AI-enabled Monitoring of Planar-Flow Casting of Amorphous Steel Foils

AI-enabled Monitoring of Planar-Flow Casting of Amorphous Steel Foils | 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 Article AI-enabled Monitoring of Planar-Flow Casting of Amorphous Steel Foils Ritaban Dutta, Michael Kellam, Daniel Liang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8042276/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Amorphous metallic foils produced by planar-flow casting (PFC) are central to emerging high-efficiency transformer and motor technologies. Despite its promise, the process remains highly sensitive to fluctuations in melt delivery, nozzle–wheel gap, and interfacial heat transfer, leading to free-surface ridging, uneven thickness, and compromised magnetic properties. Conventional monitoring, based on indirect variables or post-cast inspection, provides delayed feedback and cannot resolve the rapid dynamics of the melt–wheel interface. Here we present a data-driven framework for PFC monitoring that integrates synchronised video and infrared imaging with machine-learning analysis. A multimodal dataset of ≈ 200 runs (≈ 300,000 frames, ≈ 20 GB) was collected with front- and side-view cameras and infrared thermal imaging, paired with detailed metadata. Physics-inspired descriptors—including ridge density, reflection intensity, and thermal gradients—were extracted through automated computer vision pipelines and benchmarked against raw-frame deep learning. Analytical proxies, gradient-boosted decision trees, and multimodal fusion networks were systematically evaluated. Feature-based models achieved near-perfect gap prediction, while ridge and thermal-gradient analysis provided direct indicators of foil integrity. Deep-learning fusion models offered complementary robustness but at higher computational cost. Together, these results demonstrate a reproducible approach for linking process dynamics with foil quality, establishing a foundation for automated, closed-loop optimisation of amorphous steel-foil manufacturing. Physical sciences/Engineering Physical sciences/Materials science Physical sciences/Mathematics and computing Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Amorphous metallic alloys combine high strength, corrosion resistance, and soft-magnetic behaviour, making them attractive for energy-efficient transformers, electric motors, and advanced electronic applications. The ability to manufacture thin amorphous foils at scale is central to their industrial adoption. Planar-flow casting (PFC) is the most practical route for producing such foils, directly quenching molten alloy onto a rotating copper wheel to form continuous ribbons [ 1 – 13 ]. Despite its promise, PFC is inherently unstable. Small perturbations in melt temperature, nozzle–wheel gap, or wheel speed can destabilise the melt pool, disrupt heat transfer at the wheel interface, and generate surface ridges or thickness variations. These instabilities not only compromise foil integrity but also degrade the magnetic properties critical for industrial use. Conventional monitoring relies on bulk process variables or post-process inspection, neither of which captures the fast spatio-temporal dynamics of the melt–wheel interaction where instabilities originate [ 14 – 20 ]. High-speed video and thermal imaging now allow direct observation of melt delivery and foil formation. However, such data are high dimensional and require systematic analysis methods. Computer vision (CV) can extract interpretable descriptors—ridge density, reflection patterns, and thermal gradients—while machine learning (ML) can model their relationship to process variables and foil quality. Deep learning (DL) architectures extend this further by learning spatio-temporal representations from raw multimodal frame sequences. In this work, we introduce an integrated framework for analysing more than 200 PFC experiments using synchronised RGB and infrared video paired with detailed metadata. A structured dataset was created, annotated with process variables and quality descriptors, and analysed through both feature-based ML models and multimodal DL fusion networks. Baseline analytical proxies were benchmarked against these approaches to establish lower- and upper-bound performance. In addition, ridge-width distributions and thermal-gradient profiles were evaluated as direct video-derived measures of foil integrity [ 21 – 30 ]. The results, summarised in Figs. 1 – 5 , Supplementary Figs. 6–7 and Tables 1 – 3 , Supplementary Tables S1-S5, demonstrate how CV- and ML-based approaches can provide accurate, interpretable, and scalable monitoring of PFC. Together, these contributions establish a pathway toward reproducible and automated optimisation of amorphous steel-foil manufacturing, and lay the groundwork for industrial deployment in next-generation casting facilities [ 30 – 38 ]. Methods Dataset description A multimodal dataset was assembled from approximately 200 planar-flow casting (PFC) experiments performed under varied operating conditions. Each run included synchronised video recordings from front-view and side-view RGB cameras, with a subset also incorporating infrared (IR) thermal imaging of the emerging foil. In total, the dataset contains ≈ 300,000 frames, corresponding to ≈ 20 GB of structured video data. Each run was assigned a unique identifier and linked to a manifest file containing metadata. Recorded fields included nozzle–wheel gap, wheel speed, melt temperature, nozzle geometry, and flow proxy variables. This structure ensured that every video frame could be directly associated with corresponding process conditions [ 1 – 11 ]. The dataset spans both stable and unstable regimes, capturing variability in ridging, surface uniformity, and thermal balance. This diversity provides a foundation for developing and benchmarking predictive models that generalise across different casting conditions. A summary of dataset properties, including the number of runs, frames, experimental setups, and metadata fields, is given in Table 1 . Representative images from RGB and IR recordings are shown in Fig. 1 , while extended metadata descriptors, parameter ranges, and acquisition details are provided in Supplementary Table S1 . Experimental setup and video acquisition High-speed imaging was used to capture melt delivery, foil formation, and nozzle–wheel interactions during each casting run. Two RGB cameras were positioned to provide complementary perspectives: A front-view camera , aligned with the emerging foil, which captured ridge formation, foil width, and surface reflections [ 21 ]. A side-view camera , aligned with the nozzle–wheel interface, which recorded the gap, ribbon trajectory, and melt-pool behaviour [ 22 ]. In a subset of runs, an infrared (IR) thermal camera was mounted to capture the temperature distribution across the foil width during solidification. These IR recordings enabled frame-by-frame measurement of edge-to-centre thermal gradients, which are linked to foil stability and amorphicity [ 23 ]. All videos were acquired at 30 frames per second, with resolutions sufficient to resolve ridge spacing, reflection patterns, and foil width variations. Camera alignment was verified before each experiment, and lighting was adjusted to minimise glare and highlight surface contrast [ 24 ]. Each casting run was indexed in the dataset by its unique identifier, with video files stored in a structured hierarchy and cross-referenced with metadata logs. This arrangement allows frame-level synchronisation of visual and process information. Representative frames from the front- and side-view cameras, as well as IR recordings, are shown in Fig. 1 . Acquisition settings and camera specifications are provided in Supplementary Table S2. Ground-truth measurement and labelling Direct ground-truth measurements were collected for each casting run to ensure that predictive modelling could be benchmarked against reliable process variables. The nozzle–wheel gap was measured prior to casting using precision micrometre gauges. These values were logged in the manifest file and treated as ground-truth gap values [ 25 ]. After each run, nozzle alignment was re-checked to confirm stability and rule out drift or deformation effects. The flow proxy was derived from operator logs of melt-head height and reservoir pressure, which influence volumetric melt delivery. In selected runs, this parameter was cross validated against the measured foil mass per unit time. Although variation in flow was relatively small compared with gap, it provides an important secondary ground-truth variable [ 26 ]. Foil-quality metrics were generated directly from video recordings using automated computer-vision pipelines. From front-view recordings, ridge positions, ridge density, and ridge-width distributions were extracted. From infrared recordings, frame-level temperature profiles and cross-foil thermal gradients were calculated. By automating these annotations, reproducibility was maintained across all ≈ 300,000 frames, and subjective bias was eliminated [ 27 ]. Summary statistics of the measured process variables are provided in Table 2 . Extended distributions of gap and flow values across all runs are given in Supplementary Table S3, while cross-run variability and stability ranges are detailed in Supplementary Table S4 [ 28 ]. Annotation and labelling Each casting run was indexed by a unique identifier linking raw video, metadata, and derived annotations. Frame-level annotations were generated to ensure consistency across the dataset and reproducibility of downstream modelling. For process variables, nozzle–wheel gap and flow proxy values were assigned frame by frame using the manifest file. These values were aligned temporally with the RGB and IR recordings, allowing supervised learning without leakage between training and validation sets [ 29 ]. For foil quality, annotations were obtained using automated computer-vision pipelines. From front-view videos, ridge positions were identified using edge-detection algorithms, producing ridge-density maps and ridge-width histograms. From IR recordings, temperature gradients were computed across the foil width to capture edge–centre imbalances [ 30 ]. An illustration of these annotation workflows is provided in Fig. 1 . All annotations were stored in tabular format, aligned with the manifest file. This ensured that every frame had a consistent set of labels linking raw imagery, metadata, and derived descriptors. A consolidated summary of annotation categories is provided in Table 2 , while the detailed schema, including annotation fields and feature formats, is presented in Supplementary Table S5. Feature extraction and ridge/thermal analysis To transform raw video into structured input for modelling, a suite of physics-inspired descriptors was extracted from both RGB and infrared recordings. These descriptors capture the surface, geometric, and thermal characteristics most closely associated with casting stability and foil integrity. From front-view videos, features included ridge density, ridge spacing, ridge-width distributions, and their variation across foil width. Reflection intensity and foil width were quantified from brightness and edge contrast, providing proxies for melt stability. From side-view videos, reflection height and ribbon trajectory features were derived to capture nozzle–wheel gap behaviour. From infrared recordings, cross-foil thermal profiles were analysed to compute temperature gradients between edges and centre. These measures provide a sensitive indicator of uneven cooling, which often correlates with ridge clustering and the loss of amorphicity. All features were aggregated into frame-level datasets linked to casting run identifiers. For each run, this produced ≈ 27 structured descriptors per sample, forming the input for machine-learning and deep-learning models [ 31 – 34 ]. An overview of feature categories and their physical interpretation is summarised in Table 2 , while extended definitions and parameter ranges are provided in Supplementary Table S5. Together, these descriptors underpin both predictive modelling of gap and flow, and direct video-based evaluation of foil quality. See Supplementary Videos S1–S2 for visual demonstrations of the automated ridge-tracking and thermal-gradient analysis workflows. Baseline models As a lower-bound benchmark, we implemented simple analytical proxies that approximate process conditions directly from video streams without advanced modelling. These baselines reflect the heuristic approaches sometimes used in industrial inspection. Two proxies were developed: Gap–reflection proxy — derived from the vertical extent of reflection intensity in side-view RGB recordings, serving as a crude estimator of nozzle–wheel gap. Infrared hot-area proxy — defined as the fraction of the foil surface above a fixed temperature threshold in IR frames, used as a surrogate for flow stability. These methods were applied frame by frame to the dataset, with outputs aligned to ground-truth gap and flow values. While they captured broad trends, they lacked sensitivity to dynamic fluctuations, especially under unstable casting conditions [ 36 – 38 ]. Performance of the baseline proxies is summarised in Table 3 , including mean absolute error (MAE), root-mean-square error (RMSE), and coefficient of determination (R²). Representative results are shown in Fig. 2 , illustrating the systematic underestimation, high residual variance, and limited explanatory power of these heuristic approaches. Machine learning models (XGBoost) To benchmark predictive performance beyond analytical proxies, we trained feature-based machine-learning models using the XGBoost framework. Input data were the structured descriptors described above, including ridge density, ridge-width statistics, reflection intensity, foil width, and thermal gradients. The dataset comprised ≈ 8,800 aggregated samples, each containing 27 descriptors per cast. To avoid frame-level leakage, group-wise cross-validation was used, ensuring that entire casting runs were held out for validation. Hyperparameters — learning rate, maximum depth, number of estimators, subsampling ratios, and regularisation terms — were tuned via grid search (Supplementary Table S4). Outputs included predictions of nozzle–wheel gap (mm) and flow proxy (% change in optical-flow velocity). Models were trained with mean absolute error (MAE) as the primary loss metric, while root-mean-square error (RMSE) and coefficient of determination (R²) were also evaluated. All training was performed with deterministic folds and fixed random seeds, as detailed in Supplementary Table S2. Performance results are reported in Table 3 , demonstrating a significant improvement over baseline proxies. Representative results, including scatter plots of true vs. predicted gap, temporal tracking, calibration analysis, and feature-importance rankings, are shown in Fig. 3 . Extended diagnostics, including fold-level variability and learning curves, are provided in Supplementary Table S3. Deep learning fusion models To evaluate whether predictive accuracy could be improved by learning directly from raw multimodal frames, we implemented deep-learning (DL) fusion models. These architectures combined visual features from RGB and IR recordings with tabular descriptors in a unified framework. The model pipeline consisted of: Convolutional neural networks (CNNs) applied to RGB and IR sequences to extract spatial features of melt-pool and foil dynamics. Multilayer perceptrons (MLPs) trained on tabular descriptors, including ridge, reflection, and thermal features. A transformer encoder for fusing spatio-temporal information across modalities and capturing frame-to-frame dependencies. Training was performed in PyTorch for up to 50 epochs with early stopping. The Adam optimiser was used with standard hyperparameters, and learning rates were tuned using validation folds. Cross-validation was again grouped by casting run to ensure generalisation. Implementation details, model configurations, and training schedules are documented in Supplementary Table S2, with validation diagnostics presented in Supplementary Table S3. Performance metrics for DL fusion models are summarised in Table 3 . Representative results are shown in Fig. 4 , including scatter plots of true vs. predicted gap, temporal predictions, calibration curves, and residual distributions. These outputs demonstrate that while DL fusion achieves accuracy comparable to XGBoost for gap prediction, it requires substantially greater resources. Evaluation metrics and protocols All models were assessed using standard regression metrics: mean absolute error (MAE), root-mean-square error (RMSE), and coefficient of determination (R²). MAE was selected as the primary evaluation metric due to its robustness against outliers and ease of interpretation in physical units. RMSE was reported to highlight the impact of larger deviations, while R² quantified overall explanatory power. Model performance was compared across three levels: Baseline proxies (reflection height and infrared hot area). Feature-based machine learning (XGBoost). Multimodal deep learning fusion (CNN + MLP + transformer). To prevent data leakage, all training and validation splits were grouped by casting run, ensuring that individual runs were never split across folds. For both ML and DL models, deterministic random seeds were used to ensure reproducibility. Extended details of fold assignments, hyperparameter ranges, and training schedules are given in Supplementary Table S3 . A consolidated summary of model performance is provided in Table 3 , while detailed diagnostics, including per-fold statistics and error distributions, are reported in Supplementary Tables S3 and S4. Representative scatter, calibration, and residual plots are included in Figs. 2 – 4 . Use of AI tools Language editing assistance was provided using a large language model (ChatGPT, OpenAI), used solely for grammar and clarity improvements. All content was reviewed and approved by the authors. Results Baseline model performance Analytical proxies offered a useful lower-bound for assessing the predictive complexity of PFC monitoring. The gap–reflection proxy, derived from the vertical reflection signature in side-view frames, captured only the broad scale of nozzle–wheel gap variation. Across validation runs, it achieved a mean absolute error (MAE) of 0.12 ± 0.02 mm, with an RMSE of ≈ 0.16 mm and limited explanatory power (R² in the range 0.3–0.4). Scatter plots in Fig. 2 a demonstrate that predicted values consistently fell below the 1:1 line, indicating a systematic underestimation. Temporal traces (Fig. 2 b) show that the proxy tracked slow drifts in gap but failed to resolve rapid fluctuations. Residual analysis (Fig. 2 c–d) revealed bias and heteroscedasticity, with errors increasing disproportionately at larger gaps. The infrared hot-area proxy, designed to approximate flow through the fraction of foil surface above a fixed temperature threshold, was unable to recover meaningful correlations. Predictions of the flow proxy were essentially random with respect to ground-truth labels, yielding R² ≈ 0 and MAE values like the mean of the dataset. This confirms that hot-area measures are insensitive to subtle melt-delivery changes. Performance metrics are summarised in Table 3 , while extended error distributions across runs are reported in Supplementary Table S3. Together, these results establish that heuristic proxies cannot provide reliable estimates of casting gap or flow and therefore set the lower bound for machine-learning benchmarks. Feature-based ML performance (XGBoost) The tabular dataset consisted of ≈ 8,800 aggregated samples with 27 descriptors per cast, derived from reflection, ridge, and thermal categories (Table 2 , Supplementary Table S5). XGBoost models were trained with group-wise cross-validation by casting run, ensuring no leakage between training and validation folds. For gap prediction, XGBoost achieved mean absolute error (MAE) values of 0.009–0.011 mm and RMSE ≈ 0.013 mm across folds. Coefficients of determination reached R² = 0.96–0.97, indicating that almost all variance in ground-truth gap could be explained by the engineered features. Scatter plots in Fig. 3 a show predictions clustering tightly along the 1:1 line. Temporal predictions (Fig. 3 b) tracked rapid fluctuations in gap with error bands rarely exceeding ± 0.01 mm. Calibration curves (Fig. 3 c) confirmed unbiased behaviour across the prediction range, while feature-importance rankings (Fig. 3 d) revealed that ridge-density and ridge-variation descriptors contributed most strongly, followed by reflection height and thermal gradients. For flow proxy prediction, performance remained poor. R² values consistently fell below 0.1, with MAE ≈ 2–3% of the label range. This reflects the limited variability and indirect nature of the proxy, rather than a failure of the algorithm. Extended diagnostics, including fold-level performance and feature-ablation experiments, are provided in Supplementary Sections S2–S3. Overall, these results demonstrate that physics-informed descriptors are sufficient to achieve near-perfect gap prediction, while flow remains inadequately represented by available proxy labels. Deep learning fusion results The multimodal DL fusion model combined RGB and IR frames with tabular descriptors, using convolutional encoders, MLPs, and a transformer fusion module (Supplementary Table S2). For gap prediction, DL models achieved MAE values of 0.010–0.012 mm and RMSE ≈ 0.015 mm, with R² = 0.96–0.97, closely matching XGBoost performance. Scatter plots (Fig. 4 a) showed tight alignment with the 1:1 line, while temporal traces (Fig. 4 b) demonstrated stable tracking across casting runs. Learning curves confirmed smooth convergence without overfitting. Calibration analysis (Fig. 4 c) showed strong agreement between predicted and true values across all bins, while residual distributions (Fig. 4 d) were narrow and centred close to zero. For flow proxy prediction, DL fusion again did not improve over baseline ML approaches, with R² ≈ 0 and MAE comparable to dataset averages. This underscores the inherent limitation of proxy labels, rather than the model architecture. Performance summaries are reported in Table 3 , with fold-level diagnostics included in Supplementary Table S3. The comparison demonstrates that while DL can effectively process multimodal data streams, it does not yield meaningful gains over feature-based ML in the context of gap prediction for this dataset. Foil-quality analysis (ridge width, thermal gradients) Beyond process-variable prediction, direct video-based analysis provided quantitative measures of foil integrity. Ridge and thermal descriptors extracted through computer vision pipelines were evaluated across stable and unstable casting regimes. Ridge analysis Front-view recordings were used to compute ridge-density maps and ridge-width histograms. Under stable conditions, ridge spacing was uniform across the foil, with standard deviation typically < 0.2 mm. Unstable runs exhibited ridge clustering, leading to broadened histograms with long tails. In Fig. 5 a–b, the contrast is evident: stable runs show narrow, symmetric distributions centred around ≈ 1.5 mm, while unstable runs show skewed and multimodal distributions. These ridge features were also among the highest-ranked predictors in XGBoost (Table 2 , Fig. 3 d). Thermal analysis Infrared recordings provided cross-foil temperature profiles, which were used to calculate edge-to-centre gradients. Stable runs exhibited mild, symmetric gradients of ≈ 5–10°C, while unstable runs showed steep gradients > 25°C concentrated near foil edges. In Fig. 5 c, representative thermal profiles illustrate this difference. Scatter analysis in Fig. 5 d confirmed a strong correlation between ridge clustering and thermal imbalance runs with higher ridge density consistently coincided with larger edge gradients. Combined interpretation These results establish ridge and thermal descriptors as direct, physically interpretable indicators of foil integrity. Unlike proxy-based monitoring, they allow early detection of instability through surface and thermal signatures. A summary of ridge and thermal features is included in Table 2 , with extended ranges and definitions in Supplementary Table S5. Ablation studies To evaluate the relative contribution of feature categories, systematic ablation experiments were performed on the XGBoost models. Ridge, reflection, and thermal descriptors were selectively removed, and model performance was reassessed on held out runs. Overall results With the full feature set, the model achieved MAE ≈ 0.010 mm for gap prediction. Removing ridge features caused the largest performance degradation, with MAE increasing to ≈ 0.014 mm, corresponding to a ≈ 40% relative error increase. Excluding reflection features raised MAE by ≈ 12%, while omission of thermal descriptors increased error by ≈ 20% and reduced sensitivity to foil-quality indicators. Results are summarised in Table 4. Feature importance aggregation Feature-importance scores were also aggregated across 20 representative casting runs. As shown in Fig. 6b, ridge descriptors consistently contributed the largest share of predictive power, followed by reflection and thermal features. Ridge width variation and ridge density were ranked as the top individual predictors, consistent with their strong physical relationship to gap stability. Interpretation Together, the ablation and feature-importance results confirm that all three categories provide complementary information. Ridge descriptors dominate gap estimation, reflection features add robustness, and thermal descriptors enhance sensitivity to foil quality. None of the categories is dispensable: models trained without one class of descriptors showed reduced generalisation and increased bias. Extended ablation configurations, including feature subsets and cross-run error distributions, are reported in Supplementary Table S5. Representative results are shown in Fig. 6a–b: panel (a) summarises MAE increases under different ablation conditions, while panel (b) visualises feature-importance distributions across runs. Industrial validation To assess the transferability of our framework to production environments, we analysed factory-acquired recordings provided by an industrial partner. These included RGB views of the nozzle–wheel gap region and infrared recordings of the cooling wheel and foil during continuous operation. Unlike laboratory runs, these datasets were captured under factory-scale casting conditions, where operating variability, environmental noise, and production throughput present additional challenges. RGB analysis Factory RGB frames (Supplementary Fig. 7a–b) revealed non-uniform ridge formation across foil width, with clustering events corresponding to unstable flow delivery. Applying our ridge-detection pipeline extracted density profiles comparable to those observed in unstable laboratory runs (Fig. 7c). Ridge histograms showed broadened distributions with skewed tails, confirming instability during high-throughput operation. Thermal analysis Factory IR frames captured temperature distributions across the foil and cooling wheel (Fig. 7d–e). Analysis revealed edge-dominated gradients exceeding 25–30°C, with thermal asymmetry persisting over long-time windows. These findings mirror laboratory evidence that strong edge-to-centre gradients coincide with ridge clustering and reduced amorphicity. Quantitative profiles are summarised in Supplementary Fig. 7f, showing direct alignment with the diagnostic patterns identified in controlled experiments. Interpretation and industrial significance These results demonstrate that the ridge- and thermal-analysis pipelines generalise beyond laboratory settings. In practice, factory engineers could obtain comparable thermal profiles using commercial IR cameras, but our integrated CV–ML framework enables predictive interpretation rather than descriptive monitoring. This distinction is critical: while thermal imaging provides accessible indicators, the fusion of ridge, thermal, and metadata descriptors adds robustness against noise and variability, making duplication of the predictive models from raw factory images non-trivial. By validating on factory-scale data, we confirm that the framework not only reproduces laboratory findings but also provides diagnostic indicators relevant to magnetic property stability under real industrial operating conditions. Discussion Machine learning versus deep learning approaches Our results demonstrate that feature-based ML models, particularly XGBoost trained on physics-informed descriptors, provide state-of-the-art performance for gap prediction in planar-flow casting. With MAE values below 0.01 mm and R² consistently above 0.96 (Table 3 , Figs. 3 – 4 ), these models capture nearly all variance in ground-truth gap. By contrast, deep-learning fusion models achieve similar accuracy but at substantially higher computational cost and with more complex optimisation requirements. This outcome highlights the strength of physics-inspired descriptors: they encode key process phenomena such as ridge density, reflection geometry, and thermal imbalance, which align naturally with the underlying physics of melt–wheel dynamics. The convergence of ML and DL performance further suggests that the dataset is information-rich at the feature level. In this context, XGBoost offers a lightweight, interpretable, and industrially deployable solution, while DL fusion remains valuable for extending to new modalities or for exploratory research. Ridge and thermal descriptors as quality indicators A central outcome of this study is the confirmation that ridge, and thermal descriptors serve as direct, video-derived indicators of foil integrity. Ridge-density histograms and ridge-width distributions (Fig. 5 a–b) distinguished stable from unstable runs, while thermal profiles across the foil width (Fig. 5 c–d) revealed edge-to-centre gradients that correlated strongly with surface clustering. The ablation experiments (Supplementary Fig. 6, Table 4) reinforced this conclusion: ridge descriptors accounted for ≈ 40% of predictive power, while thermal descriptors provided complementary sensitivity to quality-related phenomena. This combination provides both predictive accuracy and physical interpretability, bridging the gap between data-driven modelling and metallurgical insight. Industrial implications Validation with factory-acquired data (Supplementary Fig. 7) demonstrated that our framework generalises to production-scale environments. Ridge clustering and thermal imbalance were captured under real industrial conditions, confirming the diagnostic value of these descriptors. Importantly, while thermal imaging can be implemented with standard factory hardware, predictive interpretation requires the integrated CV–ML framework presented here. This distinction reduces the risk of straightforward duplication by external engineers and emphasises the added value of feature integration and model fusion. From an industrial perspective, reliable monitoring of ridge and thermal instabilities addresses one of the most pressing limitations of PFC: the degradation of magnetic properties under unstable casting. By linking melt–wheel dynamics directly to foil quality, our framework provides a foundation for stabilising production and ensuring that full-width amorphous foils meet performance targets. Sensitivity and duplication risks A concern raised in discussions with industrial partners is whether factory engineers could replicate the approach using only thermal imaging. While thermal profiles alone do provide useful indicators, they are inherently descriptive. The predictive performance demonstrated here arises from integrating multiple feature categories, synchronised video, and metadata across runs. Without access to the structured dataset, annotation schema (Supplementary Tables S1–S5), and trained models, reproducing this pipeline would require significant effort and expertise. Future directions While this study establishes a reproducible dataset and framework, several opportunities remain. First, a new round of factory data collection with optimised camera placements could strengthen modelling of melt-pool instabilities and wheel–melt heat transfer. Second, integrating magnetic property measurements of cast foils would allow direct linkage between process signatures and end-use performance. Third, coupling data-driven methods with physics-based Multiphysics simulations may enable hybrid models capable of real-time control. Finally, active learning strategies could adaptively select experiments to maximise model robustness across facilities and alloy compositions. Conclusion This study introduces a reproducible, data-driven framework for monitoring planar-flow casting of amorphous steel foils. By combining synchronised RGB and infrared recordings with detailed process metadata, we constructed a structured dataset of ≈ 200 casting runs. Physics-informed features derived from ridge profiles, reflection intensity, and thermal gradients provided interpretable links between melt–wheel dynamics and foil integrity. Feature-based machine-learning models, particularly XGBoost, achieved near-perfect prediction of nozzle–wheel gap with MAE < 0.01 mm and R² ≈ 0.97. Deep-learning fusion networks offered comparable performance but required greater computational