A universal machine learned potential: from molecules to materials

Physics based simulations, such as ab-initio electronic structure calculations, can act in both a validatory and
predictive capacity but their impact is often limited by the resources required. Currently 40% of ARCHER2 is
consumed by ab-initio explicit electronic structure simulations. Machine learned potentials are
revolutionizing the way such atomic scale modelling is done, allowing the accurate simulation of materials at
length scales, time scales and throughputs that are many orders of magnitude larger than what was possible
just a decade ago. The ultimate goal of these techniques is to be applicable and out-of-the-box for all systems
from molecules to materials and able to treat charge self-consistency effects in a unified way regardless of
elemental composition. In order to balance broadness of applicability with performance on specified
problems, one promising direction is to create so-called foundation models which are fitted to a very wide
ranging database and are subsequently specialised to any given project with little additional effort. This
project will investigate the promise of this approach, and take as a starting point the recently published
"mace-mp-0" model developed in the research group of Prof. Csanyi. The project will explore novel
architectures, data fusion approaches, and transfer-learning techniques. The resulting potentials will be
validated on a variety of impactful exemplar applications relevant to IBM's activities.