Thermal decomposition of lithium-ion battery cathode materials poses a critical safety challenge for the design and management of high-energy batteries. Accurate determination of reaction kinetic parameters is essential for understanding subsequent thermal runaway behavior. A deep learning framework was developed to automatically predict kinetic parameters from multi‑rate differential thermogravimetry (DTG) curves of pristine (non-delithiated) nickel–cobalt–manganese oxide (NCM) cathode materials under reductive gas attack. By jointly fitting the Johnson–Mehl–Avrami (JMA) model to experimental benchmarks, we established material-specific kinetic compensation effect (KCE) manifolds for four commercial NCM cathode materials (NCM811, 622, 523, and 111). These manifolds guided the construction of a synthetic dataset containing 20,000 multi-rate samples, designed to emulate complex peak-shift dynamics and instrumental noise. A specialized one-dimensional convolutional neural network (1D CNN) was engineered to process synchronized three-rate DTG sequences (10, 30, and 60 K·min⁻¹), allowing the model to implicitly learn the governing kinetic laws from raw signal topology. On the synthetic training set, the multi-rate CNN achieved an exceptional R 2 of 0.94 for both activation energy ( E a ) and pre-exponential factor ( lg A ), with validation set R 2 of 0.92. Beyond numerical precision, the CNN faithfully reconstructed composition-specific KCE lines with slope deviations under 0.5%, proving its ability to capture subtle “kinetic fingerprints” without manual feature engineering. While delithiated cathodes exhibit more complex behaviors, this pristine-based foundation provides a scalable and reproducible trajectory for high-throughput screening and future transfer learning to diverse battery states of charge. By reducing diagnostic latency from hours to milliseconds, this “kinetics intelligence” approach provides a scalable and reproducible tool for high-throughput cathode screening and real-time safety assessment in industrial-scale battery management. • A one-dimensional CNN predicts activation energy (Ea) and pre-exponential factor (lgA) directly from raw DTG curves, removing the need for manual fitting or feature engineering. • A physics-guided synthetic dataset of 15,000 DTG curves was generated using JMA and KCE across multiple NCM compositions and heating rates, capturing realistic peak variations and noise. • The CNN achieves R² ≈ 0.87 for both Ea and lgA, outperforming MLP and tree-based models, with robustness to noise, baseline drift, and peak-shape variability. • Convolutional layers capture local, temperature-dependent features linked to thermal decomposition, providing mechanistic insight into performance gains. • The framework enables rapid, reproducible kinetic analysis for real-time battery safety assessment, high-throughput cathode screening, and extension to other chemistries and atmospheres.
Zhou et al. (Wed,) studied this question.