generalizability. The evaluation yielded a Mean Absolute Error (MAE) of 0.293 eV, a Root Mean Squared Error (RMSE) of 0.433 eV, and an R-squared (R\u003csup\u003e2\u003c/sup\u003e) value of 0.774. In evaluating regression performance, the coefficient of determination (R\u003csup\u003e2\u003c/sup\u003e) serves as a critical metric, representing the proportion of variance in the target variable (the band gap) that can be explained by the input features. R\u003csup\u003e2\u003c/sup\u003e values range from 0 to 1, where a value approaching 1 signifies superior predictive capability.\u003c/p\u003e \u003cp\u003eIn the fields of physics and chemistry, R\u003csup\u003e2\u003c/sup\u003e values between 0.70 and 0.99 are generally regarded as indicative of high-quality results. This expectation arises because physical data are typically more deterministic and exhibit stronger physical correlations between variables[37]. The distribution of the band gap values predicted by the machine learning model, alongside the corresponding ground truth values, is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e. The model developed in this study achieved an \u003cspan\u003e$\u003c/span\u003eR^2\u003cspan\u003e$\u003c/span\u003e value of 0.774 on the testing set, indicating that approximately 77.4% of the variance in the band gap values can be explained by the input features. The observed decline in performance on the testing data suggests a degree of slight overfitting, where the model may have overfitted to the training set, thereby limiting its generalizability to unseen data. This is likely since composition-based features alone may not fully capture the intricate electronic structures of materials, leaving more complex physical and structural properties underrepresented. Nevertheless, the GBR model developed here serves as a viable initial approach for predicting the band gap of binary materials based on their chemical composition.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e illustrates the correlation between predicted and actual band gap values, labeled with chemical formulas and absolute errors. The results are displayed via scatter plots for the training and testing phases, where the y\u0026thinsp;=\u0026thinsp;x dashed line represents ideal prediction. Most testing points lie in close proximity to the reference line, facilitating the identification of significant outliers. While the overall distribution confirms the model's accuracy, the presence of specific outliers suggests the existence of unique electronic properties or complexities that transcend the limitations of a purely compositional approach.\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 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of the GBR-predicted band gaps and the actual values\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" 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=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMaterial ID\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFormula\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eActual Bandgap (eV)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eBandgap Predicted by this model (eV)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsolute Error\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-1205297\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTb\u003csub\u003e2\u003c/sub\u003eS\u003csub\u003e3\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7331\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7329\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0002\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-556784\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZnS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.9920\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.9910\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0010\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-680145\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCdI\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.3186\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.3175\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0011\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-1038759\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMg\u003csub\u003e2\u003c/sub\u003eBi\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8771\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8755\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0016\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-27194\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSnI\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.7061\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.7080\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0019\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-624418\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCdI\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.3651\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.3677\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0026\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-422\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBeS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