Machine learning is quickly revolutionizing materials discovery and design. In soft matter, machine learning can discover hidden correlations linking building block attributes to material structure, autonomously identify and classify from millions of state points dozens of polymorphs comprising a phase diagram, find “sweet spots” in the data where complexity is likely to arise, find new crystal structures with target properties, minimize the number of simulations or experiments needed to map phase boundaries, and more. Here we give examples of each of these using random decision forests, active learning with Gaussian process regression, clustering, and artificial neural networks. For example, using unsupervised Gaussian mixture models (GMMs), supervised artificial neural networks (ANNs) and local environment descriptors, we identify and classify complex polymorphs arising from two-parameter interaction potentials. We combine recently developed techniques in inverse design called “digital alchemy” with random forest techniques to construct a model that correctly classifies the colloidal crystals of more than ten thousand polyhedral shapes into 13 different crystal structures with a predictive accuracy of 96% using only two geometric shape measures. We address the problem of efficient phase diagram sampling using active learning techniques and achieve an 80% reduction in the number of state points needed to establish the phase boundary up to a given precision in example applications. We also show how batch sampling can be used to better utilize parallel computing resources in active learning. Finally, we show how our data management framework signac supports our data-intensive research.

Machine learning for colloidal crystal structure and property discovery & inverse design

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.