Machine Learning Regression Algorithm Predicts Multi-component Crystal Configuration Energy

Some machine learning algorithm tools, such as neural networks and Gaussian process regression, are increasingly being applied to the exploration of materials. Here, we have developed a form to use this nonlinear interpolation tool to describe properties that depend on the degrees of freedom in multi-component solids. A symmetrically adapted clustering function is used to distinguish different atomic order degrees. These features are used as the input of neural networks, Gaussian process regression and other algorithmic models, and some inherent properties of materials, such as formation energy, can be reproduced by the trained machine algorithm model. We use this technique to reproduce the expansion Hamiltonian of a synthetic cluster with multi-body interaction, and calculate the formation energy of ZrO based on first principles. The form proposed in this paper and the results shown that complex multi-body interactions can be approximated by nonlinear models involving smaller clusters. The training models used in this paper to predict energy include neural networks, Gaussian process regression, random forests, and support vectors regression, using MSE and coefficient of determination to evaluate the prediction results, and adding genetic algorithms in the feature selection process can remove some redundant features and improve the prediction efficiency and accuracy. The results show that the neural network is the best algorithm model which selected in this article, the prediction effect of support vector regression is relatively inferior.