.1453\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3.1491\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0038\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-567296\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCdI\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.3218\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.3258\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0040\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-776402\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eB\u003csub\u003e10\u003c/sub\u003eH\u003csub\u003e13\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.2507\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3.2548\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0041\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-650112\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFe\u003csub\u003e3\u003c/sub\u003eO\u003csub\u003e4\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9248\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9206\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0042\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-624399\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCdI\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.3414\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.3461\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0047\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-1245096\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV\u003csub\u003e2\u003c/sub\u003eO\u003csub\u003e5\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7321\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7273\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0048\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-680205\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePbI\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.3541\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.3485\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0054\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-27267\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReP\u003csub\u003e4\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8553\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8488\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0065\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-10009\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGaTe\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7853\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7919\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0066\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-1080367\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCeSe\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9720\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9787\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0067\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-759097\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVF\u003csub\u003e3\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.6507\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.6440\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0067\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-1244954\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAl\u003csub\u003e2\u003c/sub\u003eO\u003csub\u003e3\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.5224\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3.5297\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0073\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-1080356\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCeSe\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9568\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9648\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0080\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-22009\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePbSe\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.2995\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.2910\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0085\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-1196206\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eC\u003csub\u003e9\u003c/sub\u003eCl\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.3653\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.3739\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0085\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-1279995\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMn\u003csub\u003e3\u003c/sub\u003eO\u003csub\u003e4\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8312\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8310\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0087\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-759504\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFe\u003csub\u003e10\u003c/sub\u003eO\u003csub\u003e11\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.3086\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.3179\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0093\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-22398\