A transformer guided multi modal learning framework for predictive and causal assessment of thermal runaway in high energy batteries
Machine Learning approaches from the present state either use unimodal data, unable to model elegant long spatial-temporal dependencies in warning systems or create early warning response datasets with limited quantitative interpretability sets. To address these shortcomings, this work introduces T-...
Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Article |
| Language: | en |
| Published: |
Nature Research
2025
|
| Subjects: | |
| Online Access: | https://umpir.ump.edu.my/id/eprint/47539/1/J%202025%20SR%20Sameer%20M.M.Noor%20Thermal%20Runway%20HE%20Batt%20RDU240117.pdf https://doi.org/10.1038/s41598-025-20886-x https://umpir.ump.edu.my/id/eprint/47539/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Machine Learning approaches from the present state either use unimodal data, unable to model elegant long spatial-temporal dependencies in warning systems or create early warning response datasets with limited quantitative interpretability sets. To address these shortcomings, this work introduces T-RUNSAFE, a multi-pronged, machine learning-based predictive prototype for thermal runaway assessment. The framework integrates five specialized modules: (1) ST-Former, a spatiotemporal transformer that encodes thermal gradients from thermal images and sensor logs using temporal self-attention over LSTMs, thus is superior to traditional LSTMs for capturing evolving thermal patterns; (2) FUSE-GEN, adversarial trained dual-encoder variational autoencoder, fusing acoustic emission (AE) signals and thermal embeddings into a shared latent space for earlystage internal degradation detection; (3) DEGRA-GNN, a graph attention network that capitalizes on battery electrode topology to model the spatial propagation of thermal faults; (4) CAUS-RUN, a counterfactual simulation engine employing structural causal models to attribute risk to specific spatial zones for interpretability; and (5) SENSOR-RL, a reinforcement learning module optimizing sensor sampling policies on real-time risk levels that cuts down on sensor power while still holding to detection accuracy. The experimental results show great early prediction accuracy (AUC-ROC > 0.96), high spatial degradation localization accuracy (93.5%), and a 37% decrease in power consumption of sensing. T-RUNSAFE predicts, interprets, and optimizes resource utilization for thermal runaway risk assessment. By integrating deep learning, physics-informed modeling, and causal reasoning, it enables real-time battery safety monitoring. Although challenges remain regarding sensor cost, computational overhead, and chemistry generalization, the study demonstrates the feasibility of advanced onboard battery management systems tailored for next-generation energy applications. |
|---|