Recent advances in synthesis methods have created an enormous amount of variety in the colloidal and nanoscale building blocks that are used to form new materials and devices. In contrast with atomic systems, where the basic structural units are drawn from a limited and well-understood periodic table, colloids and nanoparticles can vary in any number of ways, leading to very high-dimensional design spaces. Modern supercomputers can efficiently dispatch simulations of large regions of parameter space in parallel, but because creating order parameters for a priori unknown structures is infeasible, researchers often resort to the manual analysis of each sample — a process that places a bottleneck on scientific discovery. By using machine learning, we can drastically improve the analysis speed of these systems: unsupervised learning focuses the analysis effort of experts onto the distinct regions of identical structures in parameter space, while supervised learning enables high-throughput identification of samples. These methods can reduce the time required to analyze data sets by orders of magnitude, enabling much broader studies than were previously possible. These automated methods hold promise to expand computational materials design from a manual “gold-panning” approach — iteratively running many simulations and finding the interesting results — to a targeted, intelligent design process.
Back to Workshop IV: Synergies between Machine Learning and Physical Models