From machine-learning interatomic potentials to materials chemistry
Seminar presented by Dr. Volker Deringer (University of Cambridge, UK)
Theory and atomistic computer simulations have become cornerstones of today's materials-science research, but they are still fundamentally limited by the huge computational cost of the underlying quantum-mechanical methods. This is especially critical for surface simulations or amorphous materials, which require large system sizes to be studied with high accuracy. Recent work in the community suggests that machine learning may provide a way out: performing a high-dimensional fit to the ab initio potential-energy surface, one can "teach" interatomic potentials that reach comparable accuracy but are orders of magnitude faster. The Gaussian approximation potential (GAP) approach used here is one of many emerging techniques.
In this presentation, I will highlight three directions we have recently explored to connect GAP-based simulations with materials chemistry. First, we have shown that surface stabilities and reactivity ("graphitisation") of amorphous carbon films can be modelled by a GAP, opening the door for more detailed investigations into the material's complex surface chemistry. Second, we have shown that GAPs can be used for random structure searching, suggesting their usefulness for crystal-structure prediction. Finally, ongoing work is concerned with the modelling of disordered, carbonaceous materials for energy applications. We expect that GAP and other machine-learning approaches will find further use in exploring the atomistic structure and chemical reactivity of complex materials.
Dr. Volker Deringer, Leverhulme Early Career Fellow (c) University of Cambridge