A Hybrid State Space-Attention Architecture for Forecasting the Trajectories of Battery Degradation
Summary
- Researchers introduced DecayFormer, a compact hybrid state-space and attention architecture for forecasting early-life battery degradation trajectories.
- DecayFormer combines cycle-level selective recurrence with an attention mechanism, state-of-charge interval statistics, and a master curve head.
- Evaluated on the BatteryLife benchmark across 22 folds without fold-specific tuning, it achieved mean test MAPE values of 2.25% on Li-ion, 2.76% on CALB, 0.95% on Na-ion, and 7.19% on Zn-ion.
Why it matters
DecayFormer demonstrates that compact sequence models can accurately predict battery degradation across diverse chemistries without heavy scaling or per-fold tuning, improving computational efficiency.