resources. Baseline analytical proxies proved insufficient, confirming the value of both engineered descriptors and multimodal learning. Ridge-width distributions and thermal profiles emerged as direct indicators of foil quality, distinguishing stable and unstable casting regimes. Validation with factory-acquired data demonstrated that ridge and thermal descriptors generalise to industrial conditions. These measures address one of the most critical barriers to large-scale deployment of amorphous foils: instability-driven degradation of magnetic properties. By providing predictive, interpretable monitoring tools, this framework establishes the foundation for closed-loop, automated optimisation of PFC. Future extensions will focus on expanding datasets across facilities, capturing melt-pool and wheel–melt heat transfer in greater detail, and linking process signatures directly to magnetic performance. Together, these efforts position the approach as a practical pathway toward autonomous control of amorphous foil manufacturing. Declarations Competing Interests The authors declare no competing interests. Patent applications related to aspects of this work have been filed by CSIRO (Application No. PCT/AU2024/051282, published as WO 2025/111660; and Application No. PCT/AU2025/050062, published as WO/2025/160623). Author Contribution R.D. led the design of the study, developed the machine-learning and computer-vision framework, and drafted the manuscript. M.K. was responsible for experimental data collection and dataset assembly. D.L. contributed to experimental design, process interpretation, and manuscript refinement. All authors discussed the results, contributed to interpretation, and reviewed the manuscript. Acknowledgement This work was supported by the CSIRO Future Digital Manufacturing Fund (FDMF) and by contributions from the CSIRO Research Units (Data61 and Manufacturing). The authors thank technical staff for assistance with planar-flow casting experiments and acknowledge valuable discussions with industrial partners that informed the design of validation experiments. Data Availability The datasets generated and analysed during the current study are available from the corresponding author upon reasonable request. Custom computer-vision scripts and machine-learning pipelines developed for annotation, feature extraction, and model training are likewise available for non-commercial research purposes upon reasonable request. Data and Code Availability The datasets generated and analysed during the current study are available from the corresponding author upon reasonable request. Custom computer-vision scripts and machine-learning pipelines developed for annotation, feature extraction, and model training are likewise available for non-commercial research purposes upon reasonable request. Funding Declaration This research was supported by internal CSIRO strategic funding through the Future Digital Manufacturing Fund (FDMF) and contributions from the Data61 and Manufacturing Business Units. References Theisen, E. A. & Weinstein, S. J. An overview of planar flow casting of thin metallic glasses and its relation to slot coating of liquid films. J. Coat. Technol. Res. 19 , 49–60. https://doi.org/10.1007/s11998-021-00503-y (2022). Carpenter, J. K. & Steen, P. H. Planar-flow spin-casting of molten metals: process behaviour. J. Mater. Sci. 27 , 215–225. https://doi.org/10.1007/BF02403666 (1992). Mattson, J. W., Theisen, E. A. & Steen, P. H. Rapid solidification forming of glassy and crystalline ribbons by planar flow casting. Chem. Eng. Sci. 192 , 1198–1208. https://doi.org/10.1016/j.ces.2018.07.017 (2018). Su, Y-G., Chen, F., Wu, C-Y., Chang, M-H. & Chung, C-A. Effects of manufacturing parameters in planar flow casting process on ribbon formation and puddle evolution of Fe78Si9B13 alloy. ISIJ Int. 55 , 2383–2390 (2015). Li, D., Zhuang, J., Liu, T., Lu, Z. & Zhou, S. The pressure loss and ribbon thickness prediction in gap controlled planar-flow casting process. J. Mater. Process. Technol. 211 , 1764–1767. https://doi.org/10.1016/j.jmatprotec.2011.05.019 (2011). Karcher, C. & Steen, P. H. Solidification and downstream meniscus prediction in the planar-flow melt spinning process. Chem. Eng. Sci. 62 , 5145–5150. https://doi.org/10.1016/j.ces.2007.02.042 (2007). Cox, B. L. & Steen, P. H. Herringbone' defect formation in planar-flow melt spinning. J. Mater. Process. Technol. 213 , 1743–1752. https://doi.org/10.1016/j.jmatprotec.2013.04.010 (2013). Byrne, C., Theisen, E., Reed, B. & Steen, P. H. Capillary puddle vibrations linked to casting-defect formation in planar-flow melt spinning. Metall. Mater. Trans. B . 37 , 445–456. https://doi.org/10.1007/s11663-006-0029-4 (2006). Chung, C-A., Su, Y-G. & Chen, F. Tuning the Planar-Flow Melt-Spinning Process Subject to Operability Constraints. JOM 66, 2082–2088 (2014). https://doi.org/10.1007/s11837-014-0982-3 Cho, K. H., Lee, K. A., Kim, M. C. & Yoon, J. M. Numerical modeling of planar flow casting process. Solid State Phenom. 116–117 , 106–109 (2006). https://doi.org/10.4028/www.scientific.net/SSP.116-117.106 Das, S., Rao, P. S. & Ghosh, S. Flow CFD analysis of 2D transient melt flow from crucible onto the rotating wheel in PFMS. Mater. Today Proc. 4, 9810–9816 (2017). https://doi.org/10.1016/j.matpr.2017.06.289 Das, S., Rao, P. S. & Ghosh, S. Obtaining stable puddle and thinner ribbons during planar flow melt spinning: a 2D analysis. Mater. Today Proc. 4, 9656–9662 (2017). https://doi.org/10.1016/j.matpr.2017.06.035 Pradhan, A. & Ghosh, S. Flow dynamics in the melt puddle during planar flow melt spinning: A numerical study. Mater. Today Proc. 4, 9829–9838 (2017). https://doi.org/10.1016/j.matpr.2017.06.311 Heine, G. & Fischer, K. The effect of the main processing parameters on the geometry of melt-spun ribbons. Z. Metallkde . 75 , 403–409 (1984). Li, D. & Lu, Z. Effects of aging on the cyclical thermal shock response of a copper–beryllium alloy as a substrate of cooling wheel in planar flow casting process. Mater. Res. Express . 7 , 116511. https://doi.org/10.1088/2053-1591/abae58 (2020). Li, D., Wang, W., Liu, T., Li, L. & Lu, Z. Reducing the core losses of Fe–Si–B amorphous alloy ribbons by high cooling rate planar flow casting. Materials 15 , 894. https://doi.org/10.3390/ma15030894 (2022). Li, Y. et al. Excellent magnetic properties obtained in planar flow casting Fe–6.5%Si ribbons. Mater. Des. 235 , 112556. https://doi.org/10.1016/j.matdes.2024.112556 (2024). Zhang, Y. et al. Mechanistic investigation of thickness–controlled microstructure and performance in planar flow casting. J. Mater. Process. Technol. 324 , 118924. https://doi.org/10.1016/j.jmatprotec.2025.118924 (2025). Greer, A. L. Metallic glasses. MRS Bull. 48 , 842–854. https://doi.org/10.1557/s43577-023-00586-5 (2023). Theisen, E. A. & Weinstein, S. J. An overview of planar flow casting of thin metallic glasses (preprint). arXiv, arXiv:2104.09251. (2021). Wang, R. et al. Review of surface defect detection of steel products based on machine vision. IET Image Process. 17 , 3039–3060. https://doi.org/10.1049/ipr2.12647 (2023). Zhang, R. et al. A deep–learning–based approach for fast and robust steel surface defect classification. Pattern Recogn. 103 , 107312. https://doi.org/10.1016/j.patcog.2019.107312 (2020). Wang, Z. et al. A systematic review of deep learning approaches for surface defect detection in manufacturing. Eng. Appl. Artif. Intell. 127 , 107327. https://doi.org/10.1016/j.engappai.2023.107327 (2024). Kaya, M. & Karaköse, M. Systematic review of steel surface defect detection methods on the open datasets. LNCS 14470 , 35–58 (2024). Li, H. et al. Research on steel surface defect classification method based on deep learning. Sci. Rep. 14 , 12876. https://doi.org/10.1038/s41598-024-58643-1 (2024). Hossain, R. E. N., Lewis, J. & Moore, A. L. In–situ infrared temperature sensing for real–time defect detection in additive manufacturing. Addit. Manuf. 47 , 102328. https://doi.org/10.1016/j.addma.2021.102328 (2021). Kavale, N. S. et al. Process monitoring using synchronized path infrared thermography (SPIT). Sensors 22 , 5943. https://doi.org/10.3390/s22165943 (2022). Abdelrahman, M., Leach, R. & Maskery, I. Assisted defect detection by in–process monitoring of additive manufacturing using optical imaging and infrared thermography. Addit. Manuf. 65 , 103487. https://doi.org/10.1016/j.addma.2023.103487 (2023). García–González, A., Alkorta, J. & Sevillano, J. G. Detection of sub–superficial defects by infrared thermography in parts manufactured by additive manufacturing. Int. J. Adv. Manuf. Technol. https://doi.org/10.1007/s40964-025-01313-z (2025). Gopaluni, R. B. et al. Modern machine learning tools for monitoring and control of industrial processes: A survey. IFAC–PapersOnLine 53 , 12167–12178. https://doi.org/10.1016/j.ifacol.2020.12.2800 (2020). Lawrence, N. P. et al. Machine learning for industrial sensing and control: A survey and practical perspective. Annu. Rev. Control . 58 , 100985. https://doi.org/10.1016/j.arcontrol.2024.100985 (2024). Tercan, H. & Meisen, T. Machine learning and deep learning based predictive quality in manufacturing: a systematic review. J. Intell. Manuf. 33 , 1879–1905. https://doi.org/10.1007/s10845-022-01963-8 (2022). Lawrence, N. P., Gopaluni, R. B., Lee, J. M. & Huang, B. Machine learning for industrial sensing and control: a survey and practical perspective (preprint). arXiv, arXiv:2401.13836. (2024). Gopaluni, R. B. et al. Modern machine learning tools for monitoring and control of industrial processes: A survey (preprint). arXiv, arXiv:2209.11123. (2022). Zhang, Q. et al. Steel surface defect detection algorithm in complex background via improved vision transformers. Measurement 231 , 114547. https://doi.org/10.1016/j.measurement.2024.114547 (2024). Özkan, A. & Uçar, A. Detection of defects on metal surfaces based on deep learning. Appl. Sci. 15 , 1406. https://doi.org/10.3390/app15031406 (2025). Kubik, D. et al. Reliable machine learning models for manufacturing processes. Int. J. Adv. Manuf. Technol. https://doi.org/10.1007/s00170-025-16539-y (2025). Yan, P. et al. Learning actionable world models for industrial process control (arXiv, 2025). arXiv:2503.01411. Tables Table 1 Dataset overview. Summary of the experimental dataset used in this study, including the number of casting runs, frames analysed, experimental configurations, total data volume, and metadata fields available. Property Details Number of casting runs ~ 200 runs across multiple conditions Frames analysed (RGB + IR) ~ 300,000 frames (synchronised RGB/IR) Experimental setups Side view, front view, infrared imaging (subset) Total dataset volume ~ 20 GB structured dataset Metadata fields Gap (mm), flow proxy (% change in optical flow velocity), wheel speed (rpm), melt temperature (°C), nozzle geometry Table 2 Physics-informed descriptors used for ML. Features were extracted from video frames and grouped into ridge, thermal, and reflection descriptors, with units and physical relevance. Category Descriptor Definition / Units Physical relevance Ridge Ridge density ridges per mm Surface stability indicator Ridge Ridge width (mean) mm Foil uniformity measure Ridge Ridge width (std) mm Variation in ridge spacing Thermal Thermal gradient variance °C² Edge vs centre thermal imbalance Thermal Edge temperature °C Hot-spot indicator Reflection Reflection intensity No Unit Gap proxy from RGB Geometry Foil width mm Casting geometry constraint Table 3 Baseline vs ML model performance. Comparison of baseline proxies, XGBoost, and deep-learning fusion models for gap and flow prediction. Metrics averaged across validation folds. Model Target MAE RMSE R² Calibration error Baseline proxy Gap 0.12 mm 0.15 mm 0.05 High bias Baseline proxy Flow (% change, optical flow velocity) 0.30% 0.35% ≈ 0.00 High bias XGBoost Gap 0.010 mm 0.013 mm 0.97 Low XGBoost Flow (% change, optical flow velocity) 0.18% 0.22% 0.12 Moderate DL Fusion Gap 0.011 mm 0.014 mm 0.96 Low DL Fusion Flow (% change, optical flow velocity) 0.17% 0.21% 0.15 Moderate Additional Declarations No competing interests reported. Supplementary Files SupplementaryInformationFinal.pdf Cite Share Download PDF Status: Posted Version 1 posted 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-8042276","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":547049876,"identity":"67f63e2d-f3d2-4e37-816e-31c5f77e97c2","order_by":0,"name":"Ritaban Dutta","email":"data:image/png;base64,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","orcid":"","institution":"CSIRO Data61","correspondingAuthor":true,"prefix":"","firstName":"Ritaban","middleName":"","lastName":"Dutta","suffix":""},{"id":547049877,"identity":"ac82d460-5949-4d15-9e98-9ab496647cab","order_by":1,"name":"Michael Kellam","email":"","orcid":"","institution":"CSIRO Manufacturing","correspondingAuthor":false,"prefix":"","firstName":"Michael","middleName":"","lastName":"Kellam","suffix":""},{"id":547049878,"identity":"b8ec41c3-a27a-4565-acca-003b15351626","order_by":2,"name":"Daniel Liang","email":"","orcid":"","institution":"CSIRO Manufacturing","correspondingAuthor":false,"prefix":"","firstName":"Daniel","middleName":"","lastName":"Liang","suffix":""}],"badges":[],"createdAt":"2025-11-06 00:38:05","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8042276/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8042276/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":96377689,"identity":"2cedd815-d5c2-436e-a39a-f4b7bf105ff7","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3240270,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/032a0709b7822b085bdaa42b.png"},{"id":96377690,"identity":"89919dd5-4fd7-46e0-9345-0c0bb2035726","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6400026,"visible":true,"origin":"","legend":"","description":"","filename":"NSRPFCManuscriptDraftFinalwithFundingDeclaration.docx","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/392177e14af78752e3525d7c.docx"},{"id":96377684,"identity":"f46e4203-40e5-4646-bab3-03adb9da164d","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1129068,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/c626f31d3483883fef679e49.png"},{"id":96452896,"identity":"ee7ded2c-d283-49a1-a10a-d36e9cf7e822","added_by":"auto","created_at":"2025-11-21 09:52:50","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1037049,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/36423b147f94324842ab4aa4.png"},{"id":96453732,"identity":"60753a64-665c-4fae-9533-885a517661ab","added_by":"auto","created_at":"2025-11-21 10:01:25","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40638,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/29c04b53f38277382a4a8186.docx"},{"id":96453234,"identity":"e475c9a8-d033-4aa5-a27b-5f4c8f0bd2f8","added_by":"auto","created_at":"2025-11-21 09:58:51","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1111504,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/c237977025dee052bd5ea6a2.png"},{"id":96377691,"identity":"2f334609-3277-4b76-b343-cad99c8b9cd4","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"docx","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40873,"visible":true,"origin":"","legend":"","description":"","filename":"Table2.docx","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/9405439fe41aeb0c66f9a45e.docx"},{"id":96453077,"identity":"bf70e367-3d73-4fcc-8eb4-74b359ca8462","added_by":"auto","created_at":"2025-11-21 09:57:58","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":999052,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/82c3e03cc88150ab6856ffaa.png"},{"id":96377694,"identity":"46dc0431-c085-44ab-9f28-335184603bc0","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"docx","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40979,"visible":true,"origin":"","legend":"","description":"","filename":"Table3.docx","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/afe5073f6c1b22e07e834153.docx"},{"id":96377704,"identity":"2a91b9b7-fc5b-43d0-94df-8d7e6da55bb8","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":871523,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/cd381a4464961035ed09d871.png"},{"id":96377697,"identity":"e41d5da4-9aed-4f44-a79b-bd2689807931","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2998612,"visible":true,"origin":"","legend":"","description":"","filename":"Figure7.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/2800b91382681fdff6b42d77.png"},{"id":96454085,"identity":"9bb4ada2-5ceb-4ce4-bdeb-7866fe227144","added_by":"auto","created_at":"2025-11-21 10:02:19","extension":"json","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5907,"visible":true,"origin":"","legend":"","description":"","filename":"3e00745804a143f5873399e36b5eb317.json","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/3610f7b8b4998bd070d092f0.json"},{"id":96452883,"identity":"e550e468-26b9-4ba0-85c6-eee269b30454","added_by":"auto","created_at":"2025-11-21 09:52:18","extension":"pdf","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5087456,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryInformationFinal.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/f06f348c3c10fb85fe038890.pdf"},{"id":96377699,"identity":"cdd2aadc-e10f-4c69-ba79-506eff676c5e","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"xml","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":128192,"visible":true,"origin":"","legend":"","description":"","filename":"3e00745804a143f5873399e36b5eb3171enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/44476d320cd580e3b6c1830a.xml"},{"id":96453056,"identity":"0067f3c9-e500-40a7-8a3f-266039e3039a","added_by":"auto","created_at":"2025-11-21 09:57:50","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3240270,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/f7fa32221d93ae6c28f4d242.png"},{"id":96377720,"identity":"0d48ac15-1e61-434d-93ed-7ac88ac96a36","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1129068,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/337f83145aca168bbf30b6b3.png"},{"id":96377706,"identity":"c64a33f9-5dce-4d1a-a256-72257851fa64","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1037049,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/069613159da8addac42e72ff.png"},{"id":96377711,"identity":"540cb29d-137c-4d78-a712-5e4af2a9d898","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1111504,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/8e7c5a221e192d8037b0c367.png"},{"id":96377695,"identity":"a6a2f890-f9df-4589-ac59-cb1a9b866f57","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":999052,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/7766fb6e9b0c9df343f8ce61.png"},{"id":96453306,"identity":"af268e63-0d24-4341-a31a-2e9977823d7c","added_by":"auto","created_at":"2025-11-21 09:59:11","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":871523,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/6d1f3e8af0b1a66491384925.png"},{"id":96377734,"identity":"2032423d-72be-41cf-b609-fab369340867","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2998612,"visible":true,"origin":"","legend":"","description":"","filename":"Figure7.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/979127c0d67a82057945aacb.png"},{"id":96377705,"identity":"272c626a-e3bd-40bf-9f92-f4768349250f","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2676689,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/35e18bbc9a69ed99eef972cf.png"},{"id":96377710,"identity":"82b13aa6-ee08-47b4-ba2c-d41f3081999d","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":988188,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/5aacf07d6f99cfd013b4e56f.png"},{"id":96377729,"identity":"ef15b2fd-aca0-4cce-b0a5-56d42133defe","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":902853,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/fdcc88cde8d77a0d63381f29.png"},{"id":96377700,"identity":"2a36ae30-46c6-4115-abd6-4e028c63021b","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":961854,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/3fd878b328fa89ebe17bc02f.png"},{"id":96377712,"identity":"8982bf55-934b-4f8f-b7fc-0e38f5670ed7","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":784411,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/66ae5793ded7745a34f13b76.png"},{"id":96377736,"identity":"fbae2ba6-fe7a-46b1-b280-d2dd0d58091b","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":389100,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/1526f0b1dba7bcead08aafd4.png"},{"id":96453103,"identity":"25674786-35c5-4da2-9893-dfdec4bb263f","added_by":"auto","created_at":"2025-11-21 09:58:05","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":248861,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/b6acd0e7beec34eaf3153b08.png"},{"id":96377733,"identity":"142568f5-d085-4981-b5fc-b4e72cb75e0f","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":226984,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/1ec243729678f4172ee7e257.png"},{"id":96377709,"identity":"548fa934-8ade-4216-911e-ad4e8b6e0e05","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":241523,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/32b54315534232d2e54d98bc.png"},{"id":96454093,"identity":"0c9fc880-d9ff-496f-837c-17339ec61238","added_by":"auto","created_at":"2025-11-21 10:02:19","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":222038,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure5.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/05077bc46395c1579769dd54.png"},{"id":96377737,"identity":"9a6879f0-3347-4bf6-b4be-790d1f91c08f","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":198686,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure6.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/42a7d95a8a520c20dced8c0b.png"},{"id":96453165,"identity":"27972645-07ae-4542-9ac6-f54eb5250444","added_by":"auto","created_at":"2025-11-21 09:58:36","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":385763,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure7.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/4e99b349641814b31838ea05.png"},{"id":96377714,"identity":"06ec16ec-ffa1-4485-aaf0-7441a6c55d5b","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":398757,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/6c1562f1c76d9bae84181cb5.png"},{"id":96454101,"identity":"c6a18cc8-1211-4d83-888a-668a38a32cbc","added_by":"auto","created_at":"2025-11-21 10:02:20","extension":"png","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":258762,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/2216bcf3c4b3d209e2f3e2c2.png"},{"id":96453129,"identity":"90be4f56-ac17-4b3b-920b-73bf0c8fcfa3","added_by":"auto","created_at":"2025-11-21 09:58:22","extension":"png","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":236595,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/de5d21b4fd63122d9b4ada71.png"},{"id":96453136,"identity":"795cf31f-794f-460f-b57f-fc965ff15153","added_by":"auto","created_at":"2025-11-21 09:58:25","extension":"png","order_by":36,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":251510,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/0015a1943f3176e4b1fab4f1.png"},{"id":96453353,"identity":"3f5f5d78-ede9-4de2-8f11-6ea40f436c33","added_by":"auto","created_at":"2025-11-21 09:59:24","extension":"png","order_by":37,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":231289,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/9f5004afaa353fe52576f0e4.png"},{"id":96377731,"identity":"cb6e7480-d423-4ec5-ae5f-967c5fe125e9","added_by":"auto","created_at":"2025-11-20 11:35:25","extension":"xml","order_by":38,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":128458,"visible":true,"origin":"","legend":"","description":"","filename":"3e00745804a143f5873399e36b5eb3171structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/8781b8da80a84380ec9f7e91.xml"},{"id":96452965,"identity":"ea139682-0d09-4e1e-9e69-fc305878bc4e","added_by":"auto","created_at":"2025-11-21 09:56:23","extension":"html","order_by":39,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":143571,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/8dacde39a2006f0292578d5c.html"},{"id":96377682,"identity":"13bb809a-5457-4466-8300-52c8931dc3d4","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":3240270,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExperimental setup, dataset overview, and integrated analytics.\u003cbr\u003e\n\u003c/strong\u003e(a) Photograph of the planar-flow casting chamber showing the nozzle–wheel configuration, induction coil, and melt-delivery region. This provides the physical basis for video-based monitoring.\u003c/p\u003e\n\u003cp\u003e(b) Nozzle–melt view highlighting the melt ribbon emerging into the gap, where instabilities originate.\u003c/p\u003e\n\u003cp\u003e(c) Front-view infrared video frame with ridge annotations, where automated surface analysis quantifies ridge density, ridge width, and unevenness, providing direct video-derived quality descriptors.\u003c/p\u003e\n\u003cp\u003e(d) Side-view video frame with annotated ribbon trajectory and overlaid process parameters, including wheel speed, foil thickness, and ribbon travel, linking visual cues to operational conditions.\u003c/p\u003e\n\u003cp\u003e(e) Ridge-width distributions across casts visualised as violin/box plots, showing narrow distributions for stable runs and broad, asymmetric distributions for unstable runs.\u003c/p\u003e\n\u003cp\u003e(f) Thermal profiles across foil width averaged across runs, with variability (±σ) captured as coloured bands, illustrating systematic edge-heating during unstable casting.\u003c/p\u003e\n\u003cp\u003e(g) Gap prediction (true vs predicted, hexbin density) demonstrates feasibility of learning accurate gap estimates directly from video-derived features.\u003c/p\u003e\n\u003cp\u003e(h) Flow proxy prediction, expressed as % change in optical flow velocity, establishes a link between melt dynamics and measured flow instabilities.\u003c/p\u003e\n\u003cp\u003e(i) Distribution of gap residuals shows near-zero mean bias, validating the analytical alignment of machine-learning predictions with physical ground truth. Together, these panels illustrate the integration of experimental imaging, video-derived descriptors, and predictive analytics that underpin the dataset.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/915b3dbd507e2da0900244c3.png"},{"id":96377681,"identity":"c5692f21-3322-492b-ac67-e7f9b9038391","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1129068,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBaseline analytical performance across gap, flow, and foil quality metrics.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(a) Scatter of true vs predicted gap values using reflection-based heuristics shows systematic underestimation and widespread, confirming limited reliability.\u003c/p\u003e\n\u003cp\u003e(b) Time series of baseline predictions for a representative run reveal that while average gap trends are followed, rapid frame-to-frame dynamics are missed, reflecting poor temporal fidelity.\u003c/p\u003e\n\u003cp\u003e(c) Flow proxy predictions show weak correspondence with ground truth, illustrating that heuristic optical cues fail to track real flow changes.\u003c/p\u003e\n\u003cp\u003e(d) Residual distribution for flow proxy predictions reveals skew and bias, demonstrating structural inadequacy of baseline models.\u003c/p\u003e\n\u003cp\u003e(e) Ridge-width distributions (stable vs unstable runs) show narrow, symmetric histograms under stable casting but broad, skewed profiles when unstable, linking ridge behaviour to surface integrity.\u003c/p\u003e\n\u003cp\u003e(f) Thermal-gradient profiles across foil width highlight strong edge-dominated gradients in unstable runs, consistent with higher ridge density and reduced amorphicity. Collectively, these results establish a lower-bound benchmark for evaluating advanced ML approaches.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/004868c3b8c68e8821a4b357.png"},{"id":96453095,"identity":"1bc9e051-d863-4446-aea6-02da98d4eaab","added_by":"auto","created_at":"2025-11-21 09:58:03","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1037049,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eXGBoost feature-based machine-learning performance.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(a) Scatter of true vs predicted gap values using XGBoost shows predictions tightly aligned with the 1:1 line (R² ≈ 0.97), representing a major improvement over baselines.\u003c/p\u003e\n\u003cp\u003e(b) Time-series analysis of a representative run demonstrates that XGBoost accurately tracks fine-scale gap fluctuations, showing frame-level dynamic fidelity.\u003c/p\u003e\n\u003cp\u003e(c) Flow proxy predictions achieve improved correlation compared to heuristic methods, though variation persists, reflecting challenges in capturing subtle flow signatures.\u003c/p\u003e\n\u003cp\u003e(d) Distribution of flow proxy residuals is centred on zero, confirming removal of systematic bias seen in baseline models.\u003c/p\u003e\n\u003cp\u003e(e) Calibration curve indicates excellent statistical alignment between predicted and true values, ensuring predictions are not systematically over- or under-estimated across the range.\u003c/p\u003e\n\u003cp\u003e(f) Feature-importance ranking shows ridge density and thermal-gradient variance as dominant predictors, followed by reflection intensity, in strong agreement with the underlying casting physics. These results confirm that physics-informed descriptors yield accurate and interpretable models.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/5ea048dd599ccfba145cb9da.png"},{"id":96377688,"identity":"acdae7db-5071-46dd-94f3-b512d2ac3e47","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1111504,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDeep learning fusion performance with multimodal inputs.\u003c/strong\u003e\u003cbr\u003e\n(a) Scatter of true vs predicted gap values using deep-learning fusion (RGB, IR, and tabular features) shows close clustering along the 1:1 line, confirming high accuracy.\u003cbr\u003e\n(b) Time-series analysis across a representative run illustrates that the fusion model reliably tracks frame-level gap fluctuations with low error.\u003cbr\u003e\n(c) Flow proxy predictions improve relative to heuristic baselines but remain noisier than gap estimates, reflecting the inherent difficulty of capturing subtle melt-flow variations.\u003cbr\u003e\n(d) Residual distribution for flow proxy predictions is centred near zero with reduced variance compared to baselines, confirming removal of systematic bias.\u003cbr\u003e\n(e) Calibration curve demonstrates excellent agreement between predicted and true gap values across binned ranges, confirming statistical reliability.\u003cbr\u003e\n(f) Training and validation curves converge smoothly with no signs of overfitting, highlighting effective generalisation. While deep learning delivers accuracy comparable to feature-based XGBoost, it requires substantially higher computational resources, illustrating a trade-off between interpretability and scalability.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/e249d182b5ef6ca6ce9a6431.png"},{"id":96377683,"identity":"ee988f7f-0706-4bf1-b7b8-f88fa43b19ce","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":999052,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFoil-quality analysis using ridge and thermal descriptors.\u003c/strong\u003e\u003cbr\u003e\n(a) Ridge-density profiles across the foil width highlight stable runs with uniform spacing, while unstable runs show clustering and irregularity, indicating compromised uniformity.\u003cbr\u003e\n(b) Ridge-width distributions are narrow and symmetric during stable casting but broaden and skew under unstable conditions, reflecting degraded surface quality.\u003cbr\u003e\n(c) Thermal-gradient profiles across foil width are balanced and symmetric in stable runs, but unstable runs exhibit steep edge-dominated gradients, consistent with uneven cooling.\u003cbr\u003e\n(d) Edge-to-centre thermal gradient ratios quantify this imbalance, with unstable runs showing significantly elevated ratios, confirming the thermally driven nature of ridging.\u003cbr\u003e\n(e) Correlation between ridge density and thermal imbalance (ΔT) reveals a strong linear relationship (r ≈ 0.8), linking surface features directly to underlying thermodynamics.\u003cbr\u003e\n(f) Composite overlays of ridge maps with thermal-gradient heatmaps demonstrate the combined diagnostic power of structural and thermal cues, providing both visual and quantitative markers of foil integrity.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/65bd55fa68453a209fd26e73.png"},{"id":99789396,"identity":"8a6e6420-3c34-4e36-ad82-dce4bfb74d3b","added_by":"auto","created_at":"2026-01-08 12:49:34","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":7738729,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/84cece54-b87c-4d56-8e7f-4f600f3710ad.pdf"},{"id":96377680,"identity":"7ea4d0dc-6d81-4d17-bc53-aa4b2903fe89","added_by":"auto","created_at":"2025-11-20 11:35:24","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":5087456,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryInformationFinal.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8042276/v1/1d71de2ea508f727f0e2a123.