Discovering Crystals using Shape Matching and Machine Learning
As the effective interactions between particles at the nano- to colloidal scale can be tailored and tuned in many ways, by using DNA and polymer functionalization, depletants, and solvents, there is a vast parameter space of design choices. Using simplified coarse-grained representations of the effective interactions between particles, computer experiments are an essential research tool for both exploring the vast space, and for investigating the first principles of self-assembly. The rate at which data can be amassed through computational simulation continues to accelerate, and thus the pace of discovery becomes limited not by the rate at which data can be generated, but can be analyzed.
We show how new crystalline structures can be identified automatically from analysis of large data set. By deploying a hierarchy of pattern analysis techniques using shape matching and machine learning algorithms, local structures are extracted, classified, and then used to partition a data set into a phase diagram of similar crystals.
This method requires no a-priori knowledge of what might be present in the data set. We apply this method to two data sets generated from a parameter sweep of a particles that interact via a model double-well pair potential, the Lennard Jones Gauss potential. The data set contains both simple and complex crystals, including quasicrystals. We show how phase diagrams can be automatically generated, searches for new structures focused, and how the discovery of new types of materials can be accelerated.
 Phillips, Voth, "Discovering crystals using shape matching and machine learning," Preprint, 2013