Machine Learning-Based Prediction of Stacking Fault Energy in High-Manganese Steels: A Comparative Study of Ensemble...

Accurate prediction of the stacking fault energy (SFE) is critical for controlling deformation mechanisms, specifically transformation-induced plasticity (TRIP) and twinning-induced plasticity (TWIP), in high-manganese (high-Mn) austenitic steels, which are of growing importance in automotive and structural applications that demand exceptional strength-ductility combinations. This study presents a systematic comparative evaluation of six supervised machine learning (ML) models-Multiple Linear Regression (MLR), Random Forest (RF), Extra Trees (ETs), Gradient Boosting (GB), Support Vector Regression (SVR), and a stacking ensemble-trained on a curated, outlier-cleaned experimental database of Fe-Mn-C-Si-Al-Cr-Ni-N spanning SFE values from 5.0 to 63.0 mJ/m² (mean 23.7 ± 11.2 mJ/m²). After Z-score outlier removal (|Z| > 3) and 80/20 train-test splitting with nested 5-fold cross-validation hyperparameter optimization using GridSearchCV, ET and GB achieved training R² values of 0.988 and 0.990, respectively, confirming that SFE is highly predictable from alloy composition alone. The stacking ensemble delivered the best generalization on the independent held-out test set (test R² = 0.603, RMSE = 5.60 mJ/m², MAE = 4.86 mJ/m²), outperforming all the individual learners. Random Forest feature importance analysis identified Al (22.3%), Fe (20.5%), and Mn (17.7%) as the three most influential compositional variables, collectively explaining 60.6% of the predicted variance. Pearson correlation analysis confirmed that Al was the strongest individual linear predictor (r = +0.421, p < 0.001), whereas Fe showed a significant negative correlation (r = -0.327, p < 0.001). Mn, C, and the remaining elements showed no statistically significant linear correlations with SFE, underscoring the dominance of nonlinear compositional interactions. Composition-SFE design maps derived from the GB model delineate the TRIP/TWIP regime boundaries in the Mn-C and Mn-Al composition spaces, providing a validated computational tool for targeted high-Mn steel alloy design.