Identifying key features for predicting glass-forming ability of bulk metallic glasses via interpretable machine learning

Bulk metallic glasses (BMGs) have been receiving extensive attention in the community of physics and materials science due to their attractive properties. The traditional trial-and-error approach is inefficient in designing good BMGs, then it is imperative to elaborate a prediction scheme to accelerate the development of BGMs. This work combines a composite feature selection (c-FS) and machine learning (ML) algorithm to construct a more effective strategy for predicting the critical diameter (D_max) of BMGs. Apart from the characteristic temperatures, parameters involving domain knowledge such as thermodynamics, dynamics, and topology are also considered as the candidate features. A c-FS was devised to optimize the features, and a most efficient model was developed using the optimal subset. Our model holds the highest R² score (0.807) compared to the 29 criteria reported. In addition, the SHapley Additive exPlanation (SHAP) approach was introduced to boost the interpretability of the model. Benefiting from our model and SHAP approach, a formation rule for large-size BMGs is uncovered. The efficacy of this model provides new guidance for the design of BMGs.