DecayFormer

Trained weights of DecayFormer, a compact hybrid state space / attention model that forecasts a battery cell's full state of health (SOH) trajectory from its first 100 cycles, for all 22 aging-condition-exclusive folds of the BatteryLife benchmark as defined by BatteryMFormer. One model per fold; 791,529 parameters.

Code, reproducibility notebooks and documentation: github.com/ShreyPatel1311/DecayFormer. License: MIT.

Results of these weights

Test MAPE (%) from BatteryMFormer's released evaluation (evaluate_model.py, fp16), on each fold's own test set. Mean ± std across folds:

Domain Folds Test MAPE
Li-ion 3 2.281 ± 0.089
Zn-ion 3 6.814 ± 0.540
Na-ion 12 0.926 ± 0.282
CALB 4 2.783 ± 2.200
Fold Test MAPE Early-stopping patience
Li-ion/seed2021 2.375 30
Li-ion/seed42 2.197 30
Li-ion/seed2024 2.272 30
Zn-ion/seed2021 6.248 30
Zn-ion/seed42 6.869 30
Zn-ion/seed2024 7.324 30
Na-ion/cond400 0.660 30
Na-ion/cond401 1.346 50
Na-ion/cond402 0.843 50
Na-ion/cond403 0.834 30
Na-ion/cond404 1.001 30
Na-ion/cond405 0.831 30
Na-ion/cond406 0.487 30
Na-ion/cond407 1.463 30
Na-ion/cond408 1.006 30
Na-ion/cond409 1.138 30
Na-ion/cond410 0.721 30
Na-ion/cond411 0.785 30
CALB/cond396 5.209 30
CALB/cond397 0.946 30
CALB/cond398 0.893 30
CALB/cond399 4.082 30

Training configuration

All 22 models share one configuration: d_model 96, 3 encoder blocks, 4 attention heads, dropout 0.1; AdamW at learning rate 5e-4 with a constant schedule; batch 128 (one GPU, equal to BatteryMFormer's 2 x 64 recipe); up to 300 epochs with early stopping on validation MAPE; fp16; each fold's own seed (2021, 42 or 2024). The only per-fold difference is early-stopping patience: 50 for Na-ion cond401 and cond402, 30 for every other fold. Training ran inside BatteryMFormer's released pipeline (commit febe174) on BatteryLife v11, so data, targets, training loop and evaluation are common to all models compared with it.

Files

config.json                  architecture (DecayFormerConfig)
<Domain>/<fold>/
    model.safetensors        weights, saved from BatteryMFormer's pipeline (keys prefixed "net.")
    args.json                training arguments in that pipeline
    label_scaler             label scaler expected by that pipeline's data loader

Domains: Li-ion (seed2021, seed42, seed2024), Zn-ion (seed2021, seed42, seed2024), Na-ion (cond400 to cond411), CALB (cond396 to cond399).

Usage

from decayformer import load_fold   # pip install from the DecayFormer repository

model = load_fold("Na-ion/cond400")
# x [B, 100, 300, 4] curves, mask [B, 100], soh [B, 100, 1], cycle_feats [B, 100, 2]
trajectory = model(x, mask, soh, cycle_feats)        # [B, 5000] normalised SOH

To reproduce the table above, follow the notebooks in the repository, which evaluate these weights with BatteryMFormer's own pipeline.

Downloads last month
76
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support