Live · Battery Data Intelligence Platform

research-ready.

BatteryLake unifies the world's fragmented battery cycling datasets into one standardized, quality-assessed foundation — so models are compared on science, not on preprocessing luck.

Explore Datasets View on GitHub

batterylake.org / datasets ?quality=ready

LCO 2007_NASA_PCoE_LCO_18650_1C_1C_25TNASA Prognostics CoE 34 cells 0.96

LFP 2019_Stanford_MIT_TRI_LFP_18650_MultiC_30TStanford · MIT · TRI 124 cells 0.94

NMC 2024_Imperial_Kirkaldy_NMC_21700_MultiC_MultiTImperial College London 21 cells 0.92

LFP 2026_NTU_Ampace-Samsung_LFP-NMC_21700_2C_2C_25TNTU EEE · Singapore 16 cells 0.95

SOH 94.2%cycle 187 · B0005

Quality gate passed26 validation rules

0

Curated datasets

0

Worldwide institutions

0

Years span

2007-2026

0

Total cycles

0

Total cells

0 GB

Data volume

How it works

From raw lab exports to comparable results

Every dataset travels the same auditable path, so downstream numbers mean the same thing everywhere.

Curate & Ingest 40+ public and lab datasets, tracked with DOI-level provenance.

Standardize One ETL pipeline, one schema: metadata, time-series, cycle summaries.

Quality Gate 26 physical-plausibility and schema rules score every dataset.

Benchmark & Share Reproducible SOH / RUL runs with pinned splits and unified metrics.

Platform

One workspace for the full research loop

Browse, benchmark, and audit — each surface is built on the same standardized data core.

Benchmarks

Reproducible experiments, not preprocessing folklore

Random, temporal, and cross-cell split protocols
7 reference models — Ridge to Transformer and PINN
RMSE · MAE · MAPE reported on identical folds

Open the benchmark workbench

Split protocol — cross-cellseed 42

Train

62%

Val

18%

Test

20%

RMSE0.0121

MAE0.0094

MAPE1.08%

Quality

An auditable gate before any model sees the data

Physical plausibility: voltage windows, energy balance
Signal-level QC maps for V(t), I(t), T(t), Qd(n)
JSON reports wired into the benchmark pipeline

See the quality methodology

Quality gate — NASA PCoEready

0.96overall

Voltage range 2.0–4.5 V

Coulombic efficiency 95–105%

Temperature drift — 1 review

APIs

The same platform, programmable

RESTful endpoints with token auth
Parquet, CSV, and JSON exports
Embedded DOI citations in every response

Explore the developer console

GET api.batterylake.org/v1/datasets

import requests

resp = requests.get(
  "https://api.batterylake.org/v1/datasets",
  params={"chemistry": "LFP", "quality": "ready"}
)

for ds in resp.json()["items"]:
  print(ds["ref_name"], ds["quality_score"])

Start with research-ready battery data today.

Browse the catalog, download standardized packages, or contribute your lab's datasets to the community.

Browse Catalog Contribute Data

Platform Usage

Total Pageviews —

Dataset Downloads —

Skill Uses —

Design by

In partner with

0 datasets

Benchmark results

Dataset Task Model

Source
Experiment scope
Task definition

Models evaluated · Unranked

Linear Regression
Random Forest
XGBoost
LSTM
Transformer
CNN
PINN

Loading prediction curves…

Run your own benchmark

Choose Evaluation Task

SOH Estimation

01Select Dataset

02Select Input Signals

Configure Data Partitioning

Train

—

Validation

—

Test

—

Excluded not used in train / validation / test 0 cells

Drag cells between Train / Validation / Test or into Excluded to leave them out

Select Benchmark Models

Execute Local Training Package

01Download Training Package

Get the configured dataset, model architecture, and runtime environment.

bt_benchmark_soh_lstm.zip

02Run in Your Terminal

Unzip the package, open your terminal, and execute the following command to start local training.

unzip bt_benchmark_soh_lstm.zip
cd bt_benchmark_soh_lstm
mkdir -p data
# copy the processed dataset folder OR a .zip of it into data/
# example: cp -R /path/to/2019_Stanford_MIT_TRI_LFP_18650_MultiC_30T data/
# example: cp /path/to/2019_Stanford_MIT_TRI_LFP_18650_MultiC_30T.zip data/
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
bash run_benchmark.sh

03Upload Output Folder

When training finishes, upload the generated output folder containing your results.

