Rational design and glass-forming ability prediction of bulk metallic glasses via interpretable machine learning

The prediction accuracy of current mainstream machine learning (ML) models depends on regulating many hyperparameters. In this paper, a deep forest (DF) model with a few hyperparameters and a non-excessive dependence on super parameter regulation was applied to the prediction of glass-forming ability (GFA) of bulk metallic glasses (BMGs). Compared with these of the mainstream ML models, including Support Vector Regression (SVR), random forest (RF), gradient boosted decision trees (GBDT), k-nearest neighbor (KNN), and eXtreme gradient boosting (XGBoost), the tenfold cross-validation shows that the determination coefficient (R-2) of our suggested DF model is improved by 10.4%-74.2%. Moreover, the parameter U obtained by the SHapley Additive exPlanations (SHAP) method analysis can be used to guide the design and development of BMGs. Finally, a design and development of scheme process for BMGs that meets the expected requirements is given via parameter U and the constructed DF model.