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

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning Applications

Background:

• Machine learning (ML) models complex phenomena in magnesium (Mg) alloys, but often face hyperparameter optimization and interpretability challenges.
• These limitations reduce predictive accuracy and impede mechanistic understanding of material behavior.

Purpose of the Study:

• To develop an enhanced interpretable ML framework for Mg alloys.
• To improve predictive accuracy and gain mechanistic insights into twinning phenomena.

Main Methods:

• Integrated Optuna for automated hyperparameter tuning using tree-structured Parzen estimators.
• Employed SHapley Additive exPlanations (SHAP) for quantitative feature attribution.
• Utilized in situ EBSD tensile tests and molecular dynamics (MD) simulations on Mg alloys.

Main Results:

• Achieved significant performance improvements: F1-score gains of 6.33-11.84% (dataset T) and AUC increases up to 16.31% (dataset Y).
• Discovered a novel grain shape-orientation effect: specific elongated grain orientations (20-80°) facilitate twinning nucleation, while others (0-20°, 80-90°) suppress it.
• Linked this effect to boundary-segment orientations influencing local constraints and stress transfer.

Conclusions:

• The enhanced interpretable ML framework significantly improves predictive performance for Mg alloy twinning.
• The identified grain shape-orientation effect provides new mechanistic understanding of twinning nucleation in Mg alloys.
• This approach offers a powerful tool for materials modeling and discovery.