Machine-learning interatomic potentials: a story about how a Big Data approach compensates for our incomplete understanding of interatomic interaction

Machine learning, an approach to create models based on large amounts of data, is transforming many fields of research. This approach allows us to compensate for our incomplete understanding of a phenomenon by incorporating big data into a model. In my talk I will illustrate how this ideology works in the field of models of interatomic interaction.

Namely, in the first part of my presentation I will give a brief introduction to machine-learning potentials and present some of their success stories. The second part will be devoted to methodological foundations of machine-learning potentials and their successes. In particular, I will formulate the problem of construction of an interatomic potential as a model reduction problem with respect to a quantum-mechanical model (such as DFT). I will then discuss what modeling assumptions are made in classical potentials, identify which ones cause uncontrollable errors, and show how machine learning helps to lift some of those assumptions while still benefiting from physical knowledge. I will conclude by discussing the existing challenges related to those modeling assumptions that are difficult to lift or those cases in which common assumptions fail.