*Adapting CatBoost Machine Learning for Texture Evolution Prediction in Non-Oriented Electrical Steels: A Synthetic-Data Methodology Demonstration**
Year: 2026
Type: preprint
Abstract: # Alternative Title**Data-Driven Texture Engineering: Transferring a CatBoost Workflow from Magnetocrystalline Anisotropy to Electrical Steel Processing** ## Subtitle**A template for predicting Goss texture intensity and core loss from composition and thermomechanical processing parameters, with feature-importance interpretation and a pathway to experimental validation** ## Detailed Description This research paper presents a methodological adaptation of the CatBoost gradient-boosting framework—originally developed for predicting magnetocrystalline anisotropy from single-crystal constants \(K_1\) and \(K_2\)—to the problem of texture evolution in non-oriented (NO) electrical steels. The motivation is that traditional physically-based models of texture development are computationally expensive and require many assumptions, whereas machine learning can learn nonlinear processing–texture relationships directly from data. Because no real industrial dataset is available at this stage, the paper constructs a **212-point synthetic dataset** from simplified, physically motivated response surfaces that embed known metallurgical trends: higher annealing temperature and cold reduction promote Goss texture \(\{110\}\langle001\rangle\), while higher Al content and lower temperature favor the detrimental \(\gamma\)-fiber \(\{111\}\langle uvw\rangle\). The workflow mirrors the anisotropy study: features include composition (Si, Al, Mn, C), processing parameters (slab reheating, finish rolling, coiling, cold reduction, annealing temperature/time, heating rate), and microstructure (grain size), with targets being Goss intensity, \(\gamma\)-fiber intensity, and core loss \(P_{1.5/50}\). CatBoost is trained with depth 8, learning rate 0.03, and 5-fold cross-validation, and is compared against linear regression, random forest, and XGBoost. On the synthetic test set, CatBoost achieves the highest \(R^2\) for Goss intensity (0.84), outperforming XGBoost (0.78) and random forest (0.72). Feature importance analysis identifies annealing temperature and cold reduction as dominant predictors, consistent with metallurgical theory. A CatBoost classifier predicts dominant texture class (Goss, Cube, \(\gamma\)-fiber, random) with 78% accuracy. The paper explicitly cautions that these metrics reflect recovery of synthetic relationships, not demonstrated predictive power on real steel, and that random train/test splitting may overestimate performance. It outlines a validation pathway: collect real EBSD/XRD texture data and magnetic property measurements for 200–500 samples, use grouped cross-validation by coil/heat, hold out entire composition grades for extrapolation tests, and combine SHAP analysis with physical models. The work provides a reproducible template for data-driven texture engineering in electrical steels and a clear statement of the gap between synthetic demonstration and industrial deployment.
Author: geruganti, sudhakar
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Keywords: CatBoost, non-oriented electrical steel, Goss texture, core loss, synthetic training data, annealing temperature, cold reduction, thermomechanical processing, feature importance, texture classification, magnetocrystalline anisotropy, model validation, response surface, grain size, aluminum content, silicon content, manganese
Open Access status: green
OpenAlex ID: W7236429281
