Prediction of glass-forming ability based on multi-model fusion

Abstract

At present, it is difficult to find metals with good glass-forming ability (GFA), seriously hindering the industrial application of metallic glasses (MGs). Therefore, it is urgent to develop a regression model that provides a prompt prediction for the critical diameter (Dmax) of alloys. Most established models rely heavily on elaborate feature descriptors and single algorithms, which may limit the predictive accuracy. In this work, the Stacking fusion is adopted to construct a more efficient model (Stack-Dmax). To avoid domain knowledge involved in feature engineering, only the alloy composition is treated as the input. Experiments revealed that compared to 6 individual models, the Stack-Dmax model is more robust and accurate regarding performance; it achieves improvement up to 41.61 %, 54.76 % and 35.67 % respectively in RMSE, R2 and MAE. The Stack-Dmax model is also better than the five most recently reported models based on RF, XGB, and CNN. In practical applications, our model exhibits potential for designing new La- and Mg-based MGs. These findings suggest that the Stacking approach provides an effective GFA-prediction model only based on the composition of alloys.