Physics-Enhanced Deep Learning for Proactive Thermal Runaway Forecasting in Li-Ion Batteries
Researchers have developed a Physics-Informed Long Short-Term Memory (PI-LSTM) framework to improve the prediction of thermal runaway in lithium-ion batteries, a critical safety concern for modern energy storage systems. Traditional data-driven models often violate thermodynamic principles, while physics-based models are computationally expensive. The proposed PI-LSTM integrates heat transfer equations into the deep learning architecture via a physics-based regularization term. By utilizing multi-feature inputs such as state of charge, voltage, current, mechanical stress, and surface temperature, the model forecasts battery temperature evolution while enforcing thermal diffusion constraints. Extensive experiments across thirteen datasets demonstrated that PI-LSTM significantly outperforms standard LSTM, CNN-LSTM, and MLP models, achieving an 81.9% reduction in root mean square error and an 81.3% reduction in mean absolute error. This approach eliminates non-physical temperature oscillations and enhances generalization across diverse operating conditions. The study confirms that physics-informed deep learning provides a viable pathway for interpretable, accurate, and real-time thermal management in next-generation battery systems, bridging the gap between data-driven efficiency and physical consistency.
Wire timeline
Physics-Enhanced Deep Learning for Proactive Thermal Runaway Forecasting in Li-Ion Batteries
Researchers have developed a Physics-Informed Long Short-Term Memory (PI-LSTM) framework to improve the prediction of thermal runaway in lithium-ion batteries, a critical safety concern for modern energy storage systems. Traditional data-driven models often violate thermodynamic principles, while physics-based models are computationally expensive. The proposed PI-LSTM integrates heat transfer equations into the deep learning architecture via a physics-based regularization term. By utilizing multi-feature inputs such as state of charge, voltage, current, mechanical stress, and surface temperature, the model forecasts battery temperature evolution while enforcing thermal diffusion constraints. Extensive experiments across thirteen datasets demonstrated that PI-LSTM significantly outperforms standard LSTM, CNN-LSTM, and MLP models, achieving an 81.9% reduction in root mean square error and an 81.3% reduction in mean absolute error. This approach eliminates non-physical temperature oscillations and enhances generalization across diverse operating conditions. The study confirms that physics-informed deep learning provides a viable pathway for interpretable, accurate, and real-time thermal management in next-generation battery systems, bridging the gap between data-driven efficiency and physical consistency.
cs.AI updates on arXiv.org