Materials Modelling Across the Scales

The past two decades have seen a transformative change in atomistic modelling with the development of machine-learned interatomic potentials, which allow quantum-accurate simulations at an affordable computational cost. While the formalism of these models has converged, there still remain open questions about the optimal way to generate training databases as well as about the reliability of potentials. In this talk, I will present our efforts to automatically generate atomic databases using a combination of active learning and advanced sampling methods and how the resulting potential results in exceptionally accurate potential energy surface for Mg at a pressure range of 0-600 GPa. Relatedly, I will report how transfer learning may be used to fine-tune foundation models using a little amount of data, resulting in accurate, but application-specific potentials.