acceptodds
Under review as a conference paper at ICLR 2027

Breaking the Translation Equivariance Barrier: Coordinate Channels and Bounded Error Propagation in Multi-Physics Neural Operator Cascades

23 views

Abstract

Neural operators such as the Fourier Neural Operator (FNO) achieve remarkable accuracy in approximating solutions to partial differential equations (PDEs), but two fundamental challenges hinder their deployment in safety-critical engineering: (1) a structural approximation barrier arising from FNO's translation equivariance that prevents learning Dirichlet boundary conditions, and (2) the lack of uncertainty bounds when multiple operators are chained in multi-physics cascades. We address both. First, we prove that FNO's translation equivariance creates a fundamental approximation barrier for PDEs with non-trivial boundary conditions (Theorem 1), and that coordinate channels resolve this (Theorem 2), validated by up to 63x error reduction across five scenarios. Second, we formalize error propagation in operator cascades: we prove that when K neural operators are composed sequentially, each with conformal prediction coverage (1-alpha_k), the cascade coverage degrades due to exchangeability violations between stages (Theorem 3), and show that coordinate channels mitigate this degradation by reducing systematic boundary errors that compound across stages. We validate on a 3D electromagnetic-thermal-structural cascade, where coordinate channels reduce coupling error from 18.5% to 5.0%. For per-stage UQ, we propose Physics- Informed Conformal Prediction (PI-CP), which embeds PDE residuals into the nonconformity score—complementary to recent physics-residual CP (Gopakumar et al., 2025)—and achieves 89-91% coverage across six physics scenarios with spatially adaptive intervals. We further identify a fundamental boundary of diffusion-based UQ: on well-conditioned forward PDEs, the trained FNO's residual is too small for diffusion models to capture meaningful uncertainty (38-53% coverage). An industrial application—fiber-optic gyroscope bias instability prediction—validates the framework with 3.7-8.0x improvement over analytical baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operators, Fourier Neural Operator, Boundary Conditions, Translation Equivariance, Conformal Prediction, Uncertainty Quantification, Multi-Physics Coupling, PDE

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_GzgBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper looks at how neural operators learn boundary conditions and how errors and uncertainty propagate through multi-physics cascades. It argues that translation equivariance limits what FNO can learn when solution operators involve nontrivial Dirichlet boundary conditions, and that adding coordinate channels removes this limitation. The paper also introduces Physics-Informed Conformal Prediction (PI-CP), which uses PDE residuals to build prediction intervals that vary across space, and studies how coverage changes when operators are connected in a cascade. The experiments include several PDE problems, an electromagnetic–thermal–structural cascade, and an industrial regression task.

Critical strengths

1.Adding coordinate channels leads to large reductions in prediction error across several physical scenarios. The ablation results provide useful evidence that this design choice helps in practice. 2.The experiments examine both individual operators and cascaded systems, reporting prediction errors and coverage and comparing several uncertainty quantification methods. This gives a useful picture of performance across different settings.

Critical weaknesses

1.I had trouble connecting the coordinate-channel analysis to the models actually used in the experiments. The paper does not explain the padding or boundary handling, both of which can affect whether the full model is translation equivariant. I also found two parts of the proofs hard to follow: how Theorem 1 turns a boundary discrepancy into an interior L2 error lower bound, and how the appendix for Theorem 2 establishes the stronger claim of approximating any continuous operator. Both arguments move fairly quickly, and I would need more detail to understand why the conclusions hold. 2.The cascade coverage analysis also leaves some questions open. Theorem 3 bounds coverage changes using the distance between input distributions, but I could not tell what assumptions it needs about the conditional output distribution. The authors note that relative L2 error is only a proxy for the TV term, but the link between them needs more explanation. On the experimental side, the thermal and structural stages already have coverage below 90% when run individually. The paper shows the further drop caused by the cascade, but it does not fully explain the initial shortfall or how pooled-voxel calibration works. These missing details make it hard to see how closely the experiments support the theoretical guarantees.

