A two-step data augmentation method based on generative adversarial network for hardness prediction of high entropy alloy

Abstract

The machine learning (ML) has been widely applied in materials science research and has made a lot of contributions. However, the performance of ML model is limited by the amount of material data, and the data for some material fields are still lacking. In order to deal with data shortage for hardness prediction of high entropy alloy (HEAs), based on generative adversarial network (GAN), a data augmentation method was proposed, which is named two-step method. A thoroughly comparative analysis was carried out by applying two-step method to the hardness prediction of HEAs with 205 data and the formation energy prediction of photocatalyst with 3099 data. The Diebold-Mariano test is also adopted to compare the model performances from the statistical perspective. Results verify that two-step method is significantly superior to the previous GAN method and the newly constructed ML model shows a satisfied data augmentation performance. In addition, for photocatalyst in cases with different degrees of data shortage, it is indicated that the less the data quantity, the more significant the improvement effect, and the error is 6.1% lower than that given by the previous work even though in the case with total 3099 data. Additional 7 HEAs data were further gathered to evaluate the generalization ability of ML models, and the present model shows the excellent predictive performance again. Finally, three types of interpretability method were utilized to obtain insight into the relationship between features and label as well as the reason for two-step method to elevate ML model performance in HEAs hardness prediction.