Skip to main navigation Skip to search Skip to main content

Cross-chemistry battery state-of-health estimation from adaptive partial charging using interpretable physics-informed Efficient Kolmogorov–Arnold networks

  • Zekariyas Yohannes
  • , Yonghong Xu
  • , Yidi Wei
  • , Jian Li
  • , Boru Jia*
  • *Corresponding author for this work
  • Beijing Institute of Technology
  • Beijing Information Science & Technology University

Research output: Contribution to journal › Article › peer-review

Abstract

Accurate, robust state-of-health (SOH) estimation is essential for safe battery operation and second-life deployment, yet most existing methods depend on fixed feature extraction windows rarely available in real-world charging and offer limited interpretability or cross-chemistry generalization. To address these issues, we propose the physics-informed efficient Kolmogorov–Arnold network (PI-EKAN), which estimates SOH from adaptive partial charging. Prefix-specific Bayesian optimization (PS-BO) selects the most informative state-of-charge window for each chemistry and extracts seven voltage-based descriptors; this feed-coupled solution and dynamics networks are trained under a bounded uncertainty-weighted objective that balances the data, physics, and monotonicity constraints. Across eight datasets, PI-EKAN reaches an average MAPE of 0.466%, an RMSE of 0.529%, and an R² above 0.979, cutting MAPE by 35.6% relative to a physics-informed baseline while using only 1800 parameters, far fewer than comparable recurrent Kolmogorov–Arnold models. PS-BO identifies chemistry-specific charging windows that stay informative under restricted SOC visibility, reduced sampling frequency, and simulated battery-management-system errors. In cross-chemistry tests, adapting the pretrained model with one labeled target cell cuts average MAPE from 6.35% source-only transfer to 1.19% using hybrid low-rank EKAN-LoRA fine-tuning with half the trainable parameters across sixteen transfer pairs. Component ablation identifies the bounded adaptive weighting as the key stabilizing element, and post-hoc attribution and learned-spline analyses consistently rank voltage-shape descriptors as the leading inputs, suggesting that the model captures degradation-related structure rather than dataset artifacts. These results establish PI-EKAN as a compact, interpretable, chemistry-transferable framework for SOH estimation from the incomplete charging data real battery systems produce.

Original languageEnglish
Article number100928
JournalEnergy and AI
Volume26
DOIs
Publication statusPublished - Dec 2026
Externally publishedYes

Keywords

  • Battery state-of-health estimation
  • Interpretable battery diagnostics
  • Kolmogorov–Arnold networks
  • Partial charging
  • Physics-informed learning
  • Transfer learning

Fingerprint

Dive into the research topics of 'Cross-chemistry battery state-of-health estimation from adaptive partial charging using interpretable physics-informed Efficient Kolmogorov–Arnold networks'. Together they form a unique fingerprint.

Cite this