Drag & drop the output folder here, or click to browse Expected: metrics.json, predictions.csv, and predictions/<cell>.csv (one file per cell, every cycle)

Waiting for local output folder

Benchmark Evaluation Results

Performance Highlights

Best Model—

Test RMSE—

True SOH Pred SOH

Step 1 of 6 - Benchmark Task

How the Studio Works

01Select Dataset

02Feature Identification

03Calibrate Digital Twin

04Data Generation Setup

PIMLGeneration Engine

05Data Generation Results

Inside the Calibration Engine

PIML Objective

ℒ(θ) = ℒ_data + λ ℒ_physics

Inputs

Electrical Data Thermal Data Operating Conditions Physics Constraints

Battery measurements feeding a physics-informed neural network

Studio Workspace

Start by selecting a dataset

01

Select Dataset

02

Feature Identification

Voltage (V)

Current (A)

Temperature (°C)

Capacity (Ah)

Internal Resistance (mΩ)

SOC Range (%)

03

Calibrate Battery Digital Twin

Training LossTrain Val

0%Twin build

Calibration Progress

Select and confirm a dataset to build its digital twin

04

Synthetic Data Generation Setup

05

Synthetic Data Generation Results

Build a digital twin, adjust the setup, then generate data.

— models

0

Completed

0

In Progress

0

Pending

0

Skipped

ID	Dataset	Status	Meta	Series	Summary	QC

Format pattern · Connected

BatteryLake reference name sequencer

Inspecting Field 1 of 7

Assembled reference name


          2007_NASA_PCoE_LCO_18650_1C_1C_25T

Interaction: automatic field walkthrough; hover or focus to inspect one field.7 fields · underscore-delimited

Examples

Dataset	ref_name
NASA PCoE	2007_NASA_PCoE_LCO_18650_1C_1C_25T
Stanford-MIT-TRI	2019_Stanford_MIT_TRI_LFP_18650_MultiC_30T
NTU EEE Internal	2026_NTU_Ampace-Samsung_LFP-NMC_21700_2C_2C_25T
Imperial 21700	2024_Imperial_Kirkaldy_NMC_21700_MultiC_MultiT

Field Vocabulary

Sources (25+)

NASA_PCoE · CALCE_UMD · Stanford_MIT_TRI · Oxford_Howey · RWTH_Aachen · NTUEEE · SNL · HNEI · UL_Purdue · XJTU · KIT · Tsinghua · CMU_Bills · Stanford_Onori · TUM · Beihang · Imperial · ISU_ILCC · EVERLASTING_4TU · Mendeley · Figshare · Zenodo ...

Chemistries

LFP · NMC · NMC811 · LCO · NCA · LiIon (generic) · MultiChem

Form Factors

18650 · 21700 · Pouch · Prismatic · Cyl · Auto · EV-BMS

Getting Started

Clone the repository and install the conda environment.

git clone https://github.com/tianwen1209/BatteryLake-Benchmark-DataPrep.git
cd BatteryLake-Benchmark-DataPrep
conda create -n batterylake python=3.10
conda activate batterylake
pip install -r requirements.txt

Dataset Registry

All datasets are tracked in dataset_registry.csv with columns for dataset identity, DOI, source URL, assigned owner, ref_name, processing status, QC status, last update, and notes.

Evaluation

The benchmark evaluation framework uses evaluate.py and dataset_interface.py to run baseline models across selected datasets and split protocols.

Team

Research and engineering contributors

Project Supervisors

YW

Prof. Yonggang Wen

Principal Investigator

WH

Dr Wang Hao

Research Fellow

Core Researcher

ZT

Zhu Tianwen

PhD Candidate

Web Engineering

CY

Cai Yezi

Frontend Engineer

Affiliation

Nanyang Technological University

Singapore

Repository

BatteryLake-Benchmark-DataPrep

github.com/tianwen1209/BatteryLake-Benchmark-DataPrep

Licence

BatteryLake provides access to battery datasets, metadata, benchmarking resources, and related research materials collected from multiple sources.

Individual datasets, algorithms, and other resources may come with their own licences and citation requests. Please honour these requirements. BatteryLake will display the applicable licence and citation information whenever it is known.

Before downloading, redistributing, modifying, or using a resource, please review the licence and usage terms provided on its dataset or resource page. When source-specific terms are available, those terms take precedence over the general information provided on this page.

BatteryLake does not grant additional rights to third-party content beyond those provided by the original authors, institutions, or data owners.

Citation

If you use BatteryLake, its curated datasets, standardized outputs, benchmarking resources, or platform tools in academic work, please cite the BatteryLake paper below.

BatteryLake: Agentic, Physics-Grounded Curation of Heterogeneous Battery Aging Data and Benchmarking

Tianwen Zhu, Hao Wang, and Yonggang Wen, 2026.