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"AI-enabled Monitoring of Planar-Flow Casting of Amorphous Steel Foils","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAmorphous metallic alloys combine high strength, corrosion resistance, and soft-magnetic behaviour, making them attractive for energy-efficient transformers, electric motors, and advanced electronic applications. The ability to manufacture thin amorphous foils at scale is central to their industrial adoption. Planar-flow casting (PFC) is the most practical route for producing such foils, directly quenching molten alloy onto a rotating copper wheel to form continuous ribbons [\u003cspan additionalcitationids=\"CR2 CR3 CR4 CR5 CR6 CR7 CR8 CR9 CR10 CR11 CR12\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eDespite its promise, PFC is inherently unstable. Small perturbations in melt temperature, nozzle\u0026ndash;wheel gap, or wheel speed can destabilise the melt pool, disrupt heat transfer at the wheel interface, and generate surface ridges or thickness variations. These instabilities not only compromise foil integrity but also degrade the magnetic properties critical for industrial use. Conventional monitoring relies on bulk process variables or post-process inspection, neither of which captures the fast spatio-temporal dynamics of the melt\u0026ndash;wheel interaction where instabilities originate [\u003cspan additionalcitationids=\"CR15 CR16 CR17 CR18 CR19\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eHigh-speed video and thermal imaging now allow direct observation of melt delivery and foil formation. However, such data are high dimensional and require systematic analysis methods. Computer vision (CV) can extract interpretable descriptors\u0026mdash;ridge density, reflection patterns, and thermal gradients\u0026mdash;while machine learning (ML) can model their relationship to process variables and foil quality. Deep learning (DL) architectures extend this further by learning spatio-temporal representations from raw multimodal frame sequences.\u003c/p\u003e\u003cp\u003eIn this work, we introduce an integrated framework for analysing more than 200 PFC experiments using synchronised RGB and infrared video paired with detailed metadata. A structured dataset was created, annotated with process variables and quality descriptors, and analysed through both feature-based ML models and multimodal DL fusion networks. Baseline analytical proxies were benchmarked against these approaches to establish lower- and upper-bound performance. In addition, ridge-width distributions and thermal-gradient profiles were evaluated as direct video-derived measures of foil integrity [\u003cspan additionalcitationids=\"CR22 CR23 CR24 CR25 CR26 CR27 CR28 CR29\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe results, summarised in Figs.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, Supplementary Figs.\u0026nbsp;6\u0026ndash;7 and Tables\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, Supplementary Tables S1-S5, demonstrate how CV- and ML-based approaches can provide accurate, interpretable, and scalable monitoring of PFC. Together, these contributions establish a pathway toward reproducible and automated optimisation of amorphous steel-foil manufacturing, and lay the groundwork for industrial deployment in next-generation casting facilities [\u003cspan additionalcitationids=\"CR31 CR32 CR33 CR34 CR35 CR36 CR37\" citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eDataset description\u003c/h2\u003e\u003cp\u003eA multimodal dataset was assembled from approximately 200 planar-flow casting (PFC) experiments performed under varied operating conditions. Each run included synchronised video recordings from front-view and side-view RGB cameras, with a subset also incorporating infrared (IR) thermal imaging of the emerging foil. In total, the dataset contains\u0026thinsp;\u0026asymp;\u0026thinsp;300,000 frames, corresponding to \u0026asymp;\u0026thinsp;20 GB of structured video data.\u003c/p\u003e\u003cp\u003eEach run was assigned a unique identifier and linked to a manifest file containing metadata. Recorded fields included nozzle\u0026ndash;wheel gap, wheel speed, melt temperature, nozzle geometry, and flow proxy variables. This structure ensured that every video frame could be directly associated with corresponding process conditions [\u003cspan additionalcitationids=\"CR2 CR3 CR4 CR5 CR6 CR7 CR8 CR9 CR10\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe dataset spans both stable and unstable regimes, capturing variability in ridging, surface uniformity, and thermal balance. This diversity provides a foundation for developing and benchmarking predictive models that generalise across different casting conditions.\u003c/p\u003e\u003cp\u003eA summary of dataset properties, including the number of runs, frames, experimental setups, and metadata fields, is given in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Representative images from RGB and IR recordings are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, while extended metadata descriptors, parameter ranges, and acquisition details are provided in Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eExperimental setup and video acquisition\u003c/h3\u003e\n\u003cp\u003eHigh-speed imaging was used to capture melt delivery, foil formation, and nozzle\u0026ndash;wheel interactions during each casting run. Two RGB cameras were positioned to provide complementary perspectives:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eA \u003cb\u003efront-view camera\u003c/b\u003e, aligned with the emerging foil, which captured ridge formation, foil width, and surface reflections [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eA \u003cb\u003eside-view camera\u003c/b\u003e, aligned with the nozzle\u0026ndash;wheel interface, which recorded the gap, ribbon trajectory, and melt-pool behaviour [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eIn a subset of runs, an infrared (IR) thermal camera was mounted to capture the temperature distribution across the foil width during solidification. These IR recordings enabled frame-by-frame measurement of edge-to-centre thermal gradients, which are linked to foil stability and amorphicity [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eAll videos were acquired at 30 frames per second, with resolutions sufficient to resolve ridge spacing, reflection patterns, and foil width variations. Camera alignment was verified before each experiment, and lighting was adjusted to minimise glare and highlight surface contrast [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eEach casting run was indexed in the dataset by its unique identifier, with video files stored in a structured hierarchy and cross-referenced with metadata logs. This arrangement allows frame-level synchronisation of visual and process information. Representative frames from the front- and side-view cameras, as well as IR recordings, are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Acquisition settings and camera specifications are provided in Supplementary Table S2.\u003c/p\u003e\n\u003ch3\u003eGround-truth measurement and labelling\u003c/h3\u003e\n\u003cp\u003eDirect ground-truth measurements were collected for each casting run to ensure that predictive modelling could be benchmarked against reliable process variables.\u003c/p\u003e\u003cp\u003eThe nozzle\u0026ndash;wheel gap was measured prior to casting using precision micrometre gauges. These values were logged in the manifest file and treated as ground-truth gap values [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. After each run, nozzle alignment was re-checked to confirm stability and rule out drift or deformation effects.\u003c/p\u003e\u003cp\u003eThe flow proxy was derived from operator logs of melt-head height and reservoir pressure, which influence volumetric melt delivery. In selected runs, this parameter was cross validated against the measured foil mass per unit time. Although variation in flow was relatively small compared with gap, it provides an important secondary ground-truth variable [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eFoil-quality metrics were generated directly from video recordings using automated computer-vision pipelines. From front-view recordings, ridge positions, ridge density, and ridge-width distributions were extracted. From infrared recordings, frame-level temperature profiles and cross-foil thermal gradients were calculated. By automating these annotations, reproducibility was maintained across all \u0026asymp;\u0026thinsp;300,000 frames, and subjective bias was eliminated [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eSummary statistics of the measured process variables are provided in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Extended distributions of gap and flow values across all runs are given in Supplementary Table S3, while cross-run variability and stability ranges are detailed in Supplementary Table S4 [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003eAnnotation and labelling\u003c/h3\u003e\n\u003cp\u003eEach casting run was indexed by a unique identifier linking raw video, metadata, and derived annotations. Frame-level annotations were generated to ensure consistency across the dataset and reproducibility of downstream modelling.\u003c/p\u003e\u003cp\u003eFor process variables, nozzle\u0026ndash;wheel gap and flow proxy values were assigned frame by frame using the manifest file. These values were aligned temporally with the RGB and IR recordings, allowing supervised learning without leakage between training and validation sets [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eFor foil quality, annotations were obtained using automated computer-vision pipelines. From front-view videos, ridge positions were identified using edge-detection algorithms, producing ridge-density maps and ridge-width histograms. From IR recordings, temperature gradients were computed across the foil width to capture edge\u0026ndash;centre imbalances [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. An illustration of these annotation workflows is provided in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003eAll annotations were stored in tabular format, aligned with the manifest file. This ensured that every frame had a consistent set of labels linking raw imagery, metadata, and derived descriptors. A consolidated summary of annotation categories is provided in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, while the detailed schema, including annotation fields and feature formats, is presented in Supplementary Table S5.\u003c/p\u003e\n\u003ch3\u003eFeature extraction and ridge/thermal analysis\u003c/h3\u003e\n\u003cp\u003eTo transform raw video into structured input for modelling, a suite of physics-inspired descriptors was extracted from both RGB and infrared recordings. These descriptors capture the surface, geometric, and thermal characteristics most closely associated with casting stability and foil integrity.\u003c/p\u003e\u003cp\u003eFrom front-view videos, features included ridge density, ridge spacing, ridge-width distributions, and their variation across foil width. Reflection intensity and foil width were quantified from brightness and edge contrast, providing proxies for melt stability. From side-view videos, reflection height and ribbon trajectory features were derived to capture nozzle\u0026ndash;wheel gap behaviour.\u003c/p\u003e\u003cp\u003eFrom infrared recordings, cross-foil thermal profiles were analysed to compute temperature gradients between edges and centre. These measures provide a sensitive indicator of uneven cooling, which often correlates with ridge clustering and the loss of amorphicity.\u003c/p\u003e\u003cp\u003eAll features were aggregated into frame-level datasets linked to casting run identifiers. For each run, this produced\u0026thinsp;\u0026asymp;\u0026thinsp;27 structured descriptors per sample, forming the input for machine-learning and deep-learning models [\u003cspan additionalcitationids=\"CR32 CR33\" citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eAn overview of feature categories and their physical interpretation is summarised in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, while extended definitions and parameter ranges are provided in Supplementary Table S5. Together, these descriptors underpin both predictive modelling of gap and flow, and direct video-based evaluation of foil quality. See Supplementary Videos S1\u0026ndash;S2 for visual demonstrations of the automated ridge-tracking and thermal-gradient analysis workflows.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eBaseline models\u003c/h2\u003e\u003cp\u003eAs a lower-bound benchmark, we implemented simple analytical proxies that approximate process conditions directly from video streams without advanced modelling. These baselines reflect the heuristic approaches sometimes used in industrial inspection.\u003c/p\u003e\u003cp\u003eTwo proxies were developed:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eGap\u0026ndash;reflection proxy\u003c/b\u003e \u0026mdash; derived from the vertical extent of reflection intensity in side-view RGB recordings, serving as a crude estimator of nozzle\u0026ndash;wheel gap.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eInfrared hot-area proxy\u003c/b\u003e \u0026mdash; defined as the fraction of the foil surface above a fixed temperature threshold in IR frames, used as a surrogate for flow stability.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003eThese methods were applied frame by frame to the dataset, with outputs aligned to ground-truth gap and flow values. While they captured broad trends, they lacked sensitivity to dynamic fluctuations, especially under unstable casting conditions [\u003cspan additionalcitationids=\"CR37\" citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e\u003cp\u003ePerformance of the baseline proxies is summarised in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, including mean absolute error (MAE), root-mean-square error (RMSE), and coefficient of determination (R\u0026sup2;). Representative results are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, illustrating the systematic underestimation, high residual variance, and limited explanatory power of these heuristic approaches.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eMachine learning models (XGBoost)\u003c/h3\u003e\n\u003cp\u003eTo benchmark predictive performance beyond analytical proxies, we trained feature-based machine-learning models using the XGBoost framework. Input data were the structured descriptors described above, including ridge density, ridge-width statistics, reflection intensity, foil width, and thermal gradients.\u003c/p\u003e\u003cp\u003eThe dataset comprised\u0026thinsp;\u0026asymp;\u0026thinsp;8,800 aggregated samples, each containing 27 descriptors per cast. To avoid frame-level leakage, group-wise cross-validation was used, ensuring that entire casting runs were held out for validation. Hyperparameters \u0026mdash; learning rate, maximum depth, number of estimators, subsampling ratios, and regularisation terms \u0026mdash; were tuned via grid search (Supplementary Table S4).\u003c/p\u003e\u003cp\u003eOutputs included predictions of nozzle\u0026ndash;wheel gap (mm) and flow proxy (% change in optical-flow velocity). Models were trained with mean absolute error (MAE) as the primary loss metric, while root-mean-square error (RMSE) and coefficient of determination (R\u0026sup2;) were also evaluated. All training was performed with deterministic folds and fixed random seeds, as detailed in Supplementary Table S2.\u003c/p\u003e\u003cp\u003ePerformance results are reported in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, demonstrating a significant improvement over baseline proxies. Representative results, including scatter plots of true vs. predicted gap, temporal tracking, calibration analysis, and feature-importance rankings, are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Extended diagnostics, including fold-level variability and learning curves, are provided in Supplementary Table S3.\u003c/p\u003e\n\u003ch3\u003eDeep learning fusion models\u003c/h3\u003e\n\u003cp\u003eTo evaluate whether predictive accuracy could be improved by learning directly from raw multimodal frames, we implemented deep-learning (DL) fusion models. These architectures combined visual features from RGB and IR recordings with tabular descriptors in a unified framework.\u003c/p\u003e\u003cp\u003eThe model pipeline consisted of:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eConvolutional neural networks (CNNs)\u003c/b\u003e applied to RGB and IR sequences to extract spatial features of melt-pool and foil dynamics.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eMultilayer perceptrons (MLPs)\u003c/b\u003e trained on tabular descriptors, including ridge, reflection, and thermal features.