Improving machine learning based phase and hardness prediction of high-entropy alloys by using Gaussian noise augmented data

Developing a machine learning (ML) based high-entropy alloys (HEA) prediction model is an advanced method to improve the traditional trial-and-error experiments with a long period and high cost. However, the ML model is highly dependent on data. This paper draws on the experience of the image data augmentation approaches, and proposes a simple material data augmentation method, which aims to solve the difficulties of small samples and large noise of material data, by adding Gaussian noise to the original data to generate more “pseudo samples”. This work respectively starts by the HEA phase classification task and the hardness regression task, to study the effect of noise on data enhancement. It is found that the noise samples are different samples with new information. The noise samples enhanced data can significantly improve the test results of the models. Further testing results with the validation set that the models have never seen before, demonstrate that the regression model of medium noise samples enhanced has the best prediction accuracy (R2 = 0.954). It turns out that the data enhancement method applied in this work helps the ML models to achieve a more efficient and accurate prediction of HEA phase and hardness. Besides, this work provides a reference method for improving the ML models and designing new HEA.