- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper studies early battery lifetime prediction, where the goal is to estimate a battery’s eventual cycle life from a limited number of initial charge–discharge cycles. Its central motivation is that adjacent early cycles can have very similar marginal distributions even when their cross-cycle dependencies evolve with degradation. The authors call this phenomenon stationarity masking. The paper proposes LEAF, a framework that models transitions between local states of adjacent battery cycles rather than relying exclusively on individual-cycle representations.
The theoretical motivation combines the KL chain rule, which separates marginal and conditional discrepancies, with a projection onto a fixed-marginal feasible set. The authors evaluate LEAF on 16 battery datasets spanning Li-ion, Na-ion, and Zn-ion batteries, comparing it with eight baselines. They report improvements in mean absolute percentage error (MAPE) and 15%-Acc across four battery domains, alongside ablation, transferability, and computational-efficiency studies.
Critical strengths
S1: A useful perspective on weak early-stage degradation signals
The paper identifies a plausible limitation of marginal-distribution-based descriptions: similar distributions of individual cycles do not imply similar relationships between consecutive cycles. Shifting attention toward conditional transitions is a natural and potentially useful way to capture subtle temporal changes. The KL decomposition provides a clear statistical framing for distinguishing marginal differences from conditional differences. This gives the proposed method a coherent motivation rather than introducing an architecture without a well-defined problem.
S2: Broad empirical evaluation with component ablations
The evaluation covers 16 datasets organized into four battery domains and compares LEAF with eight methods, including general time-series models and battery-specific approaches. The reported domain-level MAPE improves over the strongest baseline by 13.0% for Li-ion, 18.7% for CALB, 19.5% for Na-ion, and 16.3% for Zn-ion. The ablations also show consistent degradation when removing conditional transitions, fixed-marginal projection, transition encoding, or temporal order. These results provide encouraging evidence that the overall design is effective and that its components may contribute complementary information.
S3: A lightweight, interpretable representation
The fixed-marginal construction explicitly separates state composition from cross-cycle dependence. Its resulting transition matrices can be inspected to identify which local state associations are enhanced or suppressed relative to independence. The paper also reports favorable computational costs compared with several more complex baselines. This combination of a relatively simple architecture, inspectable intermediate representations, and competitive predictive performance is a practical strength.
Critical weaknesses
W1: The methodological novelty may be more limited than the theoretical presentation suggests
The KL chain rule and the uniqueness of Euclidean projection onto a nonempty closed convex set are standard mathematical results. Their application here is reasonable, but the paper needs to establish more clearly what is technically new beyond applying these results to adjacent-cycle battery representations. In particular, the fixed-marginal projection guarantees a feasible joint distribution with the specified marginals, but this property alone does not establish that the resulting representation is more informative about degradation or more physically meaningful. The empirical improvements are encouraging, yet the claimed advantage of this particular projection over other ways of normalizing or correcting transitions needs stronger isolation. The key issue is not whether the mathematics is correct, but whether the proposed construction constitutes a sufficiently distinctive methodological contribution for ICLR.
W2: The evidence for “stationarity masking” and its causal role in the gains is incomplete
The paper shows examples in which conditional-transition discrepancies exceed marginal discrepancies. However, these observations do not by themselves establish that stationarity masking is widespread, that the discrepancy reflects degradation rather than noise or operating-condition variation, or that exploiting it explains the predictive improvements. The analysis uses learned local states and conditional representations, so the observed discrepancies may depend on the encoder and its training objective. A stronger demonstration would compare marginal and conditional discrepancies under controlled conditions, quantify uncertainty across batteries, and establish that conditional information adds predictive value beyond comparable representations and model capacity. The ablations support the utility of the components, but they do not fully validate the proposed explanation for why the method works.
W3: Generalization and statistical evidence need more careful qualification
The main results are promising, but the reported standard deviations come from three runs, and the manuscript does not foreground paired significance tests or confidence intervals for the improvements over the strongest baselines. This matters particularly for the smaller datasets, including CALB and Na-ion. The transfer results also show that direct transfer of the pretrained model performs poorly, while adaptation is needed; domain adaptation does not outperform the target-only baseline on CALB. Thus, the results support the usefulness of LEAF under the evaluated settings more strongly than they establish broad transferability without substantial target-domain supervision. The authors should clarify the scope of the generalization claims and provide stronger uncertainty estimates or targeted tests on genuinely unseen conditions.
Questions for authors
Q1. What is the incremental contribution of fixed-marginal projection?
The Euclidean projection preserves the two cycle marginals while adjusting the joint transition matrix. How does LEAF compare with alternative ways of obtaining a joint distribution with fixed marginals, such as entropy-regularized transport, iterative proportional fitting, or simpler normalization schemes? Can the authors show that the particular Euclidean projection is important for prediction rather than merely enforcing marginal consistency?
Q2. How robust is the evidence for stationarity masking?
The paper reports larger conditional than marginal discrepancies across multiple datasets. Are these results robust across random battery selections, different patch counts, and different encoders? Could the authors provide uncertainty estimates and a controlled comparison demonstrating that the conditional discrepancies contain degradation-related information beyond measurement noise and operating-condition differences?
Q3. How statistically reliable are the reported improvements?
Given that results are averaged over three runs, can the authors provide paired comparisons against the strongest baselines, with confidence intervals or appropriate statistical tests? Since multiple prefixes from the same battery share a lifetime label, how sensitive are the results to evaluating and aggregating performance at the battery level rather than treating all prefix predictions as separate evaluation samples?
Q4. Under what conditions does LEAF generalize to unseen batteries and domains?
The transfer experiments show that direct transfer performs poorly and that adaptation is important. Can the authors clarify how much labeled target-domain data is required to obtain the reported gains, and whether LEAF consistently outperforms target-only training under strictly matched data budgets across all target domains? How well does it perform when the target chemistry or operating protocol is genuinely unseen?
As there were no responses I will keep my scores.