Battery diagnostics
Explore saved outputs for all three estimates on a laboratory recording. Reveal the bench measurements and export a diagnostic record.
One core.
Three diagnostics.
Njal turns an electrical recording into estimates of battery recovery, resistance and capacity. The engine runs locally, with no cloud inference.
We adapt a pretrained foundation model to equipment signals, then compare its answers with independent bench measurements.
Put the model
to work.
Open diagnostics
Response to demand
In a separate recorded demonstration, inspect predictions from one low-load recording at four heavier loads. Compare predictions with measured answers.
Measured gains.
Clear comparisons.
The strongest confirmed results are lower recovery and resistance error than the conventional models selected for comparison.
NREL laboratory data · 15 physical battery units withheld from development · Four commercial families · 920 eligible paired tests
Scroll the table to compare all three models.
| Diagnostic | Njal | Unchanged foundation model | Selected conventional model |
|---|---|---|---|
| Heavy-load recovery | 10.65 mV | 21.57 mV | 25.45 mV |
| Heavy-load resistance | 0.330 mΩ | 0.373 mΩ | 0.484 mΩ |
| Capacity / nominal rating | 7.02 pp | 8.14 pp | 8.02 pp |
58% lower recovery error and 32% lower resistance error versus the selected conventional controls. The 95% paired intervals are 26–65% and 23–39%, respectively. Both comparisons pass the prespecified multiple-comparison test. Capacity is secondary; its comparison remains inconclusive.
Conventional controls were selected on development data: support-vector regression for recovery and capacity, and ridge regression for resistance. Errors are averaged within physical units, then equally across families. Repeated measurements are not counted as independent batteries.
What the comparison means.
The model receives a lower-current electrical recording and recorded temperature. Its answers are checked against separate heavier-load and discharge-capacity measurements withheld during inference. Preparation and temperature acquisition take additional time; this is not a claim of a complete diagnosis in 12 seconds.
Njal's observed errors are also lower than the unchanged foundation model in this study. Those comparisons did not pass the full prespecified superiority rule, and the adaptation advantage did not consistently replicate in other conditions. The confirmed claim is against the named conventional controls, not every existing model.
A separate 81-cell study.
A separate PulseBat study evaluated 81 previously unseen cells across three known battery families and three primary load levels. It confirmed 32% lower recovery error and 37% lower resistance error than its selected conventional comparator.
Recovery: 1.284 versus 1.896 mV. Resistance: 0.0467 versus 0.0744 mΩ. Against the strongest unchanged foundation comparator, observed reductions were 3.2% and 2.7%; neither passed the superiority test. This study uses a separate adapted model, not the NREL model above. Its sequential laboratory loads do not represent simultaneous flight conditions.
Additional transfer checks and limitations
Results vary across equipment and conditions.
A separate high-power study on 12 cells found gains over conventional controls at its primary load, while the unchanged foundation model performed better than the adapted model. A subsequent cold/hot evaluation on 11 cells also failed to establish an adaptation advantage.
On 32 additional retired-EV cells, resistance error fell 16.2% against the selected conventional control; recovery did not pass. A separate 60-cell study across five known families confirmed 18.2% lower recovery error and 23.5% lower resistance error against its conventional control. Neither established an additional advantage over the unchanged foundation model. Capacity error increased in the 60-cell study.
These are distinct laboratory studies and model variants. Their scores are not pooled, and their predictions are not presented as coming from one shared deployed core.
From laboratory evidence to equipment.
Today, the working product runs inference locally and lets a user compare its estimates with measured answers. The intended integration is diagnostic software inside equipment makers' test and maintenance tools.
Military hardware, full battery packs, embedded-device performance and maintenance outcomes have not yet been validated. The demo does not estimate flight endurance or clear equipment for use. Equipment partnerships are the next validation step.
Model, data and visual credits
The pretrained core is TabICLv2 from SODA/INRIA. Njal adds domain adaptation, signal processing and the diagnostic workflow. The underlying foundation model was not pretrained by Njal from scratch.
Primary data: Searching for a Pulse, CC BY 4.0, and its research paper. Transfer data: High-power lithium-ion battery characterization, CC BY 4.0; Tao and Ma's PulseBat / TBSI-Lijing dataset and author repository and data license. Data were transformed for these studies; the authors do not endorse Njal.
All equipment and environment geometry is original. Scenes depict the intended application, not customer hardware or military field tests. Type: original Njal Datum lettering and IBM Plex.
This presentation contains aggregate results and selected public-data recordings. Model weights, calibration data, training code and experiment recipes are kept outside the public files.
Built to become
part of the equipment.
The next step is integration with equipment makers: evaluate their signals, test against their measurements, and establish value in maintenance.
Explore the applications