Questions for authors

1.What padding and boundary handling do the experimental FNO models use? Is the full implementation translation equivariant in the sense assumed by the analysis? 2.Could you walk through how Theorem 1 obtains an interior L2 error lower bound from the boundary discrepancy? Could you also give more detail on the universal approximation claim in the appendix for Theorem 2? 3.What assumptions about the conditional output distribution are needed for Theorem 3? How do you derive the connection between the TV term and relative L2 error? 4.How are voxels pooled during calibration, and how is coverage computed? Why does coverage fall below 90% even when the stages are run individually?

2 comments
AuthorsOpenReview ↗

Response: padding note, Thm 1 derivation, Thm 3 retraction, voxel pooling

Point by point.

Q1 (padding): the 2D benchmark models (FNO2d) use zero-padding, 8 cells at non-periodic boundaries, which does break exact translation equivariance; the analysis idealizes the spectral convolution. The 3D cascade models (FNO3d) use padding=0 and are strictly equivariant, so the analysis applies to them directly. Padding contributes only boundary-local position information, not distributed coordinates, and empirically does not resolve the failure: Dirichlet Darcy without coordinate channels sits at 63.4% error. We will add an implementation note to the revision.

Q2: Theorem 1's interior bound: the boundary discrepancy ∥F(k0)−S(k0)∥≥C′∥g−gˉ∥\|F(k_0)-S(k_0)\| \ge C'\|g-\bar{g}\| on ∂Ω\partial\Omega propagates inward through the PDE's regularity estimate; CC depends on the operator and domain constants. We will write this chain out in the appendix. On the Theorem 2 appendix claim: agreed, it moves too fast; as proven it gives necessity only, and we will downgrade it.

Q3: Theorem 3 is wrong as stated; input-marginal TV is insufficient, and counterexamples from other reviews verify this. The revision restates the per-stage gap as αk+dTV(Qk,Pk)\alpha_k + d_{\mathrm{TV}}(Q_k, P_k), QkQ_k the deployment joint (predicted input, true target), with dTV(Qk,Pk)d_{\mathrm{TV}}(Q_k, P_k) bounded by input-marginal TV plus an Lk E∥y^−y∥L_k\,\mathbb{E}\|\hat{y}-y\| mechanism-response term under a TV-Lipschitz condition; relative L2L_2 enters only through conversion to absolute error. Pooled-voxel calibration pools nonconformity scores across voxels and channels; coverage is marginal over the pooled population, not per-voxel. Stage-2's 83.6% below nominal is the heavy-tailed pooled-score artifact noted in Sec. 4.6. Cascade CP will be rerun with predicted fields passed downstream; results are running now.

Based on the current manuscript, the authors’ rebuttal, and the feedback from other reviewers, I believe the paper still requires further revision. I therefore maintain my original rating of Weak Reject.

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

The paper combines coordinate-augmented Fourier neural operators, residual-normalized conformal prediction, and an analysis of uncertainty propagation through a sequential multiphysics cascade. It reports substantial accuracy improvements from coordinates in several boundary-value problems, spatially varying prediction intervals, and coverage degradation in an electromagnetic-to-thermal-to-structural pipeline.

Critical strengths
  1. The paper addresses practically relevant questions concerning boundary information, error accumulation in composed neural operators, and uncertainty calibration.
  2. The experiments span several physical settings and include informative coordinate ablations and cases where residual-based uncertainty scaling is less effective.
Critical weaknesses
  1. The cascade coverage bound in Theorem 3 does not hold under the stated assumptions. The proof uses the TV distance between input marginals to control downstream coverage, but replacing true upstream outputs with predictions can also change the conditional distribution of the downstream target. For example, let X∼Uniform[1,2]X\sim\mathrm{Uniform}[1,2], let both true stages be identity maps, and let f1(x)=3−xf_1(x)=3-x, f2(z)=zf_2(z)=z. Stage 1 admits standard split-conformal coverage, while stage 2 has zero calibration errors and C2(z)={z}C_2(z)=\{z\}, giving standalone coverage one. The predicted and true inputs to stage 2 have identical distributions, so their TV distance is zero. Nevertheless, the final interval {3−X}\{3-X\} covers the true output XX with probability zero, contradicting the claimed lower bound of 0.80.8 when both significance levels are 0.10.1. This affects a central theoretical contribution and requires more than a clarification of the proof.

