Overcoming the challenge of the data imbalance for prediction of the glass forming ability in bulk metallic glasses

Machine learning (ML) has been extensively studied in predicting the glass-forming ability of bulk metallic glasses (BMGs). Based on the current state of development of BMGs, the reported critical casting diameter (Dmax) data show an imbalance. Nevertheless, almost most of the current literature using ML to predicted Dmax has failed to consider this phenomenon, resulting in generally low prediction accuracy. Only a very small amount of literature deals with this issue only at the data level. In this work, an improved deep neural network (IDNN) model based on the DenseLoss function was proposed at the algorithmic level. The IDNN model has better generalization capability than the currently reported models by obtaining the highest R2 score of 0.841 in the test set. Our work highlights the importance of dealing with the Dmax imbalance problem to address the low accuracy of current ML models in predicting Dmax. This research work will have a significant impact on the discovery of novel BMGs.

Automated summary

This study proposes an improved deep neural network (IDNN) model based on the DenseLoss function to predict the glass-forming ability of bulk metallic glasses (BMGs). The IDNN model has better generalization capability and achieves the highest R2 score of 0.841 in the test set. This research highlights the importance of addressing the Dmax imbalance problem for improving the accuracy of ML models in predicting Dmax and has significant impact on the discovery of novel BMGs.