acceptodds
Withdrawn from ICLR 2027

Law Violation Determines How Strongly Physics Should Constrain Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A physical law helps learning only while it matches the observed dynamics; once violated, enforcing it introduces systematic bias. We show that measured law violation determines how strongly physics should constrain a model. Our rule maps an invariance-specific violation to hard constraints, soft penalties, or unconstrained learning using two thresholds calibrated on reference dynamics. Across 12 PDE systems, five conservation or symmetry classes, and nine noise and sample-size regimes, the rule selects a model family before target training. On held-out conservation and mechanism-aligned translation systems, it reduces rollout error by 12.1% relative to a validation-selected fixed CNN. After separate calibration for one Fourier neural operator, it reduces error by 12.8% across 18 conditions and requires one target fit instead of seven. A linear-Gaussian analysis recovers the same three regimes and an inverse-square law for the optimal penalty. Law violation converts constraint choice from a target-specific sweep into a quantitative decision grounded in the observed dynamics.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
physics-informed machine learning, physical constraints, conservation laws, symmetry, model selection, neural operators

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_L3viOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses the cost of choosing how to impose a physical law when learning the dynamics of systems that only approximately respect it. The architectural choice (hard constraint, soft penalty, or no constraint) is usually made experimentally. This paper argues this choice can instead be made before training: first, a scalar statistic measures, from the noisy training trajectories, how badly the law is violated, and two thresholds t1 and t2, computed experimentally, map which constraint we should use. The thresholds are fitted once per invariance class by grid search on a simple calibration system and then reused on other systems in the same class. A simple analysis gives a theoretical motivation for the three-regime structure as a bias–variance trade-off. Experiments on 12 synthetic PDEs with controlled law-breaking show that the selector reduces rollout error by about 12% relative to a fixed configuration while training a single model, but remains roughly 20% less accurate than exhaustive grid search.

Critical strengths

The first merit of the paper is proposing a solution to the aforementioned problem of choosing the right way of enforcing physical laws: instead of relying on a brute force scheme, the paper suggests a rule with a invariance class (per law) that is able to reduce the computational overhead (after calibration). The paper also tries to give a theoretical foundation and a better understanding of the technique presenting it as a bias variance tradeoff. Moreover, they design statistic that are experimentally able to predict which architectural choice we should use before the training itself.

Critical weaknesses
  1. the first weakness is that the central assumption of the paper is that the thresholds should generalize and be invariant but the paper does not justify this. No justification is given for why the thresholds should not be system dependent. The theoretical justification for the existence of 3 phases also shows that the thresholds, in that specific case, do depend on the system that is being analyzed. So what is proving the choice of the 3 phases regime has some fundamental rigor it is also suggesting that the 3 phases are not invariant within systems belonging to the same class.
  2. The evidence is confined to small synthetic experiments built to be measurable and also the generalization of the thresholds, given there is no theoretical proof of the invariance, is not extensively tested as this would need more cases for each class.
Questions for authors

It would be important to validate the paper showing more experimentally (given the absence of a theoretical understanding) how invariant are the classes. Moreover, given that the theory suggest the size of the problem has an impact on the thresholds, one should do a size study for 'all' the cases used in the paper. Moreover, it would also be important to validate this idea on larger experiments where this technique would be of major importance.

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

I'm sorry to be blunt but the paper is quite unreadable. I cannot really tell, after going through it twice, what is the object of study of the paper.

Critical strengths

I cannot say.

Critical weaknesses

The paper is completely AI generated, up to the title. While I understand that this may not be a direct reason to reject (and obviously there is no way of proving it is the case with 100% likelihood), it makes the paper completely unreadable (see below for examples), and it also signals that perhaps not a lot of thought or effort has been put into the shaping of ideas, presentation, contributions or methods developed.

The authors do not really state what problem are they addressing. It is not clear what 'learning' means in their context. Furthermore, sentences like:

  • 'Penalties and partially equivariant pathways interpolate between exact and unconstrained models' (not sure what it means for a pathway to interpolate between models)
  • 'The selector never sees the hidden breaking strength, clean target trajectories, analytic evolution operator, or candidate-model errors on the target system.' (what is the selector? what is the hidden breaking strength?)
  • 'The contribution is a pre-training rule that connects measurable physical-law violation to constraint strength. Direct conservation and symmetry defects transfer across several dynamics, while translation requires a statistic that matches the breaking mechanism.' (A pre-training rule for what? What is constraint strength? What are the models we are trying to train? What does translation stand for here?) (I could continue...)
Questions for authors

If the authors really worked on this purposefully and it represents, in their opinion, a meaningful contribution, then I suggest that they completely re-write the paper to try to make it understandable.

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

Related papers

Discussion (0)

Sign in to comment.