Hybrid ML-MCDM Framework for EV Battery End-of-Life Routing

A research-grade pipeline that predicts an electric-vehicle battery's State of Health and Remaining Useful Life from cycle data, then recommends an end-of-life route — grid-scale energy storage, home battery, component reuse, or direct recycling — under five regulatory weight regimes. Outputs a Digital Product Passport reconciled across EU Regulation 2023/1542 Annex XIII, the Global Battery Alliance Battery Pass v1.2, and India BWMR 2022 (with 2024 and 2025 amendments).

Manuscript — in preparation.

Pipeline

flowchart LR
    subgraph INGEST["Data ingestion"]
        D1[Public cell datasets<br/>BatteryLife · NASA · CALCE · Stanford<br/>1,451 cells]
        D2[Synthetic Indian cells<br/>PyBaMM + BLAST-Lite<br/>130 cells]
        D3[Regulatory PDFs<br/>EU 2023/1542 · GBA · BWMR]
        D4[MCDM literature<br/>12+ BWM/AHP papers]
    end

    PROC[Unify · clean · split<br/>1,581 cells · 7 chemistries]

    subgraph INFER["ML inference"]
        ANOM[Anomaly gate<br/>IsoForest + VAE]
        HEALTH[State of Health + Grade<br/>XGBoost + 7 chemistry specialists]
        LIFE[Remaining Useful Life<br/>XGBoost + TCN]
        SHAP[SHAP feature attributions]
    end

    subgraph DECIDE["Decision layer"]
        BWM[Fuzzy BWM weights<br/>5 regulatory regimes]
        TOPSIS[TOPSIS ranking<br/>4 EoL routes]
    end

    DPP[Digital Product Passport<br/>schema-validated JSON]

    D1 --> PROC
    D2 --> PROC
    PROC --> ANOM
    ANOM --> HEALTH
    HEALTH --> LIFE
    HEALTH --> SHAP
    D4 --> BWM
    HEALTH --> TOPSIS
    LIFE --> TOPSIS
    BWM --> TOPSIS
    D3 --> DPP
    SHAP --> DPP
    TOPSIS --> DPP

Loading

Key results

• State of Health predicted to within 2.43 pp RMSE on held-out test cells (R² 0.996); chemistry-specialist routing lifts under-represented-chemistry grade-accuracy by up to +5 pp vs a single global model.
• Remaining Useful Life predicted to within 1.92 % of the prediction range (classical) and 2.23 % (deep learning).
• Routing recommendation flips under different regulatory regimes — the framework's headline regulatory-sensitivity finding.
• Training corpus: 1,581 cells across 7 chemistries (NMC, LFP, NCA, LCO, Zn-ion, Na-ion, other) — public lab datasets (BatteryLife, NASA-PCOE, CALCE, Stanford) plus 130 synthetic Indian-context cells generated via PyBaMM and BLAST-Lite.

Quick start

git clone https://github.com/Rishabhmannu/hybrid-ml-mcdm-battery-eol.git
cd hybrid-ml-mcdm-battery-eol
pip install -r frontend/requirements.txt
streamlit run frontend/app.py

First launch downloads ~90 MB of model weights from the Hugging Face Hub and caches them locally — subsequent runs are instant.

Tech stack

• ML + decision theory — XGBoost · PyTorch · scikit-learn · SHAP · Optuna · Fuzzy BWM · TOPSIS
• Data + simulation — pandas · NumPy · PyArrow · PyBaMM · BLAST-Lite
• Frontend + reporting — Streamlit · Plotly · ReportLab · Kaleido · jsonschema
• Hosting — Hugging Face Hub (weights) · Streamlit Community Cloud (demo)

Project structure

src/             Library code (loaders, MCDM, DPP, model wrappers)
scripts/         Training, evaluation, ablation, export scripts
frontend/        Streamlit demo + PDF cell-report export
results/tables/  Small reproducibility artefacts (metrics / logs)
data/            Runtime artefacts (full corpus stays local)

Known limitations

• The training corpus is curated lab data plus 130 synthetic Indian-context cells; generalisation to real fleet data remains future work.
• Carbon-footprint values in the Digital Product Passport are chemistry-keyed literature defaults, not measured lifecycle assessments.
• The framework has not been deployed in a live battery management system, fleet management platform, or regulator submission pipeline.

Citation

@misc{kumar2026hybrid,
  author = {Kumar, Rishabh},
  title  = {Hybrid ML-MCDM Framework for EV Battery End-of-Life Routing in India},
  year   = {2026},
  url    = {https://github.com/Rishabhmannu/hybrid-ml-mcdm-battery-eol}
}

License

Released under the MIT License. Training-corpus datasets retain their original permissive licences; synthetic Indian-context cells generated for this work are released under MIT.

Contact

Rishabh Kumar — IIIT Allahabad — rishabhkumards07@gmail.com · LinkedIn · github.com/Rishabhmannu