Methods for Feature and Image Selection for Machine Learning Potentials: A Comparative Study

Machine learning interatomic potentials (MLIPs) have revolutionized atomistic simulations, enabling predictions of materials properties with first-principles precision while only taking a small fraction of DFT time cost. The reliability and transferability of MLIPs strongly depend on the quality of features and structures used to represent the underlying <i>ab initio</i> potential energy surfaces. To reduce the computational overload of expensive first-principles calculations, there is an urgent need to select informative image and features, in particular for complex materials such as high-entropy materials. Current methods for image and feature selection use linear correlations in the feature space (and output space) spanned by the training data. However, the relationship between features and properties can be highly non-linear, rendering it questionable for linear methods. Here, we propose an alternative method for image and feature selections. This approach uses auto-encoder to perform non-linear dimensional reduction. Images selection will then be conducted in the low-dimensional latent space, and the following feature selection will be realized through sensitivity analysis. We benchmark the method on first-principles data of AlCrFeNiCo high-entropy alloys. This class of high-entropy alloys is crucial in many applications as it is an emerging material with superior mechanical properties and corrosion resistance. We compare the new feature and image selection method with existing ones using atom-centered neural network potentials as models and Gaussian symmetry functions as local chemical environment descriptors.