Battery research provides a clear use case. As cells move from powering personal electronics to powering electric vehicles and grid-scale storage, the precision and precision with which they can predict cycle life becomes a key factor shaping cost, reliability, and fidelity to long-term capital planning. Integrated predictive models have rapidly advanced to reduce the burden of experimentation in making cycle life predictions. However, many approaches still rely on full lifetime measurements of the same or similar batteries, and thus struggle to reliably generalize to factors such as unfamiliar chemistry and operating conditions.
Today, Jiawei Zhang and colleagues introduce discovery learning, a framework inspired by Jerome Bruner’s discovery learning theory in educational psychology. By combining classical machine learning techniques with simulation-based inference using a pseudo-two-dimensional electrochemical model, this framework enables out-of-distribution inference of the cycle life of large-format pouch cells using historical data obtained from small cylindrical cells augmented by initial cycle testing on a subset of possible prototype groups identified by the framework as most constraining. In particular, the framework’s final predictions are generated by a Gaussian process regression model trained on pseudo-labels predicted from the framework itself rather than actual prototype cycle life data, achieving an average absolute percentage error of 7.2% across the entire set of prototypes.
