Machine Learning Interatomic Potentials: Keys to First-Principles Multiscale Modeling

Machine learning interatomic potentials (MLIPs) provide exceptional opportunities to accurately simulate atomistic systems and/or accelerate the evaluation of diverse physical properties. MLIPs moreover offer extraordinary capabilities to conduct first-principles multiscale modeling, enabling the modeling of nanostructured materials at continuum level, with quantum mechanics level of accuracy and affordable computational costs. In this chapter, we first briefly discuss conventional methods and MLIPs for studying the atomic interactions. Next, the basic concept of MLIPs, their training procedure, and technical challenges will be discussed. Later, with several examples, the bottlenecks of quantum mechanics and empirical interatomic potentials in the evaluation of materials and structural properties will be highlighted, and it will be shown that how MLIPs could efficiently address those issues. Last, the novel concept of MLIP-enabled first-principles multiscale modeling will be elaborated, and the practical prospect for the autonomous materials and structural design will be outlined.