The rapid pace of technological innovation often outstrips the rate at which materials scientists can develop better materials to accommodate these new technologies. Better designs for electric cars or aircraft demand better batteries. Challenges posed by increasingly unmanageable plastic waste streams demand development of biodegradable materials. Scientists in DSL are working to build AI software that will learn how to run simulations of materials on supercomputers to quickly identify improved materials.
Our main challenge is that the performance of even next-generation supercomputers is overwhelmed by the massive number of possible materials. For instance, evaluating the effectiveness of trillions of possible molecular materials for use in batteries would take years on a supercomputer. The AI tools we are developing will “learn” which materials represent the best use of our simulation capabilities.
Our current work focuses on using HPC to quickly train machine learning models for a variety of properties given different information about molecules. For example, we have developed convolution neural networks to predict molecular stability with a high accuracy from inexpensive, low-accuracy methods. We also are building tools that augment human creativity by generating ideas for new electrolyte molecules. Our goal is to test these ideas with our other machine learning prediction models to quickly identify new materials for better batteries.