A universal graph deep learning interatomic potential for the periodic table

Interatomic potentials (IAPs), which describe the potential energy surface of atoms, are a fundamental input for atomistic simulations. However, existing IAPs either fitted to narrow chemistries or too inaccurate general applications. Here, we report universal IAP materials based on graph neural networks with three-body interactions (M3GNet). The M3GNet was trained massive database structural relaxations performed by Materials Project over past 10 years and has broad applications in relaxation, dynamic simulations property prediction across diverse chemical spaces. About 1.8 million were identified from screening 31 hypothetical crystal structures be potentially stable against crystals energies. Of top 2000 lowest energies above hull, 1578 verified using DFT calculations. These results demonstrate machine learning-accelerated pathway discovery synthesizable exceptional properties.