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eA \u003cb\u003etransformer encoder\u003c/b\u003e for fusing spatio-temporal information across modalities and capturing frame-to-frame dependencies.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eTraining was performed in PyTorch for up to 50 epochs with early stopping. The Adam optimiser was used with standard hyperparameters, and learning rates were tuned using validation folds. Cross-validation was again grouped by casting run to ensure generalisation. Implementation details, model configurations, and training schedules are documented in Supplementary Table S2, with validation diagnostics presented in Supplementary Table S3.\u003c/p\u003e\u003cp\u003ePerformance metrics for DL fusion models are summarised in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Representative results are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, including scatter plots of true vs. predicted gap, temporal predictions, calibration curves, and residual distributions. These outputs demonstrate that while DL fusion achieves accuracy comparable to XGBoost for gap prediction, it requires substantially greater resources.\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eEvaluation metrics and protocols\u003c/h2\u003e\u003cp\u003eAll models were assessed using standard regression metrics: mean absolute error (MAE), root-mean-square error (RMSE), and coefficient of determination (R\u0026sup2;). MAE was selected as the primary evaluation metric due to its robustness against outliers and ease of interpretation in physical units. RMSE was reported to highlight the impact of larger deviations, while R\u0026sup2; quantified overall explanatory power.\u003c/p\u003e\u003cp\u003eModel performance was compared across three levels:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eBaseline proxies\u003c/b\u003e (reflection height and infrared hot area).\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eFeature-based machine learning\u003c/b\u003e (XGBoost).\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eMultimodal deep learning fusion\u003c/b\u003e (CNN\u0026thinsp;+\u0026thinsp;MLP\u0026thinsp;+\u0026thinsp;transformer).\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003eTo prevent data leakage, all training and validation splits were grouped by casting run, ensuring that individual runs were never split across folds. For both ML and DL models, deterministic random seeds were used to ensure reproducibility. Extended details of fold assignments, hyperparameter ranges, and training schedules are given in Supplementary Table \u003cb\u003eS3\u003c/b\u003e.\u003c/p\u003e\u003cp\u003eA consolidated summary of model performance is provided in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, while detailed diagnostics, including per-fold statistics and error distributions, are reported in Supplementary Tables S3 and S4. Representative scatter, calibration, and residual plots are included in Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003eUse of AI tools\u003c/h2\u003e\u003cp\u003eLanguage editing assistance was provided using a large language model (ChatGPT, OpenAI), used solely for grammar and clarity improvements. All content was reviewed and approved by the authors.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003eBaseline model performance\u003c/h2\u003e\u003cp\u003eAnalytical proxies offered a useful lower-bound for assessing the predictive complexity of PFC monitoring.\u003c/p\u003e\u003cp\u003eThe gap\u0026ndash;reflection proxy, derived from the vertical reflection signature in side-view frames, captured only the broad scale of nozzle\u0026ndash;wheel gap variation. Across validation runs, it achieved a mean absolute error (MAE) of 0.12\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02 mm, with an RMSE of \u0026asymp;\u0026thinsp;0.16 mm and limited explanatory power (R\u0026sup2; in the range 0.3\u0026ndash;0.4). Scatter plots in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea demonstrate that predicted values consistently fell below the 1:1 line, indicating a systematic underestimation. Temporal traces (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb) show that the proxy tracked slow drifts in gap but failed to resolve rapid fluctuations. Residual analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec\u0026ndash;d) revealed bias and heteroscedasticity, with errors increasing disproportionately at larger gaps.\u003c/p\u003e\u003cp\u003eThe infrared hot-area proxy, designed to approximate flow through the fraction of foil surface above a fixed temperature threshold, was unable to recover meaningful correlations. Predictions of the flow proxy were essentially random with respect to ground-truth labels, yielding R\u0026sup2; \u0026asymp; 0 and MAE values like the mean of the dataset. This confirms that hot-area measures are insensitive to subtle melt-delivery changes.\u003c/p\u003e\u003cp\u003ePerformance metrics are summarised in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, while extended error distributions across runs are reported in Supplementary Table S3. Together, these results establish that heuristic proxies cannot provide reliable estimates of casting gap or flow and therefore set the lower bound for machine-learning benchmarks.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003eFeature-based ML performance (XGBoost)\u003c/h2\u003e\u003cp\u003eThe tabular dataset consisted of \u0026asymp;\u0026thinsp;8,800 aggregated samples with 27 descriptors per cast, derived from reflection, ridge, and thermal categories (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, Supplementary Table S5). XGBoost models were trained with group-wise cross-validation by casting run, ensuring no leakage between training and validation folds.\u003c/p\u003e\u003cp\u003eFor gap prediction, XGBoost achieved mean absolute error (MAE) values of 0.009\u0026ndash;0.011 mm and RMSE\u0026thinsp;\u0026asymp;\u0026thinsp;0.013 mm across folds. Coefficients of determination reached R\u0026sup2; = 0.96\u0026ndash;0.97, indicating that almost all variance in ground-truth gap could be explained by the engineered features. Scatter plots in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea show predictions clustering tightly along the 1:1 line. Temporal predictions (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb) tracked rapid fluctuations in gap with error bands rarely exceeding\u0026thinsp;\u0026plusmn;\u0026thinsp;0.01 mm. Calibration curves (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec) confirmed unbiased behaviour across the prediction range, while feature-importance rankings (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed) revealed that ridge-density and ridge-variation descriptors contributed most strongly, followed by reflection height and thermal gradients.\u003c/p\u003e\u003cp\u003eFor flow proxy prediction, performance remained poor. R\u0026sup2; values consistently fell below 0.1, with MAE\u0026thinsp;\u0026asymp;\u0026thinsp;2\u0026ndash;3% of the label range. This reflects the limited variability and indirect nature of the proxy, rather than a failure of the algorithm. Extended diagnostics, including fold-level performance and feature-ablation experiments, are provided in Supplementary Sections S2\u0026ndash;S3.\u003c/p\u003e\u003cp\u003eOverall, these results demonstrate that physics-informed descriptors are sufficient to achieve near-perfect gap prediction, while flow remains inadequately represented by available proxy labels.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003eDeep learning fusion results\u003c/h2\u003e\u003cp\u003eThe multimodal DL fusion model combined RGB and IR frames with tabular descriptors, using convolutional encoders, MLPs, and a transformer fusion module (Supplementary Table S2).\u003c/p\u003e\u003cp\u003eFor gap prediction, DL models achieved MAE values of 0.010\u0026ndash;0.012 mm and RMSE\u0026thinsp;\u0026asymp;\u0026thinsp;0.015 mm, with R\u0026sup2; = 0.96\u0026ndash;0.97, closely matching XGBoost performance. Scatter plots (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea) showed tight alignment with the 1:1 line, while temporal traces (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb) demonstrated stable tracking across casting runs. Learning curves confirmed smooth convergence without overfitting. Calibration analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ec) showed strong agreement between predicted and true values across all bins, while residual distributions (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ed) were narrow and centred close to zero.\u003c/p\u003e\u003cp\u003eFor flow proxy prediction, DL fusion again did not improve over baseline ML approaches, with R\u0026sup2; \u0026asymp; 0 and MAE comparable to dataset averages. This underscores the inherent limitation of proxy labels, rather than the model architecture.\u003c/p\u003e\u003cp\u003ePerformance summaries are reported in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, with fold-level diagnostics included in Supplementary Table S3. The comparison demonstrates that while DL can effectively process multimodal data streams, it does not yield meaningful gains over feature-based ML in the context of gap prediction for this dataset.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003eFoil-quality analysis (ridge width, thermal gradients)\u003c/h2\u003e\u003cp\u003eBeyond process-variable prediction, direct video-based analysis provided quantitative measures of foil integrity. Ridge and thermal descriptors extracted through computer vision pipelines were evaluated across stable and unstable casting regimes.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\u003ch2\u003eRidge analysis\u003c/h2\u003e\u003cp\u003eFront-view recordings were used to compute ridge-density maps and ridge-width histograms. Under stable conditions, ridge spacing was uniform across the foil, with standard deviation typically\u0026thinsp;\u0026lt;\u0026thinsp;0.2 mm. Unstable runs exhibited ridge clustering, leading to broadened histograms with long tails. In Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea\u0026ndash;b, the contrast is evident: stable runs show narrow, symmetric distributions centred around \u0026asymp;\u0026thinsp;1.5 mm, while unstable runs show skewed and multimodal distributions. These ridge features were also among the highest-ranked predictors in XGBoost (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\u003ch2\u003eThermal analysis\u003c/h2\u003e\u003cp\u003eInfrared recordings provided cross-foil temperature profiles, which were used to calculate edge-to-centre gradients. Stable runs exhibited mild, symmetric gradients of \u0026asymp;\u0026thinsp;5\u0026ndash;10\u0026deg;C, while unstable runs showed steep gradients\u0026thinsp;\u0026gt;\u0026thinsp;25\u0026deg;C concentrated near foil edges. In Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec, representative thermal profiles illustrate this difference. Scatter analysis in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ed confirmed a strong correlation between ridge clustering and thermal imbalance runs with higher ridge density consistently coincided with larger edge gradients.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e\u003ch2\u003eCombined interpretation\u003c/h2\u003e\u003cp\u003eThese results establish ridge and thermal descriptors as direct, physically interpretable indicators of foil integrity. Unlike proxy-based monitoring, they allow early detection of instability through surface and thermal signatures. A summary of ridge and thermal features is included in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, with extended ranges and definitions in Supplementary Table S5.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\u003ch2\u003eAblation studies\u003c/h2\u003e\u003cp\u003eTo evaluate the relative contribution of feature categories, systematic ablation experiments were performed on the XGBoost models. Ridge, reflection, and thermal descriptors were selectively removed, and model performance was reassessed on held out runs.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\u003ch2\u003eOverall results\u003c/h2\u003e\u003cp\u003eWith the full feature set, the model achieved MAE\u0026thinsp;\u0026asymp;\u0026thinsp;0.010 mm for gap prediction. Removing ridge features caused the largest performance degradation, with MAE increasing to \u0026asymp;\u0026thinsp;0.014 mm, corresponding to a\u0026thinsp;\u0026asymp;\u0026thinsp;40% relative error increase. Excluding reflection features raised MAE by \u0026asymp;\u0026thinsp;12%, while omission of thermal descriptors increased error by \u0026asymp;\u0026thinsp;20% and reduced sensitivity to foil-quality indicators. Results are summarised in Table\u0026nbsp;4.\u003c/p\u003e\u003cdiv id=\"Sec23\" class=\"Section3\"\u003e\u003ch2\u003eFeature importance aggregation\u003c/h2\u003e\u003cp\u003eFeature-importance scores were also aggregated across 20 representative casting runs. As shown in Fig.\u0026nbsp;6b, ridge descriptors consistently contributed the largest share of predictive power, followed by reflection and thermal features. Ridge width variation and ridge density were ranked as the top individual predictors, consistent with their strong physical relationship to gap stability.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec24\" class=\"Section2\"\u003e\u003ch2\u003eInterpretation\u003c/h2\u003e\u003cp\u003eTogether, the ablation and feature-importance results confirm that all three categories provide complementary information. Ridge descriptors dominate gap estimation, reflection features add robustness, and thermal descriptors enhance sensitivity to foil quality. None of the categories is dispensable: models trained without one class of descriptors showed reduced generalisation and increased bias. Extended ablation configurations, including feature subsets and cross-run error distributions, are reported in Supplementary Table S5.\u003c/p\u003e\u003cp\u003eRepresentative results are shown in Fig.\u0026nbsp;6a\u0026ndash;b: panel (a) summarises MAE increases under different ablation conditions, while panel (b) visualises feature-importance distributions across runs.\u003c/p\u003e\u003cdiv id=\"Sec25\" class=\"Section3\"\u003e\u003ch2\u003eIndustrial validation\u003c/h2\u003e\u003cp\u003eTo assess the transferability of our framework to production environments, we analysed factory-acquired recordings provided by an industrial partner. These included RGB views of the nozzle\u0026ndash;wheel gap region and infrared recordings of the cooling wheel and foil during continuous operation. Unlike laboratory runs, these datasets were captured under factory-scale casting conditions, where operating variability, environmental noise, and production throughput present additional challenges.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec26\" class=\"Section3\"\u003e\u003ch2\u003eRGB analysis\u003c/h2\u003e\u003cp\u003eFactory RGB frames (Supplementary Fig.\u0026nbsp;7a\u0026ndash;b) revealed non-uniform ridge formation across foil width, with clustering events corresponding to unstable flow delivery. Applying our ridge-detection pipeline extracted density profiles comparable to those observed in unstable laboratory runs (Fig.\u0026nbsp;7c). Ridge histograms showed broadened distributions with skewed tails, confirming instability during high-throughput operation.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec27\" class=\"Section3\"\u003e\u003ch2\u003eThermal analysis\u003c/h2\u003e\u003cp\u003eFactory IR frames captured temperature distributions across the foil and cooling wheel (Fig.\u0026nbsp;7d\u0026ndash;e). Analysis revealed edge-dominated gradients exceeding 25\u0026ndash;30\u0026deg;C, with thermal asymmetry persisting over long-time windows. These findings mirror laboratory evidence that strong edge-to-centre gradients coincide with ridge clustering and reduced amorphicity. Quantitative profiles are summarised in Supplementary Fig.\u0026nbsp;7f, showing direct alignment with the diagnostic patterns identified in controlled experiments.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec28\" class=\"Section2\"\u003e\u003ch2\u003eInterpretation and industrial significance\u003c/h2\u003e\u003cp\u003eThese results demonstrate that the ridge- and thermal-analysis pipelines generalise beyond laboratory settings. In practice, factory engineers could obtain comparable thermal profiles using commercial IR cameras, but our integrated CV\u0026ndash;ML framework enables predictive interpretation rather than descriptive monitoring. This distinction is critical: while thermal imaging provides accessible indicators, the fusion of ridge, thermal, and metadata descriptors adds robustness against noise and variability, making duplication of the predictive models from raw factory images non-trivial.