With the increasing demand for applications in vehicles and energy storage systems, lithium-ion batteries have attracted extensive research interest. Thermal runaway in these batteries, which can lead to rapid temperature rise, flammable gas release, and even fires, directly imposes safety concerns. Previous studies have utilized differential scanning calorimetry (DSC) to infer thermal kinetic mechanisms for individual battery components. However, the commonly used Kissinger analysis involves oversimplified assumptions which lead to erroneous quantification of the kinetic parameters and inaccurate physical interpretation of the results. Here, we propose an extension of the recently developed Chemical Reaction Neural Network (CRNN) framework that is not limited by the simplifications from traditional optimization methods and can learn complex, multi-step thermal kinetics from DSC data. After a proof of concept, we leverage this novel approach to improve recent thermal decomposition models for nickel–cobalt–manganese oxide (NCM) cathodes. Its generality and enhanced learning capability enable improved models with better agreement to the data that account for the actual physical coupling between multi-step reaction pathways. The successful development of the CRNN approach to learn such thermal kinetic models from simple DSC data demonstrates its potential to advance thermal runaway modelling for lithium-ion batteries and other complex kinetic systems.