acceptodds
Withdrawn from ICLR 2027

Cross-Physics Mapping: Learning Operators Between Fundamentally Different Physical Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural operators have demonstrated remarkable success in learning solution mappings within individual physical systems, yet whether they can learn transformations between fields governed by fundamentally different partial differential equations remains largely unexplored. We introduce Cross-Physics Mapping (CPM), an operator-learning framework for translating heterogeneous physical fields through compatible latent representations without assuming equivalence between their governing dynamics. We formulate conditions under which cross-physics operators are well-defined and introduce a dimensionless scaling principle for constructing physically compatible training pairs. To investigate this problem, we generate 1,000 paired diffusion and wave realizations with shared latent geometries, material heterogeneities, and excitation configurations, while restricting network inputs to surface-accessible spatiotemporal observations. We systematically benchmark seven architectures, including six neural operators and a convolutional baseline, in both mapping directions. Our experiments reveal a pronounced directional asymmetry: diffusion-to-wave mapping is substantially more challenging because diffusion attenuates the high-frequency and phase-sensitive information required for reconstructing propagating wavefronts. The U-shaped Neural Operator achieves a relative error of 0.307 and an of 0.905 in this direction, whereas the Galerkin Neural Operator achieves 0.154 and 0.935, respectively, for wave-to-diffusion mapping. These results establish cross-physics operator learning as a feasible task under shared latent structure while revealing that its achievable accuracy depends critically on the information preserved by the source dynamics. CPM provides a foundation for investigating heterogeneous operator learning, physics-aware modality translation, and the identifiability of mappings between distinct physical systems.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-Informed Learning, Diffusion Wave, Cross-Physics Mapping, Virtual Wave Transform

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QkyGOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a new task aiming to learn an operator that maps the observation of one physical process to the observation of another different physical process on the same object, without access to the shared geometry, materials, or excitation. It benchmarks existing operator configurations on 1,000 simulated 2D diffusion–wave pairs with varied defects and excitations.

Critical strengths
  • The task setting is clearly described.
  • Figures 6–10 clearly summarize how the authors test different neural operator models under their proposed settings.
Critical weaknesses
  • The benchmark is narrow in scope.
  • All evaluated models are existing architectures with standard implementations; there is no methodological contribution for the defined cross-physics mapping.
  • Compared to standard tasks of neural operators, the proposed task only changes the input and output datasets (corresponding to different physical quantities). However, as neural operators are data-driven methods, it should be almost the same for training, and the paper is not highly informative in this regard.
Questions for authors
  • Could you construct more practical benchmarks to strengthen the contributions of this paper?
  • For this new task, are there any methodological improvements that could be introduced to better adapt to these specific challenges?
Reviewer_4HB7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper trains neural operators to predict one physical measurement from another, using simulated pairs of heat-diffusion and wave data from the same plate with defects. Only surface observations are used. The paper also proposes a condition for when such a mapping exists and a rule for matching the scales of the two simulations. Seven existing models are trained in both directions, and predicting heat from wave is consistently easier than predicting wave from heat.

Critical strengths

S1. Translating between observations of different physical processes is an interesting problem with a clear connection to nondestructive evaluation, where the same structure is often inspected with several modalities.

S2. The paper is candid about the limits of its experiments, such as single-seed runs, unmatched model sizes, and the different loss used for one configuration, and it interprets diffusion-to-wave predictions as conditional estimates rather than exact recovery.

Critical weaknesses

W1. The experiments do not evaluate the paper's claimed contributions.

W1.1 The paper lists a latent-variable formulation with existence and stability conditions (Proposition 1) and a dimensionless alignment rule as its main contributions (Sec. 1). However, neither is tested in the experiments.

W1.2 As a result, the experimental section (Sec. 4) reduces to a performance comparison of existing neural operators on one synthetic dataset, which does not provide evidence for the proposed framework itself.

W2. The claims are broader than what the experiments show. The title refers to learning operators between "fundamentally different physical dynamics," and Figure 1 presents thermal, THz, optical, acoustic, elastic, and quantum modalities. However, the experiments cover a single pair (diffusion and wave) on one synthetic dataset, so it is unclear whether the findings say anything about cross-physics mapping in general rather than about this specific setting.

W3. The need for direct cross-physics translation is not justified.

W3.1 The motivating application is nondestructive evaluation (Sec. 1), where the goal is to find defects. If the latent configuration (e.g., the defect) can be estimated from the source observation, this goal is already achieved, and predicting the target field becomes secondary. The paper does not explain why the target field itself is needed.

W3.2 The paper does not compare CPM with a natural two-stage alternative: first estimating the latent configuration from the source observation, then simulating the target field. This alternative gives interpretable intermediate results (where the defect is and how large it is) and target predictions that satisfy the target PDE, which the paper acknowledges CPM does not guarantee (Sec. 3.3).

Questions for authors

Q1. (W1) Could you provide experiments that directly test the claimed contributions, namely Proposition 1 and the dimensionless alignment rule?

Q2. (W2) Could you report results on at least one other pair of physical systems, or otherwise narrow the scope of the title and claims to the diffusion–wave setting?

Q3. (W3) What is the intended downstream use of the predicted target field? Could you compare CPM with a two-stage approach (estimating the latent configuration, then simulating the target) on a practical task such as defect localization?

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

This paper studies the learning of a cross-physics mapping for different types of PDEs through a shared latent configuration. It proposes matching characteristic diffusion and wave propagation lengths, and evaluates seven existing model configurations on a synthetic dataset of paired diffusion and wave boundary observations. The reported results show lower relative reconstruction errors for wave-to-diffusion prediction than for the reverse direction.

Critical strengths

The introduction is concise and motivates an interesting question.

Critical weaknesses
  1. The proposed cross-physics operator is not well-defined. Section 3.1 introduces abstract spaces and generic solution and observation operators without fully specifying the admissible latent configurations, observation spaces and norms, or coupling between modality-specific parameters. Such a definition is the foundation of this study on learning such a mapping. The abstract factorization observation does not establish the claimed PDE mapping.

  2. Proposition 1 is missing, R2R^2 is undefined. Appendix A and the conclusion refer to “Proposition 1,” although no numbered proposition is presented in the paper. The evaluation metric R2R^2 is undefined, and the ℓ2\ell_2 metric is not consistent with the notation in Section 4.

  3. The PDE dataset is insufficiently specified for reproducibility and interpretation. Section 3.2 presents generic PDEs without specifying the domain and the coefficients. Appendix C supplies a rectangular domain, initial and boundary conditions, grid dimensions, and time-step restrictions. But important ingredients remain unspecified: the coefficients.

Questions for authors

Please refer to the points in weakness.

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

Related papers

Discussion (0)

Sign in to comment.