Modelling Particle Systems with Many-body Equivariant Networks

Data Science Seminar

Christoph Ortner
UBC

Abstract

The integration of machine learning (ML) into the traditional modeling workflows is replacing decades-old ad hoc approximations (e.g., in constitutive laws) leading to new models that far outstrip their predecessors in accuracy and transferability. "Pure" ML approaches are rarely successful (so far) but remarkable results can be achieved when integrated with domain knowledge. My talk will focus on scientific machine learning for modelling particle systems, where "prior knowledge" such as locality of interaction and symmetries play key roles. I will explain the ACE (and MACE) framework for constructing many-body equivariant {neural, tensor} network architectures, which is emerging as a general platform to successfully address those challenges. The framework can be applied in a wide range of application areas, including e.g. machine learning interatomic potentials, coarse-grained molecular dynamics, reduced-order electronic structure methods, to jet-tagging or parameterizing many-body wave functions. I plan to cover some examples of theoretical results about (M)ACE as well as a selection of those examples.