TEMPEST: a machine learning framework for battery thermal prediction and Chemistry selection

Lithium-ion batteries are the cornerstone of modern portable electronics and electric vehicles. However, cell temperature profoundly influences their operational efficiency, safety, and lifetime. We present TEMPEST: Thermal Estimation and Chemistry Selection Toolkit, a data-driven framework that couples machine learning-based temperature prediction with chemistry recommendation. Our method trains long-short-term-memory (LSTM) models on three lithium-ion battery chemistries: nickel-cobalt-aluminum (NCA), nickel-manganese-cobalt (NMC), and lithium-iron-phosphate (LFP). These models are used to estimate surface temperature from voltage and current time series under controlled ambient conditions. TEMPEST consistently achieves high fidelity under varying operational load and ambient conditions, with mean absolute error below 0.5∘C for the same chemistry. The proposed approach is also extended to cross-chemistry transfer, providing stable generalization at room temperatures, with typical errors up to 2.0∘C, enabling accurate thermal monitoring without intrusive sensors. At extreme temperature conditions, the mean absolute error can rise up to 5∘C, giving higher errors for larger electrochemical mismatch between chemistries. Further, to experimentally validate the proposed method, in-lab data is collected using a standard electro-thermal battery cycler, where thermal prediction errors as low as 0.4∘C are achieved, demonstrating strong adaptation capability under limited data. When benchmarked against popular methods, our model achieves a 10% accuracy improvement while requiring approximately 20% less compute. We also introduce a battery selection algorithm that, given ambient temperature, recommends the battery chemistry expected to operate at the lowest temperature. Our results reveal that NMC and NCA cells are preferable at sub-ambient conditions, whereas LFP chemistries offer thermal stabi.