Machine Learning for Automatic Structure Analysis and Experimental Design

Abstract

Self-assembly is a ubiquitous process that holds great promise for the design and engineering of new materials and systems. While chemistry today is largely based on combining a few building blocks into molecules with desirable properties, recent advances in colloidal and nanoscale self-assembly have allowed us to move beyond the periodic table of elements to design building blocks with attributes tailored to their desired applications. With this great flexibility, however, comes a cost: both experiments and models of these systems are often laden with many tunable parameters, frustrating analysis and engineering efforts. Furthermore, it is often not known whether only a subset of these parameters is important or if observed behaviors depend on the confluence of multiple variables. In this work, we show examples of complex design spaces for colloidal and nanoscale self-assembly—including systems of far-from-equilibrium, anisotropic particles. We further show how machine learning can be applied to two major problems involved in studying self-assembly in silico: analyzing three-dimensional structure and engineering building blocks with many design variables.