Performance-Boosted Interpretable ML via Optuna-SHAP: Uncovering Orientation-Driven Twinning in Mg Alloys

Machine learning (ML) is highly effective for modeling the complex factors governing twinning in magnesium (Mg) alloys, but it is often limited by challenges in hyperparameter optimization and a lack of interpretability, which reduce predictive accuracy and hinder mechanistic understanding. In this work, we present an enhanced interpretable ML framework that integrates Optuna for automated hyperparameter tuning using tree-structured Parzen estimators and SHapley Additive exPlanations (SHAP) for quantitative feature attribution. This approach delivers significant performance improvements, including F1-score gains of 6.33-11.84% on dataset T and AUC increases of up to 16.31% on dataset Y, outperforming previous benchmarks. When applied to a custom dataset derived from in situ EBSD tensile tests on Mg alloys and complemented by molecular dynamics (MD) simulations, SHAP analysis reveals a previously unrecognized grain shape-orientation effect: elongated grains with long-axis orientations of 20-80° relative to the tensile direction facilitate twinning nucleation, whereas orientations of 0-20° or 80-90° suppress it. Combined EBSD observations and MD simulations indicate that this effect arises from changes in boundary-segment orientation combinations, which regulate local constraint conditions, stress-transfer paths, and effective boundary resistance.