Abstract: Machine learning tools combined with theoretical simulations can effectively accelerate the design of novel materials. Data-driven approaches can access the information embedded in years of experiments, perform rapid optimization of high-dimensional experimental conditions and design parameters, increase the accuracy and speed of physics-based simulations, or design new molecules and crystals automatically. By deploying automated atomistic simulations (molecular dynamics, electronic structure) to create bottom-up representations of materials, and by using those as inputs to machine learning models, we can build effective and accurate predictors. Here, we will describe recent results and ongoing work in using machine learning as the connector between multiple scales of simulation and experiment in materials design. These include
- High-throughput screening of molecular materials such as organic light emitting-diodes or small molecule battery electrolytes using electronic structure simulations
- Inverse design tools based on deep generative models for automatic chemical discovery
- Automated learning of all-atom and coarse-grained potentials for discovery of soft materials like ion-conducting polymers
- Graph-based representations that accurately predict and rationalize polymorphism in nanoporous zeolite materials.