🤖 AI

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

BatteryLake introduces an agentic, physics-grounded data lakehouse that automates the curation of heterogeneous public battery aging datasets into standardized, benchmark-ready assets through LLM-driven metadata extraction, human-in-the-loop verification, and a comprehensive open benchmark of 41 datasets.

Original authors: Tianwen Zhu, Hao Wang, Yonggang Wen

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tianwen Zhu, Hao Wang, Yonggang Wen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of battery research as a massive, chaotic library where every book is written in a different language, uses a different alphabet, and is stored in a different type of box. Some are in CSVs, some in Excel, some in MATLAB files, and others are just PDFs with the data hidden inside paragraphs of text. For years, scientists trying to predict when a battery will die (its "health" or "remaining life") have been stuck trying to manually translate and organize this mess. It's like trying to build a race car when the parts are scattered across a thousand different garages, and no one knows if a "wheel" in one garage is the same as a "tire" in another.

The paper introduces BatteryLake, a new system designed to clean up this mess and turn it into a perfectly organized, ready-to-use racing track. But here's the twist: instead of just hiring a team of humans to sort through the boxes (which is slow and hard to repeat), the authors built a team of AI agents that act like super-strict librarians.

The "No-Guessing" Librarian

The biggest problem the paper argues against is the idea that AI can just "guess" the missing facts. In the past, if an AI saw a battery name and guessed its chemical makeup, it might be right, but it could also be wrong, and that mistake would silently ruin every future experiment.

BatteryLake's agents are programmed with a strict rule: "If you can't point to the exact sentence in the source document that proves it, you must say 'I don't know'."

The Analogy: Imagine a detective who refuses to write a suspect's name on a report unless they can quote the witness's exact words. If the witness didn't say it, the detective leaves the field blank rather than making up a name.
The Result: This prevents the AI from "hallucinating" (making up) facts. If the source doesn't say the battery is made of Lithium Iron Phosphate, the agent writes "not stated" instead of guessing.

The "Human-in-the-Loop" Safety Net

The paper explicitly rules out the idea that AI should work entirely alone. Instead, they use a "selective prediction" strategy.

The Analogy: Think of the AI as a student taking a test. If the student is 99% sure of an answer, they write it down. If they are only 60% sure, they raise their hand and say, "I need a teacher to check this."
How it works: The system automatically accepts the high-confidence answers but routes the shaky ones to human experts. This saves time because humans don't have to check everything, only the parts the AI is unsure about. The paper suggests this balances speed with safety, ensuring that the final data is trustworthy.

The "26-Rule" Bouncer

Once the AI organizes the data, it has to pass through a bouncer at the door of the "Lakehouse" (the final database). This bouncer checks 26 specific rules.

The Analogy: Imagine a bouncer at a club who doesn't just check your ID (schema) but also checks if you are physically capable of dancing (physics).
The Rules: The bouncer checks things like: "Is the voltage within a realistic range?" "Does the battery's capacity get smaller over time, or did it magically grow?" (Batteries shouldn't get stronger as they age). If the data fails even one of these 26 rules—like a battery claiming to have negative voltage or impossible energy efficiency—it gets kicked out. The paper notes that generic data tools often miss these physics-based checks, but BatteryLake catches them.

The Result: A Global Benchmark

After all this cleaning, the team released a benchmark containing 41 curated datasets from over 25 institutions.

The Scale: This collection covers roughly 720 cells and 323,000 charge-discharge cycles, totaling about 9.5 GB of data.
The Variety: It includes batteries of different shapes (cylindrical, pouch, prismatic) and chemistries (LFP, NMC, LCO, NCA), spanning from 2007 to 2026.
The Goal: The paper presents this not as a solved problem for all battery research, but as a standardized "playing field." It offers 8 different baseline models (from simple math to complex AI) and 3 ways to split the data so that researchers can finally compare their results fairly.

What the Paper Doesn't Claim

The authors are careful to note that this system relies on the quality of the original sources. If the original paper or website has a mistake, the AI can't fix it; it can only faithfully copy the mistake (or abstain if it can't find the info). They also mention that running the data conversion happens on the contributor's own computer to respect privacy and licensing, meaning the platform verifies the report of the conversion rather than re-doing the heavy lifting itself.

In short, BatteryLake is a new way to turn a chaotic pile of battery data into a clean, verified, and fair playground for scientists, using AI that knows when to stop guessing and ask for help.

