AI Battery Lifecycle model assets
This repository stores versioned artifacts for the AI Battery Lifecycle Space and the canonical GitHub source repository.
Reviewer-revision v3
The revised v3 estimates current-cycle state of health (SOH) from the observable 4.0–3.6 V portion of a discharge. It replaces the pre-review v3, whose input set contained functions of the same current-cycle capacity used to define SOH. Those target proxies made the previous score unsuitable as evidence of unseen-battery performance.
The release manifest at v3/models.json is authoritative. It records:
- the ordered 18-feature prediction contract;
- the selected deployable model and complete grouped-validation metrics;
- artifact size, SHA-256 digest, and CPU inference latency;
- 20 evaluated models and the subset of serialized tabular pipelines;
- the battery-grouped split design and selection criterion.
v3/datamap.json records the target definition, public rated capacities used
for C-rate/EFC normalization, forbidden target-derived inputs, official source
checksums, and the quality-exclusion audit path. Publication tables and figures
are generated directly from row-level experiment outputs in the source
repository and copied here with their manifest.
The v3/results/ directory contains fold metrics, statistical outputs, and
environment metadata. Row-level prediction CSVs are stored as *.csv.gz to
keep the scientific audit trail complete without inflating ordinary Git blobs.
Validation scope
- NASA: five battery-grouped folds, fixed seeds 17, 42, and 2026.
- CALCE and Oxford: grouped within-dataset evaluation using the same 20-model catalog.
- External transfer: models fitted on NASA and frozen before scoring CALCE and Oxford.
- Metrics: MAE, RMSE, MAPE, R², adjusted R² where defined, residuals, absolute five-SOH-point coverage, per-battery scores, cluster-bootstrap intervals, and paired Wilcoxon tests with Holm correction.
The old root models/, v1, and v2 directories are retained for historical
compatibility. Their metrics and feature contracts are not comparable to the
reviewer-revision v3.
Intended use and limitations
These artifacts support reproducible research on laboratory battery-aging datasets. They are not certified for a battery-management system, charging control, safety decisions, warranty adjudication, or field deployment. The v3 model does not provide a validated RUL forecast or a causal charging-policy recommendation. Cross-dataset errors must be inspected before any adaptation or application to a new chemistry, manufacturer, sensor stack, or duty cycle.
Loading
from huggingface_hub import snapshot_download
local_dir = snapshot_download(
repo_id="NeerajCodz/aiBatteryLifeCycle",
repo_type="model",
)
Use the manifest’s feature_set order when constructing inputs and verify the
listed SHA-256 value before deserializing an artifact.
Citation
Please cite the accompanying manuscript after its bibliographic record is available. Until then, cite this repository URL and the exact v3 commit or tag used in the experiment.
- Downloads last month
- 29