HNEI-PakaLi-Lab/synthetic-battery-diagnosis

Synthetic battery diagnosis

Synthetic voltage-capacity curves and degradation labels for graphite/LFP, graphite/NCA and graphite/NMC.

Configurations and splits

Select lfp, nca or nmc. Each has train, test_c1, test_c2, test_c3 and references.

Chemistry	Train	Each test split
LFP	705,651	6,000
NCA	705,643	6,000
NMC	705,382	6,000

Loading

from datasets import load_dataset

repo = "HNEI-PakaLi-Lab/synthetic-battery-diagnosis"
lfp = load_dataset(repo, "lfp")
train = lfp["train"]
test_c1 = lfp["test_c1"]
stream = load_dataset(repo, "nmc", split="train", streaming=True)

For local loading, replace the repository ID with this directory's absolute path. No custom loader or trust_remote_code is needed.

The references split holds the 4 pristine curves of a chemistry, one per cell configuration (configuration = C0, C1, C2, C3, in that order). Use them as the undegraded context of a cell or as the baseline of any comparison against pristine:

• C0 is the pristine training curve: the one row of train with LLI = LAMPE = LAMNE = 0 (curve_id 0), copied unchanged.
• C1, C2 and C3 are the pristine curves of the cells of test_c1, test_c2 and test_c3. They come from V_references.mat of the test-source repository (rows 1 to 3; source_file and source_index_1based say which).

All four have all labels equal to 0, curve_id null (C1-C3) and duty_id_1based/cycle null. They are on the same qnorm grid as every other curve and keep their NaN masks (NaN voltage where the source curve is undefined, e.g. the end of the charge).

Fields and units

• sample_id is the original unique chemistry-scoped record ID. Training retains the chemistry-raw-NNNNNN format. References are chemistry-ref-c0 to chemistry-ref-c3.
• chemistry is lfp, nca or nmc.
• curve_id is the original zero-based raw index within chemistry, and is null in tests.
• configuration is C0 for source raw simulations, or C1/C2/C3 for tests. In references it is the configuration the pristine curve belongs to.
• source_file and source_index_1based trace each record to MATLAB. Raw indices identify the within-file voltage column. Test indices identify the duty-major flattened position.
• duty_id_1based and cycle identify test observations and are null in training.
• lli is loss of lithium inventory (%).
• lampe is loss of active material at the positive electrode (%).
• lamne is loss of active material at the negative electrode (%).
• qloss is capacity loss (%).
• qnorm contains normalized-capacity percentage coordinates, not measured Ah.
• voltage contains volts, paired point-for-point with qnorm.

All signals and labels use float64. Native float32 raw voltages and labels are promoted losslessly. Native float64 tests keep their original precision.

Cell configurations

The train split uses the source C0 cell configuration. The test splits use C1, C2 and C3, as listed in Table 1 of the test-source paper. LR is the negative-to-positive electrode capacity ratio (QNE/QPE). OFS is the relative electrode offset, expressed as a percentage.

Chemistry	Split	Configuration	LR	OFS (%)
LFP	train	C0	0.95	12.5
LFP	test_c1	C1	0.96	11.5
LFP	test_c2	C2	0.94	12.5
LFP	test_c3	C3	0.95	11.5
NCA	train	C0	1.05	1.5
NCA	test_c1	C1	1.06	0.5
NCA	test_c2	C2	1.04	2.5
NCA	test_c3	C3	1.05	0.5
NMC	train	C0	0.90	10.0
NMC	test_c1	C1	0.91	9.0
NMC	test_c2	C2	0.89	11.0
NMC	test_c3	C3	0.90	9.0

These are synthetic cell configurations, not measurements from physical cells. Within each chemistry, C1/C2/C3 share the same ordered labels, duties and cycles (10, 50, 100, 200, 400 and 1,000).

Repository contents

• data/ contains the train, test and references Parquet files.
• LICENSE provides the CC BY 4.0 license.
• metadata.json records source descriptions and simulation parameters. Any reference parameters in this file describe the source simulations, not extra dataset examples.
• provenance.json records pinned source files, checksums and transformations. Its source inventory includes V_references.mat, the input of the references split.
• package_manifest.json records every published file's SHA256, except the manifest itself.

Sources and license

Original training-source files are available from Matthieu Dubarry's Mendeley Data datasets. Each linked release includes its original MATLAB files and documentation.

Chemistry	Original dataset	Test-source files
LFP	Mendeley Data v3	LFP MATLAB files
NCA	Mendeley Data v1	NCA MATLAB files
NMC	Mendeley Data v1	NMC MATLAB files

The simulation-method paper describes the mechanistic approach to generate the data.

Test sets come from the DTW-Li-ion-Diagnosis paper.

The published training curves, tests and labels are released under CC BY 4.0. See LICENSE. This release adapts the source datasets to the documented Parquet format and split structure.