FMAE data repository

This repository contains processed data, cross-validation splits and source data for Multitask Battery Management with Flexible Pretraining.

Processed data

data.tar.gz.0, data.tar.gz.1, data.tar.gz.2, data.tar.gz.3 and data.tar.gz.4 are the five parts of a single compressed archive containing the processed data used for model training and evaluation. All five parts are needed to reconstruct the archive. After extraction, the data form the data/ directory used by the code, with the following subdirectories:

Directory	Datasets and use	Corresponding results
data/EV/	Processed EV1–EV6 data for self-supervised pretraining and capacity estimation; EV1, EV2, EV4, EV5 and EV6 also support anomaly detection.	Fig. 3b; EV contributions to the pretraining comparison in Fig. 5a; missing-channel anomaly detection in Fig. 5c.
data/BESS_data/	Processed BESS1 data for system-level capacity estimation.	Fig. 3b and the BESS capacity contribution to Fig. 5a.
data/lab_data/	Processed THU, KIT, MIT1 and MIT2 data for cell-level capacity estimation; KIT, MIT1 and MIT2 for remaining useful life prediction; MIT1 for internal resistance estimation.	Figs. 3a and 4, and the corresponding laboratory-task contributions to Fig. 5a.
data/NC_relaxation_datasets/	Processed voltage-relaxation data for the voltage-only capacity-estimation comparison.	Fig. 5b.

The mapping above identifies the datasets underlying the reported experiments. Related dataset-level results are also reported in Supplementary Table 2.

dataset_five_fold_splits.xlsx provides the five-fold partition information for the downstream tasks.

Source Data

source_data.xlsx contains numerical data and plotting information for the main figures and the supplementary panels listed below. Worksheet names identify the figure number, panel and quantity, allowing the corresponding source data to be located directly.

Worksheet name or prefix	Corresponding figure
Fig1a_*	Fig. 1a
Fig2a_b_*	Fig. 2a,b
fig3a_*	Fig. 3a
fig3b_*	Fig. 3b
fig4_*	Fig. 4
fig5a_*	Fig. 5a
fig5b_*	Fig. 5b
fig5c_*	Fig. 5c
SI_fig3_NE_reproduced	Supplementary Fig. 3
SI_fig4_*	Supplementary Fig. 4

The data_plot_script.zip also provides source data and scripts to reproduce figures.

Raw data

The raw industry-provided EV and BESS datasets are not shared because of industrial data confidentiality requirements. Their processed versions used in the experiments are included in the multipart archive described above.

The raw laboratory datasets are described in the following publications.

1. THU: Lu, Y. et al. Towards real-world state of health estimation, Part 1: Cell-level method using lithium-ion battery laboratory data. eTransportation 21, 100338 (2024). https://doi.org/10.1016/j.etran.2024.100338.
2. KIT and voltage-relaxation data: Zhu, J. et al. Data-driven capacity estimation of commercial lithium-ion batteries from voltage relaxation. Nature Communications 13, 2261 (2022). https://doi.org/10.1038/s41467-022-29837-w.
3. MIT1: Severson, K. A. et al. Data-driven prediction of battery cycle life before capacity degradation. Nature Energy 4, 383–391 (2019). https://doi.org/10.1038/s41560-019-0356-8.
4. MIT2: Attia, P. M. et al. Closed-loop optimization of fast-charging protocols for batteries with machine learning. Nature 578, 397–402 (2020). https://doi.org/10.1038/s41586-020-1994-5.