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFeF\u003csub\u003e3\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.1061\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3.1156\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0095\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-2371\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGa\u003csub\u003e2\u003c/sub\u003eTe\u003csub\u003e5\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7231\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7135\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0094\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-1714\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFeSi\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6966\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7067\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0101\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-680163\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCdI\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.3597\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.3486\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0111\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-2667\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCsAu\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.0233\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.0344\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0111\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-680093\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCdI\u003csub\u003e2\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.3352\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.3468\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0116\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emp-561153\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eXeF\u003csub\u003e3\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.6269\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2.6148\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.0121\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\u003eA comparative analysis between GBR-predicted band gaps and measured values is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The minimal error observed indicates the high precision of the GBR machine learning model in predicting band gaps. While the current application is limited to binary compounds, we believe the model possesses the robustness to yield accurate predictions for diverse material classes.\u003c/p\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003eThis study successfully developed an efficient machine learning model to predict the bandgap values of binary compounds based on chemical composition using the Gradient Boosting Regressor (GBR) algorithm. Evaluation results on the training dataset demonstrated exceptional performance, achieving an MAE of 0.087 eV, an RMSE of 0.118 eV, and an R\u003csup\u003e2\u003c/sup\u003e of 98.3%, indicating the model\u0026rsquo;s robust ability to capture the relationship between composition and bandgap with high precision. However, when evaluated on unseen data, the performance decreased, yielding an R\u003csup\u003e2\u003c/sup\u003e of 0.774. This decline suggests the presence of overfitting, likely due to the limitations of composition-based features in fully representing the complexity of electronic structures. Furthermore, the presence of outliers\u0026mdash;specific materials with unique electronic characteristics\u0026mdash;also influenced the predictive accuracy. Despite these challenges, the model\u0026rsquo;s performance remains well within the acceptable range for materials science applications. Visualization of the predicted versus actual values shows that most data points align closely with the reference line (y\u0026thinsp;=\u0026thinsp;x) with minimal error, notwithstanding the outliers that reflect the distinct properties of certain materials.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eConflict of interest\u003c/h2\u003e \u003cp\u003eThe authors declare no conflict of interest\u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThe authors gratefully acknowledge the financial support provided by Universitas Brawijaya through research grant contract number : 02115.52/UN10.F0/901/B/PG/2025\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eMauludi Ariesto Pamungkas established the conceptual framework, performed the formal analysis, and developed the research methodology. He was also responsible for securing resources, providing overall supervision, and leading the manuscript's review and editing process.Annida Early Rofiah developed the software and machine learning models, generated all data visualizations, and prepared the original draft of the manuscript.Agus Naba contributed through technical supervision and assisted in the final editing of the paper.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data underlying this article will be shared on reasonable request to the corresponding author. The data that support the findings of this study are openly available in Materials Project\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eO. Guesmi, M. Ben Arbia, F. Saidi, M. Ben Rabeh, and H. Maaref, \u0026ldquo;Experimental and computational studies of CCTS material: Novel design of solar cell for high efficiency,\u0026rdquo; \u003cem\u003eSolar Energy\u003c/em\u003e, vol. 262, p. 111806, Sep. 2023, doi: 10.1016/j.solener.2023.111806.