Lattice inversion potential with neural network corrections for metallic systems

Abstract

By combining the lattice inversion method and the back-propagation neural network (BPNN), we have developed a neural network-based lattice inversion potential (NN-LIP). The lattice inversion method is used to describe the pairwise interaction, and the BPNN is used to describe the many-body correction term. NN-LIP has been applied successfully to six representative noble metal systems, including Au, Ag, Pd, Pt, Ir, and Rh. The results show that NN-LIP can reproduce the results calculated by first-principles accurately, including binding energy, lattice constant, elastic constant and a series of energy curves of different lattice structured metals, greatly expanding the applicability of lattice inversion potentials. Furthermore, compared to pure machine learning potential, NN-LIP exhibits better robustness and generalization in regions not covered by the dataset used for training, which is due to the use of pairwise potential as the skeleton. NN-LIP provides a new theoretical framework for constructing high-precision potentials.