AI Performance and Health Prognosis for Degradation-Aware Battery Development

Battery development requires extensive experimental trials ranging from benchtop materials discovery, to scale up trials, to full production optimization, incurring costs that may exceed hundreds of millions of dollars. One principal barrier to a more efficient and accelerated development process is the requirement to consider the degradation of batteries with use, the experimental evaluation of which requires 12 to 18 months of laboratory cycling experiments to reach failure. Physics-based simulations of battery degradation are rapidly advancing, but the diversity and complexity of coincident degradation mechanisms precludes the complete replacement of experimental trials. Data-driven methods, however, do not require exhaustive description of degradation mechanisms, and in recent years have shown strong performance in predicting battery degradation and remaining useful life (RUL). In this presentation, we share recent work in predicting battery RUL from limited cycling for a dataset of 300 Li-ion pouch cells representing 6 cathode chemistries and a variety of anode and electrolyte compositions. RUL prognosis was performed with a mean absolute error of ~100 cycles for a model trained on all available cathode chemistries, and useful predictions were made for unseen chemistries. Furthermore, we present timeseries prognosis of a multivariate battery state of performance and health including capacity, energy, efficiency, resistance, and open circuit voltage quantities.