Uncertain lithium-ion cathode kinetic decomposition modeling via Bayesian chemical reaction neural networks

Abstract

Lithium-ion batteries are the focus of significant recent research interest due to their use in energy storage systems, electric vehicles, and other green technologies. Under various abuse conditions, these batteries undergo thermal runaway, which can lead to rapid temperature rise, flammable gas release, and fires. Current simulation approaches for thermal runaway at the full-cell scale involve utilizing kinetic models fit to experimental data from individual battery components. Despite substantial experimental uncertainty in the literature for these component experiments, prevalent models have not accounted for such uncertainty or noise. Here, we introduce uncertainty quantification for lithium-ion battery cathode thermal decomposition modeling via a novel Bayesian inference methodology. This approach leverages Chemical Reaction Neural Networks and particle-based uncertainty quantification to infer uncertain kinetic parameters while eliminating traditionally used simplifications, allowing for improvements to model accuracy and broader consideration of correlated parameters. We validated this new framework by learning an uncertain decomposition model for NCM333 (nickel-cobalt-manganese) cathode materials using differential scanning calorimetry (DSC) measurements with added synthetic noise. Then, we quantified the uncertainty in NCM811 cathode thermal runaway chemistry using experimental DSC measurements from various sources in the literature. Our methodology’s ability to account for correlated Arrhenius parameters led to much broader uncertain parameter ranges and thus more generalizability at higher temperatures. We additionally found that the NCM811 model distribution learned directly from experimental data in the literature has 4σ onset temperature ranges up to 20 °C wide and specific reaction enthalpy ranges accounting for upwards of 80% of the mean value, carrying significant implications for downstream applications. Our work bridges the gap between noisy or uncertain experimental data and practical cell-scale simulations, thereby facilitating more realistic and robust thermal runaway models that support enhanced battery safety and performance optimization.