  2. The claimed connection between prediction error and coverage degradation is not established. Remark 2 asserts that densities bounded above and below suffice to control TV distance by relative prediction error. However, on [1,2][1,2], consider the uniform density and qn(t)=1+asin⁡(2πn(t−1))q_n(t)=1+a\sin(2\pi n(t-1)), with fixed 0<a<10<a<1. Both densities have uniform positive lower and upper bounds. A quantile transport moves points by at most a/(πn)a/(\pi n), while the TV distance remains a/πa/\pi. Thus prediction error can approach zero without TV approaching zero. Consequently, the proposed implication from improved coordinate-aware predictions to improved coverage guarantees is unsupported, and replacing TV terms with empirical relative errors does not yield a certified numerical lower bound.

Questions for authors

Could you clarify how Theorem 3 accounts for changes in the downstream target’s conditional distribution when true upstream outputs are replaced by predictions? In particular, does an additional assumption exclude the equal-input-marginals counterexample described above, or would the bound need to use the joint input–target distribution?

2 comments
AuthorsOpenReview ↗

Response: counterexamples verified — Theorem 3 retraction and repair

We agree with both points, and both constructions verify.

Theorem 3 as stated is wrong. In your example (XX uniform on [1,2][1,2], identity true stages, f1(x)=3−xf_1(x)=3-x), the predicted and true stage-2 inputs have identical marginals, dTV=0d_{\mathrm{TV}}=0, yet the final interval covers the target with probability zero. Input-marginal TV cannot control downstream coverage because replacing true outputs by predictions can change the joint (input, target) relationship. We will retract the theorem.

The repair we will defend: per stage, coverage gap ≤αk+dTV(Qk,Pk)\le \alpha_k + d_{\mathrm{TV}}(Q_k, P_k), where PkP_k is the calibration joint (true input, target) and QkQ_k is the deployment joint (predicted input, true target). We then bound, up to a constant factor,

dTV(Qk,Pk)≤dTV(input marginals)+Lk E∥y^−y∥d_{\mathrm{TV}}(Q_k, P_k) \le d_{\mathrm{TV}}(\text{input marginals}) + L_k\,\mathbb{E}\|\hat{y}-y\|,

where LkL_k comes from a TV-Lipschitz condition on the downstream conditional mechanism (dTV(PY∣a,PY∣b)≤Lk∥a−b∥d_{\mathrm{TV}}(P_{Y|a}, P_{Y|b}) \le L_k\|a-b\|). Proof: a triangle step through the intermediate law in which the mechanism responds to the predicted input, plus the standard joint-TV decomposition; the full proof will go in the appendix. Two consequences worth stating plainly. First, in your example E∥3−2X∥≈1\mathbb{E}\|3-2X\| \approx 1, so the corrected bound is vacuous exactly where coverage fails, which is the honest outcome. Second, your sin-perturbation example remains consistent: the input-marginal term is not certified by prediction error; it stays an empirical quantity in the bound, and only the mechanism-response term is certified.

The cascade coverage numbers (90.9 / 83.6 / 79.5%) are measurements under our calibration protocol, not instances of a guarantee. The revision will state this and rerun cascade CP with predicted fields passed downstream.

I thank the authors for their response, but I still think that this work needs significant effort to pass the bar, thus I'll retain my score.

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

The paper investigates structural limitations of Fourier Neural Operators (FNOs) under nontrivial boundary conditions and uncertainty propagation in multi-physics cascades. It proposes coordinate-augmented FNOs and Physics-Informed Conformal Prediction (PI-CP), using PDE residuals to construct spatially adaptive prediction intervals. It presents three main theorems concerning approximation barriers, coordinate-based symmetry breaking, and cascade coverage degradation, supported by PDE benchmarks and a 3D multi-physics application.

