acceptodds
Withdrawn from ICLR 2027

Inference-Decoupled Physics Manifold Regularization for Sequential Degradation Prognostics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Integrating domain-specific physical priors into deep sequence models for degradation prognostics is fundamentally challenged by a causal availability paradox: informative thermodynamic and kinetic descriptors are retrospective, post-cycle aggregated quantities that are inaccessible during real-time streaming inference. Directly feeding these privileged descriptors into model inputs introduces non-causal dependencies that prevent online deployment, while unconstrained sequence backbones systematically fail to capture underlying phase-transition dynamics. To resolve this tension, we propose Inference-Decoupled Physics Manifold Regularization (ID-PMR). ID-PMR constructs an auxiliary manifold regression task strictly during training to regularize latent geometric representations via uncertainty-weighted optimization, and completely severs this auxiliary branch during inference—achieving zero runtime computational overhead, zero parameter growth, and zero reliance on retrospective features. Across 76 experimental configurations on four heterogeneous battery benchmarks (NASA, MATR, HUST, XJTU), ID-PMR consistently outperforms unconstrained baselines by up to 78.5% in test RMSE. Closed-loop probing confirms that latent physical interpretability recovers from near-zero to a mean R2 of 0.83. Furthermore, extensive ablations characterize a continuous manifold constraint spectrum and show that source-domain trajectory diversity is strongly associated with out-of-distribution transferability. ID-PMR provides a principled, causally compliant paradigm for infusing high-order physical priors into real-time sequence architectures.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-Informed Machine Learning, Learning Using Privileged Information, Representation Probing, Sequence Modeling, Out-of-Distribution Generalization

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_vvU8OpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a physical manifold regularization and a reweighting technique for the degradation prognostic problem. Tests are done on some battery benchmarks, by comparing the proposed method with iTransformer and causalLSTM.

Critical strengths

The tests are done on some realistic battery benchmarks. It seems that the method is better than iTransformer and causalLSTM.

Critical weaknesses

The paper is very hard to follow for readers without prior knowledge. The novelty is limited, and the comparisons are insufficient.

Questions for authors
  1. Some notations are used without explanations. E.g., (I,V,T) for the raw operational profiles, f_3, k-1, and so on. Some concepts are not defined as well, such as the latent physical variants. The author should do proofreading.
  2. The quality of Fig. 1 and 2 is low. They seem to be be AI-generated: too many elements that are trivial and non-technical. It’s not suitable for a top-tier conference like ICLR.
  3. Many sections are hard to understand for readers without prior knowledge. For example, in Section 3, what are linear probing evaluations, candidate physical descriptor, parameter-minimal linear head, physical target space, key metric f_3 (IC peak area reflecting active lithium loss)? In Section 4, what is the LUPI paradigm (there is no citation)? Can you also explain Eq. (2)? Many concepts, notations, and equations are hard to understand.
  4. The novelty seems to be limited. It seems that the innovation is to introduce an additional physical-manifold-supervised loss and some reweighting techniques. This seems lies in the conventional physics-informed learning.
  5. The baseline methods in the Experiment are limited. Only causallstm and itransformer are used. However, in terms of deep sequence models, there are many categories. The author should consider some of them.
  6. There are no limitation discussion and future work statement.
Reviewer_Psc4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a new battery health prediction approach ID-PMR. Its approach is based on an auxiliary regression of electrochemical descriptors that are selected by mutual information and linear probing, but only during training. The method incorporates uncertainty-weighted multi-task optimization. For the prediction phase, it discards the auxiliary heads. The proposed approach is evaluated on 4 battery datasets and found that the method outperforms the previous ones in terms of prediction accuracy and representation quality.

Critical strengths

This paper talks about a related practical problem, which is how to avoid using retrospective physical information during inference. The method works with various backbone networks and achieves significant improvements without adding any extra overhead during inference.

Critical weaknesses
  1. Essentially, it integrates a bunch of techniques (auxiliary supervision, privileged information, loss-weighting with uncertainty). The so-called 'manifold regularization' is basically just an auxiliary regression loss term without actual explicit manifold constraints. Also, that improvements of directly-supervised descriptors in linear probing does not mean you have really learned the physical dynamics.

  2. There isn't a clear explanation of how some part of the input sequences are created. How you make sure streaming causality is maintained during experiments? In Table 2, you see that All-40 always performs better than your Top-8 selection. But you don't provide a sufficient reason for such a trade-off.

  3. the manuscript seems to be different from the official ICLR 2027 formatting, especially on title, abstract and section headings. Furthermore, the two figures (Fig. 1&2) in your paper are very messy. Text is too small and hardly readable, the elements are arranged chaoticly. Readability is extremely poor. So it's difficult for me to understand your method and experimental results.

Questions for authors

Same as the weaknesses. How much is the paper AI-generated vs. how much you reviewed and edited it?

Reviewer_mJ7iOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper tackles physical blindness in battery SOH forecasting where models lack underlying electrochemical semantics due to the causal paradox. To resolve this, the authors introduce Inference-Decoupled Physics Manifold Regularization (ID-PMR). During training, ID-PMR attaches an auxiliary regression head to regularize latent spaces against Top-8 electrochemical features. During inference, this auxiliary head is cut off but can significantly improve SOH accuracy and latent physical decodability.

Critical strengths

Strength 1: Exploring how to leverage electrochemical physics to improve SOH prediction has practical engineering value.

Critical weaknesses

Weakness 1: Lack of Novelty and Insufficient Comparison. The core architecture of ID-PMR is indistinguishable from existing auxiliary task learning and LUPI frameworks. The authors fail to provide direct comparisons against standard auxiliary supervision baselines, relying solely on unregularized black-box models or other naive baselines. Removing the auxiliary head at deployment is also a standard engineering practice in multi-task learning rather than a distinct methodological breakthrough.

Weakness 2: Excessive Over-Packaging on claims. The manuscript heavily relies on grand terminology such as Physical Blindness and Physics Manifold Regularization without adequate theoretical or empirical backing. The improvement simply reflects supervision alignment rather than the learning of true physical mechanisms. Furthermore, Eq. (6) is a standard auxiliary MSE loss rather than a geometric manifold regularizer. And notions like "state-space coverage regularity" lack quantitative explaination.

Questions for authors

Q1. Compared to the standard paradigm of "training-phase physics feature auxiliary supervision", what fundamental algorithmic distinction does ID-PMR offer? Why is there no direct comparison against standard auxiliary supervision baselines?

Q2. Please provide a rigorous mathematical definition of a physics manifold. Why does Eq. (6), which is simply a standard regression loss, qualify as manifold regularization?

Q3. Why does an R2R^2 close to zero imply a lack of physical information rather than an inability to decode it linearly? Have non-linear probes or control experiments been conducted to rule out non-linear encoding? Or why does increased decodability serve as independent proof of learning real physical mechanisms?

Q4. Is there theoretical justification for using the product of mutual information and linear probing results as the feature selection criterion?

Q5. Given that All-40 outperforms Top-8 in Table 2, have the authors provided a comprehensive computational cost analysis encompassing feature extraction, probing, filtering, and training to justify the claimed efficiency tradeoff of choosing Top-8? Furthermore, why should varying the number of auxiliary supervised features be interpreted as a distinct manifold constraint spectrum rather than just conventional multi-task supervision?

Q6. How do the authors quantitatively define and measure the thermodynamic state-space coverage of the source domain? Why are cross-dataset transfer results sufficient to prove that transfer performance is governed by coverage rather than dataset scale or something else? And is this essentially physical intepretation of OOD?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.