Research on glass-forming ability based on transformer and tabular data transformation

Abstract

Most prior machine learning studies on glass forming ability (GFA) rely on characteristic temperature of metallic glasses (MGs). Before modeling, it is necessary to prepare MGs samples by traditional trial-and-error method, which is both time-consuming and resource-intensive. Consequently, it is necessary to design a method to predict GFA quickly and accurately before the experiment. In this paper, 111 element features are calculated as input data based on MGs composition. ADASYN, GN, and SMOGN oversampling methods are used to alleviate data shortage and skew. Investigating the performance of two prevalent deep learning approaches, transformer and tabular data transformation, in GFA research. Results reveal that oversampling technique enhances performance of each model by improving data distribution. The performance of FT-Transformer and CNN based on data transformation is significantly better than GBDT. Subsequently, the ensemble method is utilized to integrate the predictions from base models, achieving an R² score of 0.870 for ensemble model. Finally, two systems of Zr-Al-Ni and Cu-Ce-Ga are predicted, and the reliability of ensemble model is verified by experimental data. This research can estimate GFA rapidly and correctly based on MGs composition, and it is expected to guide the development of new MGs.