Critical strengths

S1. The coordinate-channel ablations demonstrate substantial empirical improvements across multiple PDE settings, while PI-CP provides a practical approach to spatially adaptive uncertainty intervals.

Critical weaknesses

W1. Theorem 1 does not establish a uniform approximation barrier across boundary conditions, while Theorem 2 conflates breaking translation equivariance with establishing the approximation capability of coordinate-augmented FNOs.

W2. Theorem 3 controls distribution shift through marginal total variation without accounting for changes in the joint relationship between predicted intermediate states and downstream targets; moreover, the proposed bound relating total variation to relative prediction error requires stronger assumptions than those stated.

Questions for authors

Q1. Can the authors clarify how the approximation constant in Theorem 1 depends on the boundary data, and provide an approximation result that supports Theorem 2?

Q2. Can the authors establish Theorem 3 under explicit assumptions on the joint deployment distribution and downstream conditional mechanisms, and justify the total-variation bound using relative prediction error?

Q3. Can the authors validate the proposed cascade coverage bound empirically and clarify whether pooled-voxel calibration satisfies the required exchangeability conditions?

2 comments
AuthorsOpenReview ↗

Response: overclaims acknowledged — Thm 1 constants, Thm 2 scope, Thm 3 retraction

Both weaknesses identify overclaims, and we will correct the paper.

W1: Theorem 1's constant CC depends on the PDE operator and domain through elliptic regularity, and on the boundary data through ∥g−gˉ∥L2(∂Ω)\|g-\bar{g}\|_{L^2(\partial\Omega)}; it is not a uniform statement across boundary-condition families, and the revision will say so. Theorem 2 as proven establishes necessity only (the equivariant class excludes the target operator). The appendix's approximation paragraph overreaches; we will downgrade it to the necessity statement plus the empirical evidence (1.2–63.4×, five scenarios).

W2: We agree the theorem is invalid as stated; a counterexample with zero input-marginal TV and zero coverage verifies this. The revision restates the per-stage gap as αk+dTV(Qk,Pk)\alpha_k + d_{\mathrm{TV}}(Q_k, P_k), QkQ_k being the deployment joint (predicted input, true target), and bounds

dTV(Qk,Pk)≤dTV(input marginals)+Lk E∥y^−y∥d_{\mathrm{TV}}(Q_k, P_k) \le d_{\mathrm{TV}}(\text{input marginals}) + L_k\,\mathbb{E}\|\hat{y}-y\|

under an explicit TV-Lipschitz condition on the downstream conditional mechanism. Prediction error enters through the mechanism-response term; the marginal term remains an empirical quantity. Full proof in the appendix.

Q3: Pooled-voxel calibration pools nonconformity scores across voxels and channels; the guarantee is marginal over the pooled population, not per-voxel. Stage-2's 83.6% below nominal is a heavy-tailed pooled-score artifact, noted in Sec. 4.6. Once the corrected bound is in place, we will compare it against the measured cascade coverage.

Thank you to the authors for the clarifications and proposed corrections. However, my primary concern remains Theorem 3, which the authors acknowledge is invalid as originally stated. The revised bound requires additional assumptions, and its proof and empirical validation remain incomplete. Given the central role of this theorem, I would like to maintain my original score.

Reviewer_X94gOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper explores coordinate inputs and uncertainty quantification for neural PDE surrogates. It argues that translation-equivariant FNOs cannot represent certain fixed-boundary solution operators and that coordinate channels remove this obstruction. Experiments report coordinate ablations, comparisons with several uncertainty methods, a 3D multiphysics cascade, and negative results for residual diffusion uncertainty estimates.

Critical strengths

One strength of this paper is that it addresses a useful engineering problem: neural surrogates must account for fixed boundaries and errors propagated between models. The coordinate ablations across multiple scenarios provide evidence that position information matters for the tested architectures.

