Research map / Large language models / AI and society
BatteryML: An Open-source Platform for Machine Learning on Battery Degradation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Presents BatteryML, an open-source platform unifying battery data representation, feature extraction, label annotation, and benchmark models (classical ML and neural nets) for RUL, SOH, and SOC prediction on public battery degradation datasets.
Problem
Battery degradation modeling lacks a unified data format, standard benchmarks, and tooling, making heterogeneous datasets hard to compare and making it difficult for ML experts and battery researchers to build models.
Approach
A one-stop open-source pipeline with a unified BatteryData representation that converts diverse cycler outputs and public datasets into a common format, plus automatic label annotation (RUL, SOH, SOC), handcrafted and raw-curve features, and a modular set of statistical and neural models benchmarked on standardized splits.
Key results
- For RUL, PCR achieves the lowest error on MATR1 (90 cycles) and SNL (200), while XGBoost is best on CRUSH (330) and Random forest best on MIX (197).
- For SOH, Random Forest achieves the lowest RMSE on MATR (0.53) and SNL (1.80), while linear models are best on CALCE (0.45) and HNEI-level results are competitive.
- For SOC, LightGBM outperforms the other methods on CALCE, HNEI, HUST, MATR, SNL and UL_PUR (e.g., 0.284 RMSE on HUST), and neural networks such as LSTM and GRU perform much worse.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use tree-based models instead of linear models for SOH estimation | BatteryML SOH models · SOH datasets | RMSE | Mixed |
| Use deep learning models instead of traditional methods for SOH estimation | BatteryML SOH models · SOH datasets | RMSE | Hurt |
| Use LightGBM instead of other methodologies for SOC prediction | BatteryML SOC models · SOC datasets | RMSE | Mixed |
| Use linear models instead of deep learning models for SOC prediction | BatteryML SOC models · SOC datasets | RMSE | Helped |
| Fit models on raw QdLinear curves instead of manually crafted features | RUL prediction on CRUSH and MIX · CRUSH and MIX | RMSE | Helped |
| Use XGBoost or Random Forest instead of other models | RUL prediction · CRUSH and MIX | RMSE | Helped |
| Use Gaussian Process instead of PCR or PLSR | RUL prediction · RUL datasets | RMSE | Hurt |
| Use CNN for RUL prediction | BatteryML CNN · MATR1 | RMSE: 76, 67, 64, 74, 60, 82, 65, 79, 367, 78 | Mixed |
Surprising findings
- Neural networks (LSTM, GRU, MLP, CNN, Transformer) generally fail to beat traditional tree-based or linear models on SOH and SOC tasks.
- CNN results on MATR1 are highly sensitive to random seed, with one seed yielding RMSE 367 versus about 60-82 for the others.
Limitations
- No universally optimal model; neural networks are unstable across random seeds.
- SOH and SOC labels are approximations (observed discharge capacity rather than reference performance test results).
- Some public datasets were excluded as unsuitable for RUL estimation.
Experimental setup
Datasets: CALCE HNEI HUST MATR RWTH SNL UL_PUR
Benchmarks: MATR1 MATR2 HUST SNL CLO CRUH CRUSH MIX CALCE HNEI RWTH UL_PUR
Metrics: RMSE mean error across 10 seeds with standard deviation
Baselines: Dummy regressor Variance model Discharge model Full model Ridge regression PCR PLSR Gaussian process XGBoost Random forest MLP CNN
Builds on: BEEP scikit-learn PyTorch XGBoost LightGBM
Closest papers by meaning
- NOCE-Net: Representation Learning for Battery Operational Context via Nested Sequence Modelling (NeurIPS 2026 · similarity 0.87)
- Towards Automatic Power Battery Detection: New Challenge Benchmark Dataset and Baseline (CVPR 2024 · similarity 0.82)
- The Catechol Benchmark: Time-series Solvent Selection Data for Few-shot Machine Learning (NeurIPS 2025 · similarity 0.80)
- Energy-based Automated Model Evaluation (ICLR 2024 · similarity 0.80)
- CheMixHub: Datasets and Benchmarks for Chemical Mixture Property Prediction (NeurIPS 2025 · similarity 0.80)
- Geometric Mixture Models for Electrolyte Conductivity Prediction (NeurIPS 2025 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.