Treffer: Prediction of supercooled liquid region of Fe-based metallic glasses by deep learning: Prediction of supercooled liquid region...

In this study, the supercooled liquid region (ΔTx) of Fe-based metallic glasses was predicted using a convolutional neural network (CNN), which is one of the popular deep-learning models. ΔTx is defined as Tx-Tg, where Tx is the onset crystallization temperature and Tg is the glass temperature. ΔTx can be considered a characteristic indicator of thermally stable bulk metallic glass. CNN-guided supervised learning was performed using only the compositional features without structural information and physics-informed features. The chemical formula in the material dataset was used to generate 140 features with the Python libraries 'Pymatgen' and 'Matminer'. The regression performance to predict ΔTx was evaluated using the determination coefficient (R2 score) and root-mean-squared errors (RMSE). To establish the CNN network, various reconstructed shapes of 140 compositional features were employed as the input vectors, such as 10 × 14, 14 × 10, 7 × 20, and 20 × 7. Among them, the 14 × 10 compositional feature vectors exhibited the best prediction performance for ΔTx with an R2 score of 0.924 and RMSE of 7.7 K. As a case study, the trained CNN model was applied to predict ΔTx of recently developed Fe-based metallic glasses, corresponding to the unseen dataset. The predictions for the unseen dataset were found to be consistent with the reported experimental results for ΔTx, indicating that the composition-based CNN model opens a new path to searching for promising Fe-based metallic glasses with high glass-forming ability. [ABSTRACT FROM AUTHOR]