Representation learning for data-driven analysis of soft matter simulations

The constantly expanding universe of synthetic chemistry has driven a sustained effort in molecular simulation to understand and leverage these approaches to produce new functional materials. Molecular simulation plays a key role in facilitating the understanding of new systems and identifying promising areas of the vast design space in which to expend experimental resources. In this context, quantitative descriptors of ordering within the simulated system become critical to both our physical understanding of the materials and the precise control over those systems through rational and data-driven design. In this talk, I will present my work in unsupervised machine learning to characterize local and global ordering in soft matter systems. The methods rely on the automated discovery of a low-dimensional latent space from the collection of observed environments to provide relevant order parameters across a wide range of systems.

I demonstrate the methods on colloidal crystallization, ice crystals, binary mesophases, and copolymer aggregates to illustrate its broad applicability. I also show that the spatiotemporal evolution of systems in the learned latent space is smooth and continuous, despite being derived from isolated snapshots rather than dynamic trajectories. In each case, the learned collective variables can give insight into the physical nature of the system at hand, without extensive parameter tuning or development of new functional forms. Finally, the learned collective variables are used in a supervised learning context to predict the relation between design variables and self-assembled structure. A systematic analysis of surrogate model architecture is considered and the merits of each are explored. These predictive models are then used to successfully select candidates that yield targeted structure. I will conclude with new developments and future work in the use of unsupervised and self-supervised learning for soft matter design.