\u003c/li\u003e\n\u003cli\u003eB. R. Sutherland, \u0026ldquo;Solar Materials Find Their Band Gap,\u0026rdquo; \u003cem\u003eJoule\u003c/em\u003e, vol. 4, no. 5, pp. 984\u0026ndash;985, May 2020, doi: 10.1016/j.joule.2020.05.001.\u003c/li\u003e\n\u003cli\u003eN. Papež, R. Dallaev, Ş. Ţălu, and J. Ka\u0026scaron;tyl, \u0026ldquo;Overview of the Current State of Gallium Arsenide-Based Solar Cells,\u0026rdquo; \u003cem\u003eMaterials\u003c/em\u003e, vol. 14, no. 11, p. 3075, Jun. 2021, doi: 10.3390/ma14113075.\u003c/li\u003e\n\u003cli\u003eC. Battaglia, A. Cuevas, and S. De Wolf, \u0026ldquo;High-efficiency crystalline silicon solar cells: status and perspectives,\u0026rdquo; \u003cem\u003eEnergy Environ. Sci.\u003c/em\u003e, vol. 9, no. 5, pp. 1552\u0026ndash;1576, 2016, doi: 10.1039/C5EE03380B.\u003c/li\u003e\n\u003cli\u003eY. Wang \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;A review of gallium phosphide nanophotonics towards omnipotent nonlinear devices,\u0026rdquo; \u003cem\u003eNanophotonics\u003c/em\u003e, vol. 13, no. 18, pp. 3207\u0026ndash;3252, Aug. 2024, doi: 10.1515/nanoph-2024-0172.\u003c/li\u003e\n\u003cli\u003eS. Jiang \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Study on Light Extraction from GaN-based Green Light-Emitting Diodes Using Anodic Aluminum Oxide Pattern and Nanoimprint Lithography,\u0026rdquo; \u003cem\u003eSci Rep\u003c/em\u003e, vol. 6, no. 1, p. 21573, Feb. 2016, doi: 10.1038/srep21573.\u003c/li\u003e\n\u003cli\u003eH. D. Jabbar, M. A. Fakhri, and M. Jalal AbdulRazzaq, \u0026ldquo;Gallium Nitride \u0026ndash;Based Photodiode: A review,\u0026rdquo; \u003cem\u003eMaterials Today: Proceedings\u003c/em\u003e, vol. 42, pp. 2829\u0026ndash;2834, 2021, doi: 10.1016/j.matpr.2020.12.729.\u003c/li\u003e\n\u003cli\u003eA. Aldalbahi \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;A new approach for fabrications of SiC based photodetectors,\u0026rdquo; \u003cem\u003eSci Rep\u003c/em\u003e, vol. 6, no. 1, p. 23457, Mar. 2016, doi: 10.1038/srep23457.\u003c/li\u003e\n\u003cli\u003eM. Banari, N. Memarian, I. Concina, and A. Vomiero, \u0026ldquo;UV photodetector study based on Ce: ZnO nanostructures with different concentration of Ce dopant,\u0026rdquo; \u003cem\u003eOptical Materials\u003c/em\u003e, vol. 146, p. 114576, Dec. 2023, doi: 10.1016/j.optmat.2023.114576.\u003c/li\u003e\n\u003cli\u003eS. M. Abd El-Azeem and S. M. El-Ghanam, \u0026ldquo;Comparative study of gallium nitride and silicon carbide MOSFETs as power switching applications under cryogenic conditions,\u0026rdquo; \u003cem\u003eCryogenics\u003c/em\u003e, vol. 107, p. 103071, Apr. 2020, doi: 10.1016/j.cryogenics.2020.103071.\u003c/li\u003e\n\u003cli\u003eQ. Yu, \u0026ldquo;Comparative Analysis of Sic and Gan: Third-Generation Semiconductor Materials,\u0026rdquo; \u003cem\u003eHSET\u003c/em\u003e, vol. 81, pp. 484\u0026ndash;490, Jan. 2024, doi: 10.54097/2q3qyj85.\u003c/li\u003e\n\u003cli\u003eM. I. Idris and A. B. Horsfall, \u0026ldquo;3D structures for silicon carbide transistors utilising Al 2 O 3 as a gate dielectric,\u0026rdquo; \u003cem\u003eMaterials Science in Semiconductor Processing\u003c/em\u003e, vol. 128, p. 105727, Jun. 2021, doi: 10.1016/j.mssp.2021.105727.\u003c/li\u003e\n\u003cli\u003eP. H. M. Andrade, C. Volkringer, T. Loiseau, A. Tejeda, M. Hureau, and A. Moissette, \u0026ldquo;Band gap analysis in MOF materials: Distinguishing direct and indirect transitions using UV\u0026ndash;vis spectroscopy,\u0026rdquo; \u003cem\u003eApplied Materials Today\u003c/em\u003e, vol. 37, p. 102094, Apr. 2024, doi: 10.1016/j.apmt.2024.102094.\u003c/li\u003e\n\u003cli\u003eV. Mishra \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Diffuse reflectance spectroscopy: An effective tool to probe the defect states in wide band gap semiconducting materials,\u0026rdquo; \u003cem\u003eMaterials Science in Semiconductor Processing\u003c/em\u003e, vol. 86, pp. 151\u0026ndash;156, Nov. 2018, doi: 10.1016/j.mssp.2018.06.025.\u003c/li\u003e\n\u003cli\u003eN. Roisin, M.-S. Colla, R. Scaffidi, T. Pardoen, D. Flandre, and J.-P. Raskin, \u0026ldquo;Band gap reduction in highly-strained silicon beams predicted by first-principles theory and validated using photoluminescence spectroscopy,\u0026rdquo; \u003cem\u003eOptical Materials\u003c/em\u003e, vol. 144, p. 114347, Oct. 2023, doi: 10.1016/j.optmat.2023.114347.\u003c/li\u003e\n\u003cli\u003eX. Zheng, A. J. Cohen, P. Mori-S\u0026aacute;nchez, X. Hu, and W. Yang, \u0026ldquo;Improving Band Gap Prediction in Density Functional Theory from Molecules to Solids,\u0026rdquo; \u003cem\u003ePhys. Rev. Lett.\u003c/em\u003e, vol. 107, no. 2, p. 026403, Jul. 2011, doi: 10.1103/PhysRevLett.107.026403.\u003c/li\u003e\n\u003cli\u003eI. Jihad, M. H. S. Anfa, S. M. Alqahtani, and F. H. Alharbi, \u0026ldquo;DFT-PBE band gap correction using machine learning with a reduced set of features,\u0026rdquo; \u003cem\u003eComputational Materials Science\u003c/em\u003e, vol. 244, p. 113153, Sep. 2024, doi: 10.1016/j.commatsci.2024.113153.