Efficient selection of linearly independent atomic features for accurate machine learning potentials

Machine learning potentials are promising in atomistic simulations due to their comparable accuracy to first-principles theory but much lower computational cost. However, the reliability, speed, and transferability of atomistic machine learning potentials depend strongly on the way atomic configurations are represented. A wise choice of descriptors used as input for the machine learning program is the key for a successful machine learning representation. Here we develop a simple and efficient strategy to automatically select an optimal set of linearly-independent atomic features out of a large pool of candidates, based on the correlations that are intrinsic to the training data. Through applications to the construction of embedded atom neural network potentials for several benchmark molecules with less redundant linearly-independent embedded density descriptors, we demonstrate the efficiency and accuracy of this new strategy. The proposed algorithm can greatly simplify the initial selection of atomic features and vastly improve the performance of the atomistic machine learning potentials.