Chemical Reaction Neural Networks for Fitting Accelerating Rate Calorimetry Data

TL;DR

This work addresses the challenge of accurately modeling thermal runaway kinetics in lithium-ion batteries by learning Arrhenius-type reaction parameters from ARC data using Chemical Reaction Neural Networks (CRNNs). By replacing linearized fitting with gradient-based optimization of stage-specific CRNNs, the authors achieve superior fits for both two-stage and four-stage multi-step kinetics, and demonstrate the impact by integrating the learned parameters into 3D ARC and oven simulations. The results show that CRNN-derived parameters better capture temperature trajectories, heat release, and runaway timing, improving predictive fidelity and enabling scalable, physics-informed safety analyses. The approach offers a flexible, automated pathway to extract kinetic information across cell chemistries and operating conditions, supporting safer battery design through simulation-driven insight.

Abstract

As the demand for lithium-ion batteries rapidly increases there is a need to design these cells in a safe manner to mitigate thermal runaway. Thermal runaway in batteries leads to an uncontrollable temperature rise and potentially fires, which is a major safety concern. Typically, when modelling the chemical kinetics of thermal runaway calorimetry data ( e.g. Accelerating Rate Calorimetry (ARC)) is needed to determine the temperature-driven decomposition kinetics. Conventional methods of fitting Arrhenius Ordinary Differential Equation (ODE) thermal runaway models to Accelerated Rate Calorimetry (ARC) data make several assumptions that reduce the fidelity and generalizability of the obtained model. In this paper, Chemical Reaction Neural Networks (CRNNs) are trained to fit the kinetic parameters of N-equation Arrhenius ODEs to ARC data obtained from a Molicel 21700 P45B. The models are found to be better approximations of the experimental data. The flexibility of the method is demonstrated by experimenting with two-equation and four-equation models. Thermal runaway simulations are conducted in 3D using the obtained kinetic parameters, showing the applicability of the obtained thermal runaway models to large-scale simulations.

Figures (8)

• Figure 1: ARC experimental testing (\ref{['fig:TC']}) Location of thermocouples on the cell surface to measure cell temperature (\ref{['fig:setup']}) Cell in holder inside ARC chamber (\ref{['fig:post']}) Burnt out cell post ARC test (\ref{['fig:temp_rate_from_exp']}) Heat rate versus temperature (\ref{['fig:time_temp_from_exp']}) Temperature evolution with time
• Figure 2: Division of data into stages, showing the start and end temperature for each stage (\ref{['fig:temp_rate_display']}) A two-stage fit (\ref{['fig:stage_division_4_stage']}) A four-stage fit
• Figure 3: (\ref{['fig:stage_1_linear_fit_display']}) Linear fit of stage 1 (\ref{['fig:temp_rate_2_stage_linearization']}) Two-stage fit from linearization method
• Figure 4: (\ref{['fig:CRNN_stage_i']}) CRNN subnetwork diagram for a single stage. The orange nodes denote trainable variables updated via gradient-based optimization, given by $\boldsymbol{\theta_{i}}=[A_{i},E_{a,i},h_{i},m_{i},n_{i}]$. The blue nodes represent system state variables integrated via a differentiable ODE integrator. (\ref{['fig:CRNN_workflow_sub']}) Computation graph for a single CRNN training step. Every CRNN subnetwork (Figure \ref{['fig:CRNN_stage_i']}) is assimilated to set up the ODE system, which is embedded into the training loop to learn $\boldsymbol{\theta_{i}}$.
• Figure 5: (\ref{['fig:temp_before_2_stage']}) Temperature evolution from linear fitting (\ref{['fig:temp_after_2_stage']}) Temperature evolution after CRNN training (\ref{['fig:rate_before_2_stage']}) Heat rate evolution from linear fitting (\ref{['fig:rate_after_2_stage']}) Heat rate evolution after CRNN training