\u003c/li\u003e\n\u003cli\u003eT. Ko, T. Park, M. Kim, and K. Min, \u0026ldquo;Enhancing predictions of experimental band gap using machine learning and knowledge transfer,\u0026rdquo; \u003cem\u003eMaterials Today Communications\u003c/em\u003e, vol. 41, p. 110717, Dec. 2024, doi: 10.1016/j.mtcomm.2024.110717.\u003c/li\u003e\n\u003cli\u003eP. R. Regonia, C. M. Pelicano, R. Tani, A. Ishizumi, H. Yanagi, and K. Ikeda, \u0026ldquo;Predicting the band gap of ZnO quantum dots via supervised machine learning models,\u0026rdquo; \u003cem\u003eOptik\u003c/em\u003e, vol. 207, p. 164469, Apr. 2020, doi: 10.1016/j.ijleo.2020.164469.\u003c/li\u003e\n\u003cli\u003eR. Huang, C. McMahan, B. Herrin, A. McLain, B. Cai, and S. Self, \u0026ldquo;Gradient boosting: A computationally efficient alternative to Markov chain Monte Carlo sampling for fitting large Bayesian spatio-temporal binomial regression models,\u0026rdquo; \u003cem\u003eInfectious Disease Modelling\u003c/em\u003e, vol. 10, no. 1, pp. 189\u0026ndash;200, Mar. 2025, doi: 10.1016/j.idm.2024.09.008.\u003c/li\u003e\n\u003cli\u003eM. Seyedrahimi-Niaraq, H. Mahdiyanfar, and M. H. Olyaee, \u0026ldquo;Comparison of the performance of gradient boost, linear regression, decision tree, and voting algorithms to separate geochemical anomalies areas in the fractal environment,\u0026rdquo; \u003cem\u003eArtificial Intelligence in Geosciences\u003c/em\u003e, vol. 6, no. 2, p. 100156, Dec. 2025, doi: 10.1016/j.aiig.2025.100156.\u003c/li\u003e\n\u003cli\u003eL. F. M. Sepulveda \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Forecasting of individual electricity consumption using Optimized Gradient Boosting Regression with Modified Particle Swarm Optimization,\u0026rdquo; \u003cem\u003eEngineering Applications of Artificial Intelligence\u003c/em\u003e, vol. 105, p. 104440, Oct. 2021, doi: 10.1016/j.engappai.2021.104440.\u003c/li\u003e\n\u003cli\u003eD. A. Otchere, T. O. A. Ganat, J. O. Ojero, B. N. Tackie-Otoo, and M. Y. Taki, \u0026ldquo;Application of gradient boosting regression model for the evaluation of feature selection techniques in improving reservoir characterisation predictions,\u0026rdquo; \u003cem\u003eJournal of Petroleum Science and Engineering\u003c/em\u003e, vol. 208, p. 109244, Jan. 2022, doi: 10.1016/j.petrol.2021.109244.\u003c/li\u003e\n\u003cli\u003eD. Yang \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Bandgap prediction and design of halide double perovskites via ensemble machine learning,\u0026rdquo; \u003cem\u003eChemical Engineering Journal\u003c/em\u003e, vol. 525, p. 169979, Dec. 2025, doi: 10.1016/j.cej.2025.169979.\u003c/li\u003e\n\u003cli\u003eF. J. Kusuma \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Direct band gap prediction of single and double perovskite using cost-sensitive ensemble learning,\u0026rdquo; \u003cem\u003eJournal of Alloys and Compounds\u003c/em\u003e, vol. 1037, p. 182102, Aug. 2025, doi: 10.1016/j.jallcom.2025.182102.\u003c/li\u003e\n\u003cli\u003eH. He, Y. Wang, Y. Qi, Z. Xu, Y. Li, and Y. Wang, \u0026ldquo;From prediction to design: Recent advances in machine learning for the study of 2D materials,\u0026rdquo; \u003cem\u003eNano Energy\u003c/em\u003e, vol. 118, p. 108965, Dec. 2023, doi: 10.1016/j.nanoen.2023.108965.\u003c/li\u003e\n\u003cli\u003eM. Alsalman, S. M. Alqahtani, and F. H. Alharbi, \u0026ldquo;Bandgap energy prediction of senary zincblende III\u0026ndash;V semiconductor compounds using machine learning,\u0026rdquo; \u003cem\u003eMaterials Science in Semiconductor Processing\u003c/em\u003e, vol. 161, p. 107461, Jul. 2023, doi: 10.1016/j.mssp.2023.107461.\u003c/li\u003e\n\u003cli\u003eH. Lamouadene, M. El Kassaoui, M. El Yadari, A. El Kenz, and A. Benyoussef, \u0026ldquo;Exploring modeling techniques for predicting band gaps of Doped-ZnO: A Machine learning approach,\u0026rdquo; \u003cem\u003eChemical Physics\u003c/em\u003e, vol. 591, p. 112603, Mar. 2025, doi: 10.1016/j.chemphys.2025.112603.\u003c/li\u003e\n\u003cli\u003eM. H. Saeed \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Determination of bandgap of period 3, 4, and 5 transition metal dopants on zinc oxide using an artificial neural network based approach,\u0026rdquo; \u003cem\u003eChemometrics and Intelligent Laboratory Systems\u003c/em\u003e, vol. 242, p. 104983, Nov. 2023, doi: 10.1016/j.chemolab.2023.104983.\u003c/li\u003e\n\u003cli\u003eS. M. Alqahtani, A. Q. Alsayoud, and F. H. Alharbi, \u0026ldquo;Structures, band gaps, and formation energies of highly stable phases of inorganic ABX\u003csub\u003e3\u003c/sub\u003e halides: A = Li, Na, K, Rb, Cs, Tl; B = Be, Mg, Ca, Ge, Sr, Sn, Pb; and X = F, Cl, Br, I,\u0026rdquo; \u003cem\u003eRSC Adv.\u003c/em\u003e, vol. 13, no. 13, pp. 9026\u0026ndash;9032, 2023, doi: 10.1039/D3RA00185G.\u003c/li\u003e\n\u003cli\u003eI. Vurgaftman, J. R. Meyer, and L. R. Ram-Mohan, \u0026ldquo;Band parameters for III\u0026ndash;V compound semiconductors and their alloys,\u0026rdquo; \u003cem\u003eJournal of Applied Physics\u003c/em\u003e, vol. 89, no. 11, pp. 5815\u0026ndash;5875, Jun. 2001, doi: 10.1063/1.1368156.\u003c/li\u003e\n\u003cli\u003eF. R. Boer and D. G. Pettifor, \u003cem\u003eThe Structures of binary compounds\u003c/em\u003e. in Cohesion and structure, no. v. 2. Amsterdam New York: North-Holland Sole distributors for the USA and Canada, Elsevier Science Pub. Co, 1989.