Accelerating Battery Research, Development and Deployment Using Machine Learning

Batteries are complex multi-physics systems, with weak observabilty into their underlying physical processes. Physics models have made great strides to answer engineering questions and accelerate scale-up of batteries for 100-kWh-scale electric vehicles and MWh-scale grid energy storage. Yet many challenges still exist. First, with 50+ parameters, electrochemical models of conventional Li-ion batteries are still cumbersome to identify, requiring 3+ months of experiments - a pace too slow to keep up with new product development. Developers often compromise with empirical models that do not directly predict operating limits. Second, complex chemo-mechanical-coupled physics are not fully understood or modeled accurately enough. This slows development of next generation emerging electrode/electrolytes and impedes lifetime prediction/optimization for conventional Li-ion. Semi-empirical reduced-order models predict lifetime, but their supporting aging characterization experiments take 9+ months. Third, battery data comes in many forms, ranging from microscopy/spectroscopy to electrochemical. Streamlined approaches are needed to interpret all of these disparate data sources, distill them, and guide research, design, and deployment decisions.<br/><br/>This talk presents recent progress in application of machine learning to electrochemical and microscopy data, as well as storing that data and supporting models in accessible, searchable formats. Super-resolution generative adversarial networks quantify cracking and damage processes from scanning electron microscopy images of polycrystalline cathodes. Machine learning algorithms interpret electrochemical data to indentify physics-based electrochemical models and reduced-order lifetime models. A new battery data hub stores battery R&D data to better coordinate the efforts of researchers and provide uniform access to data and analysis tools.