I will report on the recent developments in this rapidly advancing field of machine learned interatomic potentials. While in the 2010s kernel based methods and the equivalent shallow neural networks have already changed how large scale simulation is done, new exciting developments include linear potentials such as MTP and ACE, as well as equivariant message passing networks. We now understand all these models as part of a larger, unified design space, and this new insight has lead to the development of model architectures such as MACE, which has unprecedented accuracy and extrapolation power.
Back to Workshop II: Scale-Bridging Materials Modeling at Extreme Computational Scales
Copyright. All Rights Reserved.