Artificial intelligence meets soft matter: Machine learning applications for the study of colloidal self-assembly

Abstract

Colloidal systems consist of microscopic particles — called colloids — that are dispersed in a solvent. Because of the continuous collisions with the much smaller particles in the solvent, colloids undergo Brownian motion and diffuse randomly through the solvent. As a consequence, they can spontaneously self-assemble into a wide variety of different phases, including fluids, glasses, crystals, and even quasicrystals. This exploration of phase space and self-assembly is effectively a colloidal analogue of the phase behaviour typically observed in atomic and molecular systems.

One of the fundamental tools for studying colloidal self-assembly is represented by computer simulations, which allow us to explore in detail the self-assembly process – from how specific interactions relate to the outcome of self-assembly, to the properties of the self-assembled materials. However, as the colloidal systems of interest become more complex, new tools are required to speed up computer simulations of such systems and analyse their properties. In this thesis, we develop new machine learning algorithms both to speed up simulation codes, as well as to help tackle a variety of open problems in the study of colloidal self-assembly. One common theme of this thesis is the development of structural order parameters that can be used to identify the products of self-assembly, reduce the numerical cost associated with the prediction of the properties and the behaviour (e.g., energy and dynamics) of colloidal systems, and even to inverse-design soft materials with a desired structure.