🔋 BatteryLake

The Open Data Foundation for Battery Prognostics & Health Management

One naming standard. One ETL pipeline. One quality gate. Battery data, research-ready.

Explore the Platform · Who It Serves · Platform Tour · How to Use · Data Schema · Benchmarks · Contributing

---

💡 Why BatteryLake?

Battery degradation research is held back by a data fragmentation crisis. Every lab releases cycling data in its own format — its own column names, units, sampling rates, and file conventions — scattered across Zenodo, Figshare, Mendeley, OSF, institutional servers, and paper supplements. Before anyone trains a single model, weeks disappear into downloading, decoding, cleaning, and reconciling incompatible files. Worse: because everyone preprocesses differently, published accuracy numbers are rarely comparable across papers.

BatteryLake removes that barrier. We curate, standardize, and quality-check battery cycling datasets from laboratories worldwide into a single uniform schema, then expose them through an interactive catalog, a reproducible benchmark suite, and open APIs — so models for State-of-Health (SOH) estimation and Remaining-Useful-Life (RUL) prediction are finally compared on science, not on preprocessing luck.

Fair, reproducible comparison across 31 institutions and 19 years of battery aging research.

---

🎯 Who BatteryLake Serves

BatteryLake is built as community infrastructure for three overlapping research communities:

	You are…	BatteryLake gives you…
🔬	A battery researcher studying degradation, chemistry, or cell design	A curated catalog of 40+ aging datasets across LFP / NMC / LCO / NCA chemistries and 7 form factors — searchable by chemistry, C-rate, temperature, and institution, with DOI-level provenance for every source. Publish your own data through the contribution pipeline and make it instantly reusable.
🤖	An ML researcher building SOH / RUL / degradation models	Standardized benchmark tasks with pinned train/val/test splits, a reference model suite (Ridge → Random Forest → XGBoost → MLP → LSTM → Transformer → PINN), and unified RMSE / MAE / MAPE reporting — so your new architecture gets an honest, like-for-like comparison.
📊	A data scientist / engineer who needs clean time-series at scale	One schema for everything: metadata.json + timeseries.parquet + cycle_summary.csv per dataset, loadable in two lines of pandas. Machine-readable quality reports, a machine-readable naming standard, and REST APIs for programmatic access.

If your work touches battery data from more than one source, BatteryLake is designed to save you weeks.

---

📸 Platform Tour

🗂️ Dataset Catalog — browse, filter, download

Search 40+ curated datasets by chemistry, form factor, category, domain, and duty profile. Every dataset card carries ETL status, cell/cycle counts, and links to both source and processed data.

🛡️ Quality Assessment — an auditable gate before training

Every dataset is scored across four dimensions — completeness, consistency, accuracy, validity — backed by 26 physical-plausibility and schema rules (voltage windows, coulombic-efficiency bounds, capacity monotonicity, timestamp integrity…). Each dataset ships a machine-readable quality_report.json that approves or blocks downstream training.

⚡ Developer Console — the same platform, programmable

Explore REST endpoints interactively: query the dataset catalog, stream per-cell time-series, launch benchmark runs, and submit feature recipes. Responses embed provenance and quality scores by default.

---

📦 What's Inside

Metric	Value	Notes
Curated datasets	40	Public archives + internal NTU experiments
Individual cells	1,819	Across all chemistries and form factors
Charge/discharge cycles	2.24M	Standardized, per-cycle aligned
Data volume	245 GB	Source data as published by the originating labs
Source institutions	31	Laboratories worldwide
Publication span	2007 – 2026	19 years of battery aging research

These figures are computed from the dataset catalog in js/main.js (the Home page derives them at load time), so this table should be refreshed whenever datasets are added.

Chemistries: LFP · NMC · NMC811 · LCO · NCA · multi-chemistry Form factors: 18650 · 21700 · Pouch · Prismatic · Cylindrical · Automotive · EV-BMS Sources include: NASA PCoE · CALCE (UMD) · Stanford–MIT–TRI · Oxford (Howey) · RWTH Aachen · Sandia (SNL) · HNEI · XJTU · KIT · Tsinghua · CMU · Stanford (Onori) · TUM · Beihang · Imperial College · Iowa State (ILCC) · EVERLASTING (4TU) · NTU EEE — and more.

---

🏷️ Naming Standard

Every dataset receives a machine-readable reference name encoding its key experimental parameters — so a filename alone tells you what you are loading:

YYYY _ SOURCE _ CHEMISTRY _ FORMFACTOR _ CHRGC _ DCHRGC _ TEMPT

Field	Meaning
YYYY	Publication / collection year
SOURCE	Lab or institution (e.g. NASA_PCoE, Imperial_Kirkaldy)
CHEMISTRY	LFP / NMC / LCO / NCA / MultiChem
FORMFACTOR	18650 / 21700 / Pouch / Prismatic …
CHRGC / DCHRGC	Charge / discharge C-rates (or MultiC)
TEMPT	Ambient test temperature in °C (or MultiT)

2007_NASA_PCoE_LCO_18650_1C_1C_25T
2019_Stanford_MIT_TRI_LFP_18650_MultiC_30T
2024_Imperial_Kirkaldy_NMC_21700_MultiC_MultiT
2026_NTUEEE_Ampace-Samsung_LFP-NMC_21700_2C_2C_25T

---

🧱 Standardized Data Schema

Heterogeneous raw exports are converted through one reproducible ETL pipeline:

