Machine learning interatomic potentials for atomistic modeling of materials

Coordinators : Simona Ispas and Thibault Sohier)

Matter, whether amorphous or crystalline, is made up of atoms interacting with each other. Atomistic simulations have contributed to the comprehension of the intricate relationships between the underlying atomic structure and physical properties of various systems (e.q. sound velocity, heat and electronic conductivities as well as many spectroscopic quantities).

Full quantum-mechanical descriptions are often quite accurate, but they quickly become very computationally costly as the number of atoms grows. Empirical descriptions using interatomic potentials allow very fast simulations, but they lack accuracy and need to be parametrized carefully.

In this context, machine learning offers an interesting compromise. Interatomic potentials can be learned from a training set of accessible quantum-mechanical simulations on relatively small systems, then applied to much larger problems. We are working on this approach to simulate various properties of both disordered, amorphous systems and complex crystal made up of different 2D materials (van der Waals heterostructures).