BibTeX

@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}
}

When using an individual dataset, algorithm, model, or other contributed resource, please also cite its original authors and follow any citation instructions shown on the corresponding resource page.

Citing both BatteryLake and the original resource helps ensure that the platform contributors, dataset creators, and research institutions receive appropriate credit.

Upload a dataset to assess

No file selected.

Precomputed catalog report

Sample CSVs to try uploading

Loading…

Example Result dataset_03_quality_assesment.json · Assessed 14:32 · example values Ready 1 warning

0.97

Completeness

Missing channels2.1%

0.95

Consistency

Sequence issues3 flags

0.92

Accuracy

Physical checks5 / 6 pass

1.00

Validity

Schema errors0

0.96

Overall Score

Ready 1 warning

Physical Plausibility Checks

High-signal checks that determine whether a dataset can safely enter preprocessing and benchmark training.

5 passed - 1 warning

Voltage Range Validation
All cell voltages within 2.0V-4.5V nominal operating range for the stated chemistry.
Energy Balance Check
Charge/discharge energy integral consistency; coulombic efficiency remains within 95-105% per cycle.
Capacity Monotonicity
Degradation trajectory follows expected non-increasing trend with allowable recovery windows.
Temperature Consistency
Cell surface temperature must remain within 5°C of stated test condition; one dataset needs review.
Timestamp Integrity
Monotonically increasing timestamps with no negative intervals or unreasonable gaps above 24h.
Current Direction Consistency
Charge and discharge current signs follow one convention throughout the dataset.

Diagnostic Output

quality_report.json

{
  "dataset_id": "dataset_03",
  "quality_score": {
    "completeness": 0.97,
    "consistency": 0.95,
    "accuracy": 0.92,
    "validity": 1.00
  },
  "overall": 0.96,
  "gate": "ready_with_warning",
  "checks": [
    { "name": "voltage_range", "passed": true },
    { "name": "temperature_consistency", "status": "review" },
    { "name": "capacity_mono", "passed": true }
  ],
  "generated_at": "2026-04-28T12:00:00Z"
}

01Set Up the Processing Skill

01

Download the skill package

Save batterylake-processing.zip to your BatteryLake data repository.

View SKILL.md Processing standard

See package contents

02Choose a Dataset and Send the Prompt to Your Agent

01Choose a dataset

inRaw_Dataset/…

outProcessed_Dataset/…
02Paste the prompt to your agent
The agent reads the standard and the dataset TODO, then continues from the first unfinished gate.
03Watch it work
- Progress lives in status.jsonone line per gate, plus TODO.md
- Interrupted runs resumefinished steps are skipped when fingerprints match
- Raw files stay untouchedundecodable data is reported, never invented
Check the result below

While processing, your agent runs these five checks

InventoryList every file and archive member, hash each source.inventoried
SemanticsConfirm cells, units, clocks, cycle boundaries and labels.adapter design confirmed
ConversionWrite every source field to the canonical layer.converted
FidelityCompare raw and standard values record by record.canonical_validated
EquivalenceRaw and standard loaders train to the same result.benchmark_verified

Your agent produces these files and follows these rules

What the skill produces

click an item

Rules the skill never skips

click to expand

Identify files by content Magic bytes, complete payloads and EOF are checked. A .pkl may hold several streams; extensions are only a hint.
No cleaning in the fidelity layer No resampling, interpolation, clipping, float32 down-casting or dropped diagnostics. Those belong to a named task view.
Capacities are not interchangeable Health capacity, partial-DoD throughput, RPT capacity, rated capacity and author SOH stay separate. Unclear labels are withheld, not guessed.
EOL is an observed event The last row is not EOL. Sparse RPTs may only bound an interval, and right-censoring is recorded instead of a fake RUL.
No leakage across splits The same physical cell, duplicate copies and overlapping windows stay on one side; scalers are fit on training data only.
Gates advance only with evidence inventoried, converted, canonical_validated and benchmark_verified are separate; none is inferred from the one before.

03Upload status.json to Check Benchmark Readiness

Drop status.json or a dataset folder here Processed_Dataset/<category>/dataset_xx/status.json

Try a real one:

Example dataset_21 needs_semantic_review

Developer console

Explore endpoints as a working platform surface, not a brochure.

Choose a route from the library, inspect its parameters, preview the request, then run a simulated call to see the exact response shape the platform returns.

Live routes0

Formats0

Auth modeToken

Dataset catalog endpoint

Designed for search, filtering, and benchmark preparation. The response returns only datasets that pass the required ETL and quality gates.