\u003c/p\u003e\u003cp\u003eBy validating on factory-scale data, we confirm that the framework not only reproduces laboratory findings but also provides diagnostic indicators relevant to magnetic property stability under real industrial operating conditions.\u003c/p\u003e\u003c/div\u003e"},{"header":"Discussion","content":"\u003cdiv id=\"Sec30\" class=\"Section2\"\u003e\u003ch2\u003eMachine learning versus deep learning approaches\u003c/h2\u003e\u003cp\u003eOur results demonstrate that feature-based ML models, particularly XGBoost trained on physics-informed descriptors, provide state-of-the-art performance for gap prediction in planar-flow casting. With MAE values below 0.01 mm and R\u0026sup2; consistently above 0.96 (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), these models capture nearly all variance in ground-truth gap. By contrast, deep-learning fusion models achieve similar accuracy but at substantially higher computational cost and with more complex optimisation requirements. This outcome highlights the strength of physics-inspired descriptors: they encode key process phenomena such as ridge density, reflection geometry, and thermal imbalance, which align naturally with the underlying physics of melt\u0026ndash;wheel dynamics.\u003c/p\u003e\u003cp\u003eThe convergence of ML and DL performance further suggests that the dataset is information-rich at the feature level. In this context, XGBoost offers a lightweight, interpretable, and industrially deployable solution, while DL fusion remains valuable for extending to new modalities or for exploratory research.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec31\" class=\"Section2\"\u003e\u003ch2\u003eRidge and thermal descriptors as quality indicators\u003c/h2\u003e\u003cp\u003eA central outcome of this study is the confirmation that ridge, and thermal descriptors serve as direct, video-derived indicators of foil integrity. Ridge-density histograms and ridge-width distributions (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea\u0026ndash;b) distinguished stable from unstable runs, while thermal profiles across the foil width (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec\u0026ndash;d) revealed edge-to-centre gradients that correlated strongly with surface clustering. The ablation experiments (Supplementary Fig.\u0026nbsp;6, Table\u0026nbsp;4) reinforced this conclusion: ridge descriptors accounted for \u0026asymp;\u0026thinsp;40% of predictive power, while thermal descriptors provided complementary sensitivity to quality-related phenomena. This combination provides both predictive accuracy and physical interpretability, bridging the gap between data-driven modelling and metallurgical insight.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec32\" class=\"Section2\"\u003e\u003ch2\u003eIndustrial implications\u003c/h2\u003e\u003cp\u003eValidation with factory-acquired data (Supplementary Fig.\u0026nbsp;7) demonstrated that our framework generalises to production-scale environments. Ridge clustering and thermal imbalance were captured under real industrial conditions, confirming the diagnostic value of these descriptors. Importantly, while thermal imaging can be implemented with standard factory hardware, predictive interpretation requires the integrated CV\u0026ndash;ML framework presented here. This distinction reduces the risk of straightforward duplication by external engineers and emphasises the added value of feature integration and model fusion.\u003c/p\u003e\u003cp\u003eFrom an industrial perspective, reliable monitoring of ridge and thermal instabilities addresses one of the most pressing limitations of PFC: the degradation of magnetic properties under unstable casting. By linking melt\u0026ndash;wheel dynamics directly to foil quality, our framework provides a foundation for stabilising production and ensuring that full-width amorphous foils meet performance targets.\u003c/p\u003e\u003cdiv id=\"Sec33\" class=\"Section3\"\u003e\u003ch2\u003eSensitivity and duplication risks\u003c/h2\u003e\u003cp\u003eA concern raised in discussions with industrial partners is whether factory engineers could replicate the approach using only thermal imaging. While thermal profiles alone do provide useful indicators, they are inherently descriptive. The predictive performance demonstrated here arises from integrating multiple feature categories, synchronised video, and metadata across runs. Without access to the structured dataset, annotation schema (Supplementary Tables S1\u0026ndash;S5), and trained models, reproducing this pipeline would require significant effort and expertise.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec34\" class=\"Section3\"\u003e\u003ch2\u003eFuture directions\u003c/h2\u003e\u003cp\u003eWhile this study establishes a reproducible dataset and framework, several opportunities remain. First, a new round of factory data collection with optimised camera placements could strengthen modelling of melt-pool instabilities and wheel\u0026ndash;melt heat transfer. Second, integrating magnetic property measurements of cast foils would allow direct linkage between process signatures and end-use performance. Third, coupling data-driven methods with physics-based Multiphysics simulations may enable hybrid models capable of real-time control. Finally, active learning strategies could adaptively select experiments to maximise model robustness across facilities and alloy compositions.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study introduces a reproducible, data-driven framework for monitoring planar-flow casting of amorphous steel foils. By combining synchronised RGB and infrared recordings with detailed process metadata, we constructed a structured dataset of \u0026asymp;\u0026thinsp;200 casting runs. Physics-informed features derived from ridge profiles, reflection intensity, and thermal gradients provided interpretable links between melt\u0026ndash;wheel dynamics and foil integrity.\u003c/p\u003e\u003cp\u003eFeature-based machine-learning models, particularly XGBoost, achieved near-perfect prediction of nozzle\u0026ndash;wheel gap with MAE\u0026thinsp;\u0026lt;\u0026thinsp;0.01 mm and R\u0026sup2; \u0026asymp; 0.97. Deep-learning fusion networks offered comparable performance but required greater computational resources. Baseline analytical proxies proved insufficient, confirming the value of both engineered descriptors and multimodal learning. Ridge-width distributions and thermal profiles emerged as direct indicators of foil quality, distinguishing stable and unstable casting regimes.\u003c/p\u003e\u003cp\u003eValidation with factory-acquired data demonstrated that ridge and thermal descriptors generalise to industrial conditions. These measures address one of the most critical barriers to large-scale deployment of amorphous foils: instability-driven degradation of magnetic properties. By providing predictive, interpretable monitoring tools, this framework establishes the foundation for closed-loop, automated optimisation of PFC.\u003c/p\u003e\u003cp\u003eFuture extensions will focus on expanding datasets across facilities, capturing melt-pool and wheel\u0026ndash;melt heat transfer in greater detail, and linking process signatures directly to magnetic performance. Together, these efforts position the approach as a practical pathway toward autonomous control of amorphous foil manufacturing.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eCompeting Interests\u003c/h2\u003e\u003cp\u003eThe authors declare no competing interests. Patent applications related to aspects of this work have been filed by CSIRO (Application No. PCT/AU2024/051282, published as WO 2025/111660; and Application No. PCT/AU2025/050062, published as WO/2025/160623).\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eR.D. led the design of the study, developed the machine-learning and computer-vision framework, and drafted the manuscript. M.K. was responsible for experimental data collection and dataset assembly. D.L. contributed to experimental design, process interpretation, and manuscript refinement. All authors discussed the results, contributed to interpretation, and reviewed the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThis work was supported by the CSIRO Future Digital Manufacturing Fund (FDMF) and by contributions from the CSIRO Research Units (Data61 and Manufacturing). The authors thank technical staff for assistance with planar-flow casting experiments and acknowledge valuable discussions with industrial partners that informed the design of validation experiments.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets generated and analysed during the current study are available from the corresponding author upon reasonable request. Custom computer-vision scripts and machine-learning pipelines developed for annotation, feature extraction, and model training are likewise available for non-commercial research purposes upon reasonable request.\u003c/p\u003e\n\u003ch3\u003eData and Code Availability\u003c/h3\u003e\n\u003cp\u003eThe datasets generated and analysed during the current study are available from the corresponding author upon reasonable request. Custom computer-vision scripts and machine-learning pipelines developed for annotation, feature extraction, and model training are likewise available for non-commercial research purposes upon reasonable request.\u003c/p\u003e\u003ch2\u003eFunding Declaration\u003c/h2\u003e\u003cp\u003eThis research was supported by internal CSIRO strategic funding through the Future Digital Manufacturing Fund (FDMF) and contributions from the Data61 and Manufacturing Business Units.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eTheisen, E. A. \u0026amp; Weinstein, S. J. An overview of planar flow casting of thin metallic glasses and its relation to slot coating of liquid films. \u003cem\u003eJ. Coat. Technol. Res.\u003c/em\u003e \u003cb\u003e19\u003c/b\u003e, 49\u0026ndash;60. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11998-021-00503-y\u003c/span\u003e\u003cspan address=\"10.1007/s11998-021-00503-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCarpenter, J. K. \u0026amp; Steen, P. H. Planar-flow spin-casting of molten metals: process behaviour. \u003cem\u003eJ. Mater. Sci.\u003c/em\u003e \u003cb\u003e27\u003c/b\u003e, 215\u0026ndash;225. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/BF02403666\u003c/span\u003e\u003cspan address=\"10.1007/BF02403666\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (1992).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMattson, J. W., Theisen, E. A. \u0026amp; Steen, P. H. Rapid solidification forming of glassy and crystalline ribbons by planar flow casting. \u003cem\u003eChem. Eng. Sci.\u003c/em\u003e \u003cb\u003e192\u003c/b\u003e, 1198\u0026ndash;1208. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ces.2018.07.017\u003c/span\u003e\u003cspan address=\"10.1016/j.ces.2018.07.017\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSu, Y-G., Chen, F., Wu, C-Y., Chang, M-H. \u0026amp; Chung, C-A. Effects of manufacturing parameters in planar flow casting process on ribbon formation and puddle evolution of Fe78Si9B13 alloy. \u003cem\u003eISIJ Int.\u003c/em\u003e \u003cb\u003e55\u003c/b\u003e, 2383\u0026ndash;2390 (2015).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi, D., Zhuang, J., Liu, T., Lu, Z. \u0026amp; Zhou, S. The pressure loss and ribbon thickness prediction in gap controlled planar-flow casting process. \u003cem\u003eJ. Mater. Process. Technol.\u003c/em\u003e \u003cb\u003e211\u003c/b\u003e, 1764\u0026ndash;1767. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jmatprotec.2011.05.019\u003c/span\u003e\u003cspan address=\"10.1016/j.jmatprotec.2011.05.019\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2011).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKarcher, C. \u0026amp; Steen, P. H. Solidification and downstream meniscus prediction in the planar-flow melt spinning process. \u003cem\u003eChem. Eng. Sci.\u003c/em\u003e \u003cb\u003e62\u003c/b\u003e, 5145\u0026ndash;5150. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ces.2007.02.042\u003c/span\u003e\u003cspan address=\"10.1016/j.ces.2007.02.042\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2007).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCox, B. L. \u0026amp; Steen, P. H. Herringbone' defect formation in planar-flow melt spinning. \u003cem\u003eJ. Mater. Process. Technol.\u003c/em\u003e \u003cb\u003e213\u003c/b\u003e, 1743\u0026ndash;1752. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jmatprotec.2013.04.010\u003c/span\u003e\u003cspan address=\"10.1016/j.jmatprotec.2013.04.010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2013).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eByrne, C., Theisen, E., Reed, B. \u0026amp; Steen, P. H. Capillary puddle vibrations linked to casting-defect formation in planar-flow melt spinning. \u003cem\u003eMetall. Mater. Trans. B\u003c/em\u003e. \u003cb\u003e37\u003c/b\u003e, 445\u0026ndash;456. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11663-006-0029-4\u003c/span\u003e\u003cspan address=\"10.1007/s11663-006-0029-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2006).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChung, C-A., Su, Y-G. \u0026amp; Chen, F. Tuning the Planar-Flow Melt-Spinning Process Subject to Operability Constraints. JOM 66, 2082\u0026ndash;2088 (2014). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11837-014-0982-3\u003c/span\u003e\u003cspan address=\"10.1007/s11837-014-0982-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCho, K. H., Lee, K. A., Kim, M. C. \u0026amp; Yoon, J. M. Numerical modeling of planar flow casting process. \u003cem\u003eSolid State Phenom.\u003c/em\u003e \u003cb\u003e116\u0026ndash;117\u003c/b\u003e, 106\u0026ndash;109 (2006). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.4028/www.scientific.net/SSP.116-117.106\u003c/span\u003e\u003cspan address=\"https://doi.org/10.4028/www.scientific.net/SSP.116-117.106\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDas, S., Rao, P. S. \u0026amp; Ghosh, S. Flow CFD analysis of 2D transient melt flow from crucible onto the rotating wheel in PFMS. Mater. Today Proc. 4, 9810\u0026ndash;9816 (2017). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.matpr.2017.06.289\u003c/span\u003e\u003cspan address=\"10.1016/j.matpr.2017.06.289\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDas, S., Rao, P. S. \u0026amp; Ghosh, S. Obtaining stable puddle and thinner ribbons during planar flow melt spinning: a 2D analysis. Mater. Today Proc. 4, 9656\u0026ndash;9662 (2017). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.matpr.2017.06.035\u003c/span\u003e\u003cspan address=\"10.1016/j.matpr.2017.06.035\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePradhan, A. \u0026amp; Ghosh, S. Flow dynamics in the melt puddle during planar flow melt spinning: A numerical study. Mater. Today Proc. 4, 9829\u0026ndash;9838 (2017). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.matpr.2017.06.311\u003c/span\u003e\u003cspan address=\"10.1016/j.matpr.2017.06.311\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHeine, G. \u0026amp; Fischer, K. The effect of the main processing parameters on the geometry of melt-spun ribbons. \u003cem\u003eZ. Metallkde\u003c/em\u003e. \u003cb\u003e75\u003c/b\u003e, 403\u0026ndash;409 (1984).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi, D. \u0026amp; Lu, Z. Effects of aging on the cyclical thermal shock response of a copper\u0026ndash;beryllium alloy as a substrate of cooling wheel in planar flow casting process. \u003cem\u003eMater. Res. Express\u003c/em\u003e. \u003cb\u003e7\u003c/b\u003e, 116511. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1088/2053-1591/abae58\u003c/span\u003e\u003cspan address=\"10.1088/2053-1591/abae58\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi, D., Wang, W., Liu, T., Li, L. \u0026amp; Lu, Z. Reducing the core losses of Fe\u0026ndash;Si\u0026ndash;B amorphous alloy ribbons by high cooling rate planar flow casting. \u003cem\u003eMaterials\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 894. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/ma15030894\u003c/span\u003e\u003cspan address=\"10.3390/ma15030894\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi, Y. et al. Excellent magnetic properties obtained in planar flow casting Fe\u0026ndash;6.5%Si ribbons. \u003cem\u003eMater. Des.\u003c/em\u003e \u003cb\u003e235\u003c/b\u003e, 112556. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.matdes.2024.112556\u003c/span\u003e\u003cspan address=\"10.1016/j.matdes.2024.112556\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang, Y. et al. Mechanistic investigation of thickness\u0026ndash;controlled microstructure and performance in planar flow casting. \u003cem\u003eJ. Mater. Process. Technol.\u003c/em\u003e \u003cb\u003e324\u003c/b\u003e, 118924. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jmatprotec.2025.118924\u003c/span\u003e\u003cspan address=\"10.1016/j.jmatprotec.2025.118924\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGreer, A. L. Metallic glasses. \u003cem\u003eMRS Bull.\u003c/em\u003e \u003cb\u003e48\u003c/b\u003e, 842\u0026ndash;854. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1557/s43577-023-00586-5\u003c/span\u003e\u003cspan address=\"10.1557/s43577-023-00586-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTheisen, E. A. \u0026amp; Weinstein, S. J. An overview of planar flow casting of thin metallic glasses (preprint). arXiv, arXiv:2104.09251. (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang, R. et al. Review of surface defect detection of steel products based on machine vision. \u003cem\u003eIET Image Process.\u003c/em\u003e \u003cb\u003e17\u003c/b\u003e, 3039\u0026ndash;3060. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1049/ipr2.12647\u003c/span\u003e\u003cspan address=\"10.1049/ipr2.12647\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang, R. et al. A deep\u0026ndash;learning\u0026ndash;based approach for fast and robust steel surface defect classification. \u003cem\u003ePattern Recogn.\u003c/em\u003e \u003cb\u003e103\u003c/b\u003e, 107312. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.patcog.2019.107312\u003c/span\u003e\u003cspan address=\"10.1016/j.patcog.2019.107312\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang, Z. et al. A systematic review of deep learning approaches for surface defect detection in manufacturing. \u003cem\u003eEng. Appl. Artif. Intell.\u003c/em\u003e \u003cb\u003e127\u003c/b\u003e, 107327. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.engappai.2023.107327\u003c/span\u003e\u003cspan address=\"10.1016/j.engappai.2023.107327\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKaya, M. \u0026amp; Karak\u0026ouml;se, M. Systematic review of steel surface defect detection methods on the open datasets. \u003cem\u003eLNCS\u003c/em\u003e \u003cb\u003e14470\u003c/b\u003e, 35\u0026ndash;58 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi, H. et al. Research on steel surface defect classification method based on deep learning. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 12876. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-024-58643-1\u003c/span\u003e\u003cspan address=\"10.1038/s41598-024-58643-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHossain, R. E. N., Lewis, J. \u0026amp; Moore, A. L. In\u0026ndash;situ infrared temperature sensing for real\u0026ndash;time defect detection in additive manufacturing. \u003cem\u003eAddit. Manuf.\u003c/em\u003e \u003cb\u003e47\u003c/b\u003e, 102328. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.addma.2021.102328\u003c/span\u003e\u003cspan address=\"10.1016/j.addma.2021.102328\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKavale, N. S. et al. Process monitoring using synchronized path infrared thermography (SPIT). \u003cem\u003eSensors\u003c/em\u003e \u003cb\u003e22\u003c/b\u003e, 5943. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/s22165943\u003c/span\u003e\u003cspan address=\"10.3390/s22165943\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAbdelrahman, M., Leach, R. \u0026amp; Maskery, I. Assisted defect detection by in\u0026ndash;process monitoring of additive manufacturing using optical imaging and infrared thermography. \u003cem\u003eAddit. Manuf.\u003c/em\u003e \u003cb\u003e65\u003c/b\u003e, 103487. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.addma.2023.103487\u003c/span\u003e\u003cspan address=\"10.1016/j.addma.2023.103487\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGarc\u0026iacute;a\u0026ndash;Gonz\u0026aacute;lez, A., Alkorta, J. \u0026amp; Sevillano, J. G. Detection of sub\u0026ndash;superficial defects by infrared thermography in parts manufactured by additive manufacturing. \u003cem\u003eInt. J. Adv. Manuf. Technol.\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s40964-025-01313-z\u003c/span\u003e\u003cspan address=\"10.1007/s40964-025-01313-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGopaluni, R. B. et al. Modern machine learning tools for monitoring and control of industrial processes: A survey. \u003cem\u003eIFAC\u0026ndash;PapersOnLine\u003c/em\u003e \u003cb\u003e53\u003c/b\u003e, 12167\u0026ndash;12178. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ifacol.2020.12.2800\u003c/span\u003e\u003cspan address=\"10.1016/j.ifacol.2020.12.2800\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLawrence, N. P. et al. Machine learning for industrial sensing and control: A survey and practical perspective. \u003cem\u003eAnnu. Rev. Control\u003c/em\u003e. \u003cb\u003e58\u003c/b\u003e, 100985. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.arcontrol.2024.100985\u003c/span\u003e\u003cspan address=\"10.1016/j.arcontrol.2024.100985\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTercan, H. \u0026amp; Meisen, T. Machine learning and deep learning based predictive quality in manufacturing: a systematic review. \u003cem\u003eJ. Intell. Manuf.\u003c/em\u003e \u003cb\u003e33\u003c/b\u003e, 1879\u0026ndash;1905. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10845-022-01963-8\u003c/span\u003e\u003cspan address=\"10.1007/s10845-022-01963-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLawrence, N. P., Gopaluni, R. B., Lee, J. M. \u0026amp; Huang, B. Machine learning for industrial sensing and control: a survey and practical perspective (preprint). arXiv, arXiv:2401.13836. (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGopaluni, R. B. et al. Modern machine learning tools for monitoring and control of industrial processes: A survey (preprint). arXiv, arXiv:2209.11123. (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang, Q. et al. Steel surface defect detection algorithm in complex background via improved vision transformers. \u003cem\u003eMeasurement\u003c/em\u003e \u003cb\u003e231\u003c/b\u003e, 114547. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.measurement.2024.114547\u003c/span\u003e\u003cspan address=\"10.1016/j.measurement.2024.114547\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e\u0026Ouml;zkan, A. \u0026amp; U\u0026ccedil;ar, A. Detection of defects on metal surfaces based on deep learning. \u003cem\u003eAppl. Sci.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 1406. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/app15031406\u003c/span\u003e\u003cspan address=\"10.3390/app15031406\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKubik, D. et al. Reliable machine learning models for manufacturing processes. \u003cem\u003eInt. J. Adv. Manuf. Technol.\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00170-025-16539-y\u003c/span\u003e\u003cspan address=\"10.1007/s00170-025-16539-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYan, P. et al. \u003cem\u003eLearning actionable world models for industrial process control\u003c/em\u003e (arXiv, 2025). arXiv:2503.01411.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\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\u003eDataset overview. Summary of the experimental dataset used in this study, including the number of casting runs, frames analysed, experimental configurations, total data volume, and metadata fields available.\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\u003eProperty\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDetails\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNumber of casting runs\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e~\u0026thinsp;200 runs across multiple conditions\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFrames analysed (RGB\u0026thinsp;+\u0026thinsp;IR)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e~\u0026thinsp;300,000 frames (synchronised RGB/IR)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eExperimental setups\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSide view, front view, infrared imaging (subset)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTotal dataset volume\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e~\u0026thinsp;20 GB structured dataset\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMetadata fields\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eGap (mm), flow proxy (% change in optical flow velocity), wheel speed (rpm), melt temperature (\u0026deg;C), nozzle geometry\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 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePhysics-informed descriptors used for ML. Features were extracted from video frames and grouped into ridge, thermal, and reflection descriptors, with units and physical relevance.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\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=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCategory\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDescriptor\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eDefinition / Units\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003ePhysical relevance\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRidge\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRidge density\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eridges per mm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSurface stability indicator\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRidge\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRidge width (mean)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003emm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eFoil uniformity measure\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRidge\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRidge width (std)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003emm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eVariation in ridge spacing\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eThermal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eThermal gradient variance\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u0026deg;C\u0026sup2;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eEdge vs centre thermal imbalance\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eThermal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eEdge temperature\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u0026deg;C\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eHot-spot indicator\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eReflection\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eReflection intensity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eNo Unit\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eGap proxy from RGB\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGeometry\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFoil width\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003emm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eCasting geometry constraint\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=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eBaseline vs ML model performance. Comparison of baseline proxies, XGBoost, and deep-learning fusion models for gap and flow prediction. Metrics averaged across validation folds.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTarget\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eMAE\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRMSE\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eR\u0026sup2;\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eCalibration error\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBaseline proxy\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eGap\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.12 mm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.15 mm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eHigh bias\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBaseline proxy\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFlow (% change, optical flow velocity)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.30%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.35%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u0026asymp;\u0026thinsp;0.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eHigh bias\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eXGBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eGap\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.010 mm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.013 mm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.97\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eLow\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eXGBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFlow (% change, optical flow velocity)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.18%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.22%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eModerate\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDL Fusion\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eGap\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.011 mm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.014 mm\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eLow\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDL Fusion\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFlow (% change, optical flow velocity)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.17%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.21%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.15\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eModerate\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8042276/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8042276/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAmorphous metallic foils produced by planar-flow casting (PFC) are central to emerging high-efficiency transformer and motor technologies. Despite its promise, the process remains highly sensitive to fluctuations in melt delivery, nozzle\u0026ndash;wheel gap, and interfacial heat transfer, leading to free-surface ridging, uneven thickness, and compromised magnetic properties. Conventional monitoring, based on indirect variables or post-cast inspection, provides delayed feedback and cannot resolve the rapid dynamics of the melt\u0026ndash;wheel interface.\u003c/p\u003e\u003cp\u003eHere we present a data-driven framework for PFC monitoring that integrates synchronised video and infrared imaging with machine-learning analysis. A multimodal dataset of \u0026asymp;\u0026thinsp;200 runs (\u0026asymp;\u0026thinsp;300,000 frames, \u0026asymp;\u0026thinsp;20 GB) was collected with front- and side-view cameras and infrared thermal imaging, paired with detailed metadata. Physics-inspired descriptors\u0026mdash;including ridge density, reflection intensity, and thermal gradients\u0026mdash;were extracted through automated computer vision pipelines and benchmarked against raw-frame deep learning. Analytical proxies, gradient-boosted decision trees, and multimodal fusion networks were systematically evaluated.\u003c/p\u003e\u003cp\u003eFeature-based models achieved near-perfect gap prediction, while ridge and thermal-gradient analysis provided direct indicators of foil integrity. Deep-learning fusion models offered complementary robustness but at higher computational cost. Together, these results demonstrate a reproducible approach for linking process dynamics with foil quality, establishing a foundation for automated, closed-loop optimisation of amorphous steel-foil manufacturing.\u003c/p\u003e","manuscriptTitle":"AI-enabled Monitoring of Planar-Flow Casting of Amorphous Steel Foils","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-20 11:35:19","doi":"10.21203/rs.3.rs-8042276/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"fba06fcb-8935-4214-b405-bb414d30fed4","owner":[],"postedDate":"November 20th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":58200433,"name":"Physical sciences/Engineering"},{"id":58200434,"name":"Physical sciences/Materials science"},{"id":58200435,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2026-01-02T04:08:52+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-20 11:35:19","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8042276","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8042276","identity":"rs-8042276","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}