\u003c/li\u003e\n\u003cli\u003eP. Villars, K. Cenzual, J. Daams, Y. Chen, and S. Iwata, \u0026ldquo;Data‐Driven Atomic Environment Prediction for Binaries Using the Mendeleev Number. Part 1. Composition AB.,\u0026rdquo; \u003cem\u003eChemInform\u003c/em\u003e, vol. 35, no. 23, p. chin.200423004, Jun. 2004, doi: 10.1002/chin.200423004.\u003c/li\u003e\n\u003cli\u003eJ. C. Phillips, \u0026ldquo;Ionicity of the Chemical Bond in Crystals,\u0026rdquo; \u003cem\u003eRev. Mod. Phys.\u003c/em\u003e, vol. 42, no. 3, pp. 317\u0026ndash;356, Jul. 1970, doi: 10.1103/RevModPhys.42.317.\u003c/li\u003e\n\u003cli\u003eH. Watanabe and L. Lu, \u0026ldquo;Space-group theory of photonic bands,\u0026rdquo; \u003cem\u003ePhys. Rev. Lett.\u003c/em\u003e, vol. 121, no. 26, p. 263903, Dec. 2018, doi: 10.1103/PhysRevLett.121.263903.\u003c/li\u003e\n\u003cli\u003eM. P. Polak, P. Scharoch, and R. Kudrawiec, \u0026ldquo;The effect of isovalent doping on the electronic band structure of group IV semiconductors,\u0026rdquo; \u003cem\u003eJ. Phys. D: Appl. Phys.\u003c/em\u003e, vol. 54, no. 8, p. 085102, Feb. 2021, doi: 10.1088/1361-6463/abc503.\u003c/li\u003e\n\u003cli\u003eA. Gupta, T. S. Stead, and L. Ganti, \u0026ldquo;Determining a Meaningful R-squared Value in Clinical Medicine,\u0026rdquo; \u003cem\u003eAcademic Medicine \u0026amp; Surgery\u003c/em\u003e, Oct. 2024, doi: 10.62186/001c.125154.\u003c/li\u003e\n\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":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-computational-electronics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jcel","sideBox":"Learn more about [Journal of Computational Electronics](https://www.springer.com/journal/10825)","snPcode":"10825","submissionUrl":"https://submission.nature.com/new-submission/10825/3","title":"Journal of Computational Electronics","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Bandgap, Binary Compounds, Machine Learning, Gradient Boosting Regression","lastPublishedDoi":"10.21203/rs.3.rs-8819164/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8819164/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study developed a predictive model for the bandgap of binary materials using a machine learning approach based on the Gradient Boosting Regressor (GBR). The model showed exceptional performance on the training set, achieving a Mean Absolute Error (MAE) of 0.087, a Root Mean Squared Error (RMSE) of 0.118, and an R\u003csup\u003e2\u003c/sup\u003e of 98%. While the model's performance on the unseen test data was lower, with an R\u003csup\u003e2\u003c/sup\u003e of 77%, this still indicates a strong capability to predict bandgaps, explaining over 77% of the variation. The performance drop suggests a minor degree of overfitting to the training data. Visualization of the results shows that while most predictions align closely with the ideal reference line, a few outliers lead to significant errors for certain materials. This is likely because compositional features alone are insufficient to capture the intricate physical properties governing these materials.\u003c/p\u003e","manuscriptTitle":"Bandgap Prediction of Binary Compounds via a Machine Learning Approach Utilizing Gradient Boosting Regression","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-13 10:48:52","doi":"10.21203/rs.3.rs-8819164/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-16T15:40:24+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-23T08:12:02+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"154773683928492698964308088894884976774","date":"2026-02-12T19:06:36+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-10T08:11:09+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-09T13:30:46+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-09T13:29:58+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Computational Electronics","date":"2026-02-08T04:41:25+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-computational-electronics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jcel","sideBox":"Learn more about [Journal of Computational Electronics](https://www.springer.com/journal/10825)","snPcode":"10825","submissionUrl":"https://submission.nature.com/new-submission/10825/3","title":"Journal of Computational Electronics","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"0a85fccb-5409-4bf8-a9a1-d2cf4619b37d","owner":[],"postedDate":"February 13th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-05-04T16:03:14+00:00","versionOfRecord":{"articleIdentity":"rs-8819164","link":"https://doi.org/10.1007/s10825-026-02547-y","journal":{"identity":"journal-of-computational-electronics","isVorOnly":false,"title":"Journal of Computational Electronics"},"publishedOn":"2026-04-27 15:57:47","publishedOnDateReadable":"April 27th, 2026"},"versionCreatedAt":"2026-02-13 10:48:52","video":"","vorDoi":"10.1007/s10825-026-02547-y","vorDoiUrl":"https://doi.org/10.1007/s10825-026-02547-y","workflowStages":[]},"version":"v1","identity":"rs-8819164","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8819164","identity":"rs-8819164","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}