*Hyperparameter Sensitivity of CatBoost for Texture Prediction in Non-Oriented Electrical Steels: Effects of Learning Rate and Maximum Tree Depth on Predictive Performance**
Year: 2026
Type: preprint
Abstract: * ## Alternative Title**Tuning CatBoost for Data-Driven Texture Engineering: A Systematic Study of Learning Rate and Tree Depth Using a Synthetic Electrical-Steel Dataset** ## Subtitle**A grid-search analysis (depth 4–10, learning rate 0.01–0.10) identifying depth 8 and learning rate 0.03 as optimal, with overfitting diagnostics, iteration effects, and practical tuning recommendations for small tabular materials datasets** ## Detailed Description This research paper investigates how two critical CatBoost hyperparameters—**learning rate** and **maximum tree depth**—affect the model’s ability to predict Goss texture intensity \(\{110\}\langle001\rangle\) and core loss \(P_{1.5/50}\) in non-oriented electrical steels. The study uses the same 212-point synthetic dataset from the companion methodology paper, where targets are generated from physically motivated response surfaces that embed known metallurgical trends (annealing temperature and cold reduction promote Goss; Al and low temperature promote \(\gamma\)-fiber). A full grid search over tree depth \(\{4,6,8,10\}\) and learning rate \(\{0.01,0.03,0.10\}\) is performed with 500 iterations and 5-fold cross-validation. Results show a clear depth–learning-rate interaction: shallow trees (depth 4) underfit even at high learning rate (best test \(R^2 = 0.77\)), while deep trees (depth 10) with learning rate 0.10 overfit, with a train–test gap of 0.15. The optimal configuration is **depth 8 with learning rate 0.03**, achieving test \(R^2 = 0.84\), RMSE = 0.33 m.r.d., and MAE = 0.27 m.r.d. Learning rate 0.01 requires more than 500 iterations to converge; extending to 1000–1500 iterations raises test \(R^2\) from 0.78 to 0.83 but does not surpass the depth-8/LR-0.03 result. Core loss prediction shows the same optimal configuration (test \(R^2 = 0.81\)). A comparison with tuned Random Forest and XGBoost confirms CatBoost’s advantage (best test \(R^2 = 0.84\) vs 0.79 and 0.74, respectively). The paper reports full grid results, iteration curves, train–test gap analysis, and practical tuning rules: start with depth 6–8 and LR 0.03; avoid depth 10 with LR 0.10 on small datasets; monitor the train–test gap (above 0.10 indicates overfitting); and use cross-validation for configuration selection. Limitations are stated explicitly: the data are synthetic, the training set is small (170 samples), and random splitting may overestimate performance. For real industrial data, grouped cross-validation by coil/heat and extrapolation tests on held-out composition grades are recommended. The work provides a reproducible hyperparameter-tuning template for data-driven texture engineering in electrical steels and a clear statement of the gap between synthetic demonstration and experimental validation.
Author: geruganti, sudhakar
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Keywords: CatBoost, hyperparameter optimization, learning rate, Goss texture, non-oriented electrical steel, core loss prediction, texture engineering, overfitting, annealing temperature, cold reduction, aluminum content, random forest comparison, XGBoost comparison, grid search, response surface, gradient boosting decision tree
Open Access status: green
OpenAlex ID: W7235880334
