Fast and reliable validation of new designs in complex physical systems such as batteries is critical to accelerating technological innovation. However, battery development remains bottlenecked by the high time and energy costs required to evaluate the lifetime of new designs 1 , 2 . Notably, existing lifetime forecasting approaches require datasets containing battery lifetime labels for target designs to improve accuracy and cannot make reliable predictions before prototyping, thus limiting rapid feedback 3 , 4 . Here we introduce Discovery Learning, a scientific machine learning approach that integrates active learning 5 , physics-guided learning 6 and zero-shot learning 7 into a human-like reasoning loop, drawing inspiration from educational psychology. Discovery Learning can learn from historical battery designs and reduce the need for prototyping, thereby predicting the lifetime of new designs from minimal experiments. To test Discovery Learning, we present industrial-grade battery data comprising 123 large-format lithium-ion pouch cells, including diverse material–design combinations and cycling protocols. Trained on public datasets of cell designs different from ours, Discovery Learning achieves 7.2% test error in predicting cycle life using physical features from the first 50 cycles of 51% of cell prototypes. Under conservative assumptions, this results in savings of 98% in time and 95% in energy compared with conventional practices. Discovery Learning represents a key advance in accurate and efficient battery lifetime prediction and, more broadly, helps realize the promise of machine learning to accelerate scientific discovery 8 .
Zhang et al. (2026) studied this question.