Gerardo Campos-Villalobos (Utrecht University) Machine Learning Potentials for Colloidal Systems

In recent years, we have witnessed enormous advances in machine learning (ML) or data- driven approaches for constructing accurate and efficient potentials for computer simulations. These approaches have been mainly applied to atomistic simulations. Nowadays, colloidal systems represent a class of materials consisting of mesoscopic particles of different compositions, functionalities and shapes. Such nanoparticles can self-assemble in a range of different two- and three-dimensional superstructures exhibiting an incredibly large surface-to-volume ratio, which make them perfectly suited for many technological applications. Atomistic simulations of such systems are severely limited by the length- and time-scales that can be achieved with present-day computers. Hence, computational studies of their phase- and self-assembling behaviour relies heavily on the use of coarse-grained (CG) models based on effective CG interactions. In this talk, we introduce a multi-scale ML approach to construct accurate and computationally-efficient CG many-body interaction potentials for complex colloidal systems composed of particles of arbitrary shape and with highly anisotropic interactions. In our proposed approach, either the CG forces or many-body effective interactions, which are extracted from reference fine-grained simulations, are represented by linear models constructed in terms of structural descriptors of local particle environments. We show that this simple yet accurate coarse-graining framework may enable the characterisation, understanding, and prediction of the structure and phase-behaviour of relevant soft-matter systems by direct and efficient simulations.