Deep learning-based phase prediction of high-entropy alloys: Optimization, generation, and explanation

Identifying phase information of high-entropy alloys (HEAs) can be helpful as it provides useful information such as anticipated mechanical properties. Recently, machine learning methods are attracting interest to predict phases of HEAs, which could reduce the effort for designing new HEAs. As research direction is in its infancy, there is still plenty of room to develop machine learning models to improve the prediction accuracy and further guide the design of HEAs. In this work, we employ deep learning-based methods regarding optimization, generation, and explanation, for enhancing the performance and identifying key design parameters for phase prediction of HEAs. We first establish regularized deep neural networks for predicting HEA phases and optimize hyper-parameters concerning model architecture, training, and regularization. To overcome data shortage of HEAs, we then focus on developing conditional generative adversarial network for generating additional HEA samples. We observe the augmentation from our generative model significantly improves model performance, achieving prediction accuracy of 93.17%. Lastly, we concentrate on understanding contributions of design parameters to identifying solid solution (SS) phase as an example. Our work delivers guidance not only for developing a reliable deep learning-based phase prediction model, but for explaining significant design parameters to assist design of novel HEAs.