A fuller battery history can improve estimates—but not predict the future

The value of looking back

A charge–discharge curve offers a snapshot of a battery during one cycle. It can reveal a great deal, but it does not by itself describe how the cell has changed across its recorded life. That distinction matters when researchers try to compare archives assembled under different protocols and operating conditions.

In a preprint, Jing Wang, Shiyuan Piao and Fugee Tsung propose STEK, a framework that combines an endpoint cycle’s voltage, current and capacity curves with additional structural information. Its intended use is retrospective archive harmonization: estimating attributes of recorded cycles when standardized curves survive but some derived annotations or target-defining raw fields are missing or inconsistent.

The distinction between retrospective and real-time use is essential. STEK’s lifecycle summary can include cycles recorded after the cycle being evaluated. It is therefore not a method for estimating a battery’s state using only information available at that moment, nor is it a forecast of future cycling.

A terrain, a snapshot and a cautious correction

The framework builds a cell-level “terrain” from incremental capacity, or dQ/dV: how charge capacity changes with voltage. Arranged across archived cycles, these profiles summarize changes in curve shape over the cell’s recorded trajectory. A separate endpoint-derived proxy summarizes the evaluated cycle’s curves and their changes.

STEK uses learned gates, informed by descriptor statistics and availability, to combine these structural inputs. It then applies bounded adjustments to the curve representation and a gated prediction correction, retaining an endpoint-curve pathway as a fallback. The gates are quality proxies, not calibrated measures of confidence.

Battery type and optional material-text metadata also condition a training-time consistency constraint across the voltage, current and capacity representations. These metadata can reflect source-level labels rather than verified, cell-specific molecular structures. They do not directly enter the prediction head.

The authors call the constraint Koopman-inspired because it uses a structured linear map in a transformed representation space. But the channels are not successive moments in time: the map is not presented as a physical model of battery dynamics, and the method performs no future-cycle rollout.

What the comparisons show

The evaluation uses MIX-battery, a collection drawn from ten sources. After eligibility filtering, the main experiment includes 472 cells and 2,341 endpoint examples, with cells separated across training, validation and test sets. The evaluated endpoints are cycles 96–100; each example contains one endpoint curve, not 96 cycles of input. Results are reported across three training seeds.

The benchmark’s aggregate score averages normalized errors across seven predefined outputs; lower is better. With a TimesNet backbone, full STEK scored 12.42, compared with 13.29 for endpoint-only TimesNet and 15.08 for endpoint-only iTransformer. These are comparisons of complete settings: STEK has archived and auxiliary context that the endpoint-only models do not. Its advantage cannot be attributed to architecture alone.

A more targeted comparison gave the fusion controls the same auxiliary inputs. Full STEK’s score of 12.42 was 18.4% lower than direct concatenation’s 15.22 and 5.2% lower than quality-unaware feature modulation’s 13.10. The direct-concatenation control was not exactly parameter-matched. Across three tested backbones, aggregate scores improved, although TimesNet’s resistance-proxy error worsened relative to its endpoint-only version.

The contribution of the Koopman-inspired constraint was smaller. Removing it changed the aggregate score from 12.42 to 12.74 while improving the resistance proxy. Comparisons with alternative maps also yielded modest differences with overlapping variability, so the evidence supports an aggregate benefit more cautiously than a universal one.

Knowing when not to learn

The seven outputs are not seven independent hidden battery states. Two are normalized capacity summaries recoverable from a valid endpoint capacity channel; another is simply the complement of state of health. Other outputs depend on historical references or cross-signal information.

For the two directly recoverable summaries, simple label-free rules substantially outperformed STEK once a safeguard handled four endpoints with a missing discharge segment. This is a useful reminder: when the relevant raw information is available, calculation can be preferable to a learned approximation. Conversely, exact reconstruction of the throughput proxy from the model’s resampled input is unavailable because that input omits timestamps.

In separate tests holding out each of three data sources, the full retrospective setting improved aggregate scores over its curve-only counterpart, but not every target improved: resistance-proxy error worsened on one held-out source. When both structural descriptors were masked at test time, the aggregate score rose from 12.42 to 12.71, consistent with a useful fallback. That test addressed missing descriptors, not noisy or misleading ones.

STEK also adds computational cost, and reported prediction-pass calculations exclude offline terrain construction and material-text embedding. The preprint’s strongest case is therefore specific: carefully incorporating archived context can improve aggregate estimation in a heterogeneous retrospective benchmark. Whether the approach works as well with incomplete or corrupted archives—or in a genuinely prospective setting—remains open.