Machine learning-driven prediction of phase transformation regions of TTT diagram for carbon and low alloy steels via high-throughput boosting methods

Time–temperature–transformation (TTT) diagrams are essential for describing isothermal phase transformation kinetics in carbon and low-alloy steels and for designing heat-treatment schedules. However, their experimental construction requires extensive metallographic characterization over wide temperature–time domains, limiting rapid compositional screening. In this study, ensemble boosting algorithms were systematically evaluated for predicting four critical phase boundaries: Pearlite Start (Ps), Pearlite End (Pe), Bainite Start (Bs), and Bainite End (Be). A dataset comprising 93 experimentally reported TTT diagrams was digitized and preprocessed using strict correlation filtering and normalization procedures. Independent regression models were developed with 10-fold cross-validation and evaluated using both conventional random splitting and a stricter steel-wise compositional split strategy to assess generalization across unseen steel compositions. Among the studied models, including AdaBoost, Gradient Boosting, LightGBM, XGBoost, and CatBoost, the CatBoost model exhibited the most consistent performance, achieving (R 2 ≥ 0.97 for pearlitic transformations (RMSE as low as 12 °C) and stable accuracy for bainitic regions (RMSE ≈ 22 °C). SHAP analysis identified transformation time, carbon, and nickel as dominant predictors, consistent with established transformation theory. The model was further validated using a steel-wise compositional split, and TTT curves for 15 independent (unseen) steel compositions were predicted. Validations against experimentally reported TTT diagrams demonstrate strong agreement between reconstructed and reported TTT curves, including accurate prediction of the transformation nose. These results demonstrate that boosting-based models provide an accurate and interpretable framework for rapid estimation of TTT diagrams, thereby helping to reduce experimental effort and supporting alloy and heat-treatment design