Structure Prediction and Discovery of Molecular Crystals with Enhanced Electronic Properties*

Molecular crystals often exhibit polymorphism, the crystallization of the same molecule in several different structures. Crystal structure may profoundly influence the physical and chemical properties. We combine first principles simulations with machine learning (ML) and optimization algorithms to predict the structure of molecular crystals and discover molecular crystals with enhanced electronic and optical properties. Our crystal structure prediction (CSP) workflow begins by estimating the unit cell volume using a machine learned model based on the volume enclosed by the molecule's packing-accessible surface and molecular topological fragments, which capture the bonding environments of the atoms in the molecule and the types of inter-molecular interactions the molecules may form. Next, the random structure generator, Genarris, performs preliminary configuration space screening by random sampling with physical constraints. Genarris generates structures with a distribution around the target volume in all space groups compatible with the molecular point group symmetry and the requested number of molecules per unit cell, including space groups with molecules occupying special Wyckoff positions. The "raw" pool is clustered and down-selected based on considerations of diversity and energy to form an initial populaiton for the GAtor genetic algorithm (GA). A GA uses the evolutionary principle of survival of the fittest to perform global optimization. GAtor's special features are: A massive parallelization scheme enables effective utilization of high performance computing resources; A variety of breeding operators (crossover and mutation) tailored for molecular crystals provide a balance between exploration and exploitation; Evolutionary niching, performed by using ML for dynamic clustering and then using a cluster-based fitness function, helps overcome initial pool biases and selection biases by steering the GA to under-explored regions of the configuration space; Property-based fitness functions enable inverse design of crystal structures with target properties. In this talk, our CSP workflow will be demonstrated for select cases.