Available

GET/v1/datasets?chemistry=LCO&quality=ready

chemistryLCO | LFP | NMC

formatCSV | Parquet

split_readytrue

dataset_id

cell_count

cycle_count

quality_score

response · application/json200 OK

Request preview

import requests

resp = requests.get(
  "https://api.batterylake.org/v1/datasets",
  headers={"Authorization": "Bearer BT_TOKEN"},
  params={"chemistry": "LCO", "quality": "ready"}
)

for dataset in resp.json()["items"]:
  print(dataset["name"], dataset["quality_score"])

Application workbench

Applications are grouped by what the user is trying to do, not by abstract product names.

3 surfaces

BenchmarkExperiment Builder

Pick datasets, split cells, select models, and run comparable SOH/RUL benchmarks.

FeaturesFormula Studio

Upload custom signal formulas and compute cycle-level model inputs from raw curves.

ExplainInterpretation Lab

Inspect feature importance, attention maps, and degradation signals after training.

Open API

Data Download API

Access harmonized, quality-assessed datasets via RESTful endpoints with embedded citations, metadata, and provenance tracking.

RESTful endpoints for dataset listing and download
Parquet, CSV, and JSON export formats
Embedded DOI citations and provenance
Selective cell/cycle range queries

Open API

Benchmark API

Run standardized ML benchmarks across curated datasets with reproducible train/val/test splits and unified evaluation metrics.

SOH estimation and RUL prediction tasks
5+ baseline models (LR, RF, XGBoost, LSTM, Transformer)
Random, temporal, and cross-cell split protocols
RMSE, MAE, MAPE unified metric suite

Coming Soon

Feature Engineering

Build custom feature extraction pipelines with interactive notebooks. Extract dQ/dV features, health indicators, and domain-specific signals for advanced analytics.

dQ/dV peak detection and tracking
Health indicator extraction (IC, DV)
Statistical feature pipeline
EIS parameter fitting

Coming Soon

Model Interpretation

Understand model predictions with explainability tools. Visualize attention maps, feature importance, and identify key battery aging patterns driving predictions.

SHAP feature importance analysis
Attention map visualization
Gradient-based attribution
Integrated gradients for deep models

Platform Roadmap

Data Download API

Open API — harmonized dataset access with provenance and citations.

Available

Benchmark API

Open API — reproducible SOH / RUL experiment execution.

In Progress

Feature Engineering

End-to-end application — dQ/dV, health indicators, formula features.

Planned

Model Interpretation

End-to-end application — SHAP, attention maps, attribution.

Planned

01Describe0 / 12 required fields
02Raw datano link yet
03Submit0 / 5 ready

01

Describe the dataset

Only what the catalog needs and what the processing skill needs to convert your files. Your draft stays in this browser.

Dataset metadata

For the processing skill

02

Provide the raw data

Host the original cycler exports where the team can download them and paste the link. Keep the original file names; do not resample or clean.

Download link to the raw data * Zenodo and Figshare are free for research data and give you a DOI; a OneDrive, Google Drive or institutional share link works too. Include every file of the release (all cells, all batches).

Quick-check a sample optionalDrop one CSV or JSON excerpt (a cell, a few thousand rows). The quality engine runs locally and notes the detected channels for the reviewers. Already ran the processing skill?Add status.json

03

Submit

Opens a GitHub issue prefilled with your metadata, notes, checklist and data link; attach nothing else. The team takes it from there.

Readiness

0 / 5

Open GitHub issue

After submission the team runs the batterylake-processing skill on your files. You get a status.json and a quality report, and the dataset appears in the catalog under your reference name with your DOI and license.

research-ready.

Reproducible experiments, not preprocessing folklore

An auditable gate before any model sees the data

The same platform, programmable

Start with research-ready battery data today.

Benchmark results

Models evaluated · Unranked

Point-value RUL is not available

Run your own benchmark

Battery Digital Twin Studio

How the Studio Works

Studio Workspace

Select Dataset

Feature Identification

Calibrate Battery Digital Twin

Synthetic Data Generation Setup

Synthetic Data Generation Results

Examples

Field Vocabulary

Getting Started

Dataset Registry

Evaluation

Research and engineering contributors

Nanyang Technological University

BatteryLake-Benchmark-DataPrep

Licence

Citation

01Set Up the Processing Skill

Download the skill package

02Choose a Dataset and Send the Prompt to Your Agent

While processing, your agent runs these five checks

Your agent produces these files and follows these rules

03Upload status.json to Check Benchmark Readiness

Platform Roadmap

Describe the dataset

Provide the raw data

Submit