Technical Summary: BatteryLake

Problem Statement
Public battery aging datasets are critical for advancing battery health management (SOH estimation and RUL prediction), yet their practical utility is severely constrained by poor data engineering. Despite the existence of millions of charge-discharge cycles across dozens of repositories, three structural barriers prevent large-scale, reproducible research:

Format and Schema Heterogeneity: Raw data arrives in diverse formats (CSV, MATLAB, HDF5, proprietary cycler exports) with inconsistent units, conflicting sign conventions for current, and dozens of aliases for the same physical measurement.
Metadata Dispersion: Essential experiment-defining facts (chemistry, capacity, protocols, temperature) are rarely embedded in data files but are instead scattered across landing pages, READMEs, and companion papers, requiring manual recovery.
Benchmark Non-Reproducibility: Studies employ different parsing routines, cycle segmentation methods, and train/test splits, making model comparisons difficult to reproduce or directly compare.

While Large Language Models (LLMs) offer potential for automation, their naive application introduces a critical failure mode: the plausible fabrication of metadata (hallucination), which silently corrupts downstream analysis. Existing data integration tools lack the domain-specific physical semantics required to disambiguate battery signals.

Methodology: The BatteryLake Framework
The authors propose BatteryLake, a governed lakehouse and benchmark that automates the curation of heterogeneous battery data through an agentic, evidence-grounded framework. The system operates on three core pillars:

Evidence-Grounded Extraction with Explicit Abstention:
- Two-Stage Provenance-Ranked Extraction: An LLM agent extracts metadata from source pages (e.g., Zenodo, GitHub) and companion papers in order of authority.
- Verbatim Grounding: Every extracted value must be supported by a verbatim quote from the source text. If evidence is unavailable, the agent must abstain (outputting "not stated") rather than impute values based on world knowledge.
- Conflict Resolution: Conflicts between sources are surfaced to human reviewers rather than being silently resolved.
Agentic Converter Synthesis with Validation-Guided Repair:
- Role Classification & Mapping: Agents classify file roles (measurement vs. metadata) and synthesize schema mappings grounded in headers, units, and value statistics.
- Generate-Validate-Repair Loop: The agent synthesizes a dataset-specific converter program. This program is executed locally on the contributor's machine, and its output is checked against a machine-readable validator. If the output fails, the error report is fed back to the agent to repair the converter. This loop repeats up to $k$ times.
- Local Execution: Raw data never leaves the contributor's machine, preserving privacy and licensing constraints.
Selective Human Verification and Physical Validation:
- Selective Prediction: Human review is framed as a selective prediction problem. Decisions with high confidence scores are auto-accepted, while low-confidence extractions are routed to domain experts. The system trades off automatic coverage against residual error risk.
- The 26-Rule Validation Gate: Admitted data must pass a gate covering four dimensions:
  - Validity: Schema compliance and key integrity.
  - Completeness: Coverage of required channels and cycles.
  - Consistency: Monotonic timestamps and sign conventions.
  - Accuracy (Physical Plausibility): Electrochemical constraints, such as voltage bounds, coulombic efficiency ranges ( $95\%-105\%$ ), and non-increasing capacity fade trajectories.

Key Contributions

Problem Formalization: The paper formalizes battery dataset onboarding as evidence-grounded structured extraction with abstention and schema mapping with converter program synthesis. It casts human review as selective prediction with an explicit risk–coverage objective.
Agentic Curation Framework: A novel two-stage extraction algorithm and a generate-validate-repair loop for converter synthesis, governed by a 26-rule validation gate that includes physical plausibility checks.
Governed Lakehouse and Open Benchmark: The release of a canonical four-layer data model (Metadata, Cell, Cycle Summary, Time Series) and a benchmark comprising 41 curated datasets from over 25 institutions.
- Scope: Covers various chemistries (LFP, NMC, LCO, NCA), form factors (18650, 21700, pouch, prismatic), and aging regimes (cycle, drive-cycle, impedance, calendar, thermal abuse).
- Standardization: Includes standardized SOH/RUL tasks, three split protocols (random, temporal, cross-cell), and eight baseline model families (Linear, RF, XGBoost, MLP, CNN, LSTM, Transformer, PINN).

Results and Deliverables
The platform currently hosts 41 registered datasets, with 12 fully curated datasets containing approximately 720 cells and 323,000 charge-discharge cycles (9.5 GB of data). The curation process ensures that every admitted dataset carries machine-checkable curation evidence, aligning with FAIR principles. The benchmark provides reproducible training packages that pin configurations, seeds, and data splits, allowing for fair comparison across the eight baseline model families.

Significance
The paper argues that trustworthy scientific data curation at scale requires coupling LLM flexibility with specific safeguards: evidence grounding with abstention, selective human verification, and executable physical validation. BatteryLake addresses a general pattern in scientific data engineering: converting the "long tail" of small, heterogeneous, under-documented datasets into governed, analysis-ready collections. By replacing trust in either humans or models with an auditable pipeline where residual risk is measured and tunable, the framework offers a template for governing scientific data beyond the battery domain. The authors position BatteryLake not just as a dataset, but as a system that enables reproducible, reliable, and openly shareable battery research.