Raw Data → Download & Validate → ETL Processing → Quality Gate → Split Protocol → Training → Evaluation

Each processed dataset is published as four artifacts:

File	Contents
metadata.json	Cell-level metadata: chemistry, form factor, nominal capacity, source lab, test conditions, publication DOI
timeseries.parquet	Per-cycle, per-cell measurements: voltage_V, current_A, temperature_C, capacity_Ah, timestamp
cycle_summary.csv	Cycle-level aggregates: discharge/charge capacity, coulombic efficiency, internal resistance, cycle number
dataset_note.md	Human-readable provenance, known issues, preprocessing decisions, usage recommendations

Loading any dataset is the same two lines, regardless of which lab it came from:

import pandas as pd

ts     = pd.read_parquet("2007_NASA_PCoE_LCO_18650_1C_1C_25T/timeseries.parquet")
cycles = pd.read_csv("2007_NASA_PCoE_LCO_18650_1C_1C_25T/cycle_summary.csv")

soh = cycles["discharge_capacity_Ah"] / cycles["discharge_capacity_Ah"].iloc[0]

---

📊 Benchmark Tasks

BatteryLake defines standardized prognostics tasks with fixed, reproducible split protocols:

• 🔋 SOH Estimation — predict state-of-health (capacity fade) from cycling signals
• ⏳ RUL Prediction — predict remaining useful life (cycles to end-of-life)

Split protocols: random · temporal · cross-cell — always with pinned seeds. Reference model suite: classical baselines (Ridge, Random Forest, XGBoost) → deep sequence models (MLP, LSTM, Transformer) → physics-informed neural networks (PINN). Metrics: RMSE · MAE · MAPE, reported on identical folds.

Every benchmark run exports a portable training package, so results can be reproduced outside the platform.

---

🚀 Getting Started

# Clone the repository
git clone https://github.com/tianwen1209/batterylake.git
cd batterylake

# Serve the interactive catalog locally
python3 -m http.server 8000
# then open http://localhost:8000

Or visit the hosted platform directly to browse datasets, inspect quality reports, configure benchmark experiments, and download both source and processed data.

---

🧭 How to Use

A typical workflow, end to end:

1 · Find your data

Open Datasets, then search or filter by chemistry, form factor, category, domain, or duty profile. Click any dataset card to inspect its metadata, Processed Data Files Check, and notes — the detail view links to both the original source (DOI) and the processed download.

2 · Load it in Python

Every processed dataset follows the same schema (see Standardized Data Schema):

import pandas as pd

ts     = pd.read_parquet("<ref_name>/timeseries.parquet")   # V / I / T per cycle per cell
cycles = pd.read_csv("<ref_name>/cycle_summary.csv")        # one row per cycle

3 · Run a benchmark

On the Benchmarks page: select a dataset → choose features → set the train/val/test split → pick a model, then download the generated training package and run it locally:

unzip bt_benchmark_soh_lstm.zip
cd bt_benchmark_soh_lstm
mkdir -p data
# copy the processed dataset folder into data/ first, e.g.:
# cp -R /path/to/2019_Stanford_MIT_TRI_LFP_18650_MultiC_30T data/

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
bash run_benchmark.sh

When the run finishes, upload the outputs/ folder back on the same page to view metrics, per-cell prediction trajectories, and model comparisons.

4 · Check quality before you train

Every dataset's Quality Assessment report (completeness / consistency / accuracy / validity + physical-plausibility checks) is available on the platform and as machine-readable JSON — use it to decide whether a dataset enters your training set.

5 · Go programmatic

Use the APIs & Applications console to query the catalog, stream cell time-series, and launch benchmark runs from code.

---

🤝 Contributing

BatteryLake grows with the community. Ways to contribute:

• 📤 Contribute a dataset — use the Contribute page on the platform; all submissions pass through the same ETL and quality-assessment pipeline, and retain full attribution + DOI linkage to your original publication.
• 🧪 Add a model or task — reference implementations live in the benchmark suite; new baselines and prognostics tasks are welcome.
• 🐛 Report issues — schema problems, quality-rule false positives, or catalog corrections via GitHub Issues.

Every dataset keeps attribution to its original source. Please cite the underlying datasets when you use them.

---

👥 Team

Developed at Nanyang Technological University, Singapore.

Group	Members
Project Supervisors	Prof. Yonggang Wen (Principal Investigator) · Dr Wang Hao (Research Fellow)
Core Researcher	Zhu Tianwen (PhD Candidate)
Web Engineering	Cai Yezi (Frontend Engineer)

---

📌 Citation

If BatteryLake supports your research, please cite:

@misc{zhu2026batterylakeagenticphysicsgroundedcuration,
      title={BatteryLake: Agentic, Physics-Grounded Curation of Heterogeneous Battery Aging Data and Benchmarking}, 
      author={Tianwen Zhu and Hao Wang and Yonggang Wen},
      year={2026},
      eprint={2607.09762},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2607.09762}, 
}

---

🙏 Acknowledgements

BatteryLake stands on the shoulders of the open battery-data community. We gratefully acknowledge every laboratory and repository whose published datasets make this platform possible — NASA PCoE, CALCE, the Stanford–MIT–TRI consortium, Oxford, RWTH Aachen, Sandia, HNEI, and many more. Each dataset retains attribution to its original source and DOI.

---

BatteryLake — turning scattered battery data into a shared scientific foundation.

Home page statistics

Home metrics describe complete original releases rather than downloaded subsets. See the counting methodology and source audit (中文) before updating the catalog.