Critical weaknesses
  • The authors should verify the cascade guarantee, as it is not valid under the stated assumptions. Appendix A controls changes in an input-output coverage event using only input-marginal total variation. Upstream prediction can change the conditional target relationship even when the input marginal is unchanged. Bounded positive densities also do not justify Remark 2's linear bound from relative prediction error to total variation.

  • The cascade experiments do not match the stated pipeline presented. The supplied scripts inject random perturbations, scaled by upstream ground-truth error, rather than consistently passing predicted fields downstream. Fattening correlated voxel scores for calibration does not establish the exchangeability required by the advertised finite-sample guarantee. These issues weaken both the cascade validation and attribution of its improvement.

Questions for authors
  1. How does Theorem 3 handle a symmetric input XX, true identity stages, and predictors f1(X)=−Xf_1(X)=-X, f2(z)=zf_2(z)=z? Input-marginal TV is zero, but an exact second-stage interval misses the original target almost surely.
  2. What additional assumptions justify the TV-to-relative-error inequality? Upper and lower density bounds alone do not control oscillatory density differences.
  3. As mentioned in the weaknesses, the cascade experiments do not match the stated pipeline. The supplied scripts inject random perturbations, scaled by upstream ground-truth error, rather than consistently passing predicted fields downstream. This needs clarification.
1 comment
AuthorsOpenReview ↗

Response: pipeline mismatch confirmed — corrected script and theorem repair

You read the code correctly: the CP-cascade script injects error-scaled multiplicative noise around ground truth instead of passing predicted fields downstream. That does not match the stated sequential pipeline, and we will fix it. Our audit of both scripts found two further issues, which we state here upfront: (1) in the coupling-error experiment (Table 3: 18.5% → 5.0%), only the thermal→structural stage passes predicted fields downstream (structural coupled 6.67% vs standalone 2.41%); the EM→thermal stage perturbs the thermal source with error-scaled noise, because in our data generator the thermal source is drawn independently of the EM field (the EM→thermal coupling is a modeling assumption the generator does not realize, and the revision will align the text). (2) The injection script's coverage numbers are a stress test under error-level-matched perturbations, not a validation of the sequential pipeline. A corrected script (run_genuine.py) is in the anonymous repository: it treats stage 2 as an explicit no-coupling control and passes predicted fields downstream at the genuinely coupled stage with unit-correct normalization; the revision will report its results.

W1, Q1, Q2: we agree. Theorem 3 and Remark 2 are wrong as stated, and your symmetric example verifies. The revision retracts both. The replacement states the per-stage coverage gap as αk+dTV(Qk,Pk)\alpha_k + d_{\mathrm{TV}}(Q_k, P_k), QkQ_k being the deployment joint (predicted input, true target), with

dTV(Qk,Pk)≤dTV(input marginals)+Lk E∥y^−y∥d_{\mathrm{TV}}(Q_k, P_k) \le d_{\mathrm{TV}}(\text{input marginals}) + L_k\,\mathbb{E}\|\hat{y}-y\|,

the second term from a TV-Lipschitz downstream mechanism (dTV(PY∣a,PY∣b)≤Lk∥a−b∥d_{\mathrm{TV}}(P_{Y|a}, P_{Y|b}) \le L_k\|a-b\|). In your example E∥X−f1(X)∥≈1\mathbb{E}\|X - f_1(X)\| \approx 1, so the corrected bound is vacuous precisely where coverage fails. Absolute prediction error, not marginal TV, carries the error dependence.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised PDF: corrected Theorem 3, honest cascade protocol, change list

We have uploaded a revised PDF addressing the discussion-phase feedback. Changes, per reviewer:

Theorem 3 (Mdzz, vLkv, X94g; W1/W2, Q1-Q2). Retracted as stated; your counterexamples verify. The revision restates the per-stage coverage gap as αk+dTV(Qk,Pk)\alpha_k + d_{\mathrm{TV}}(Q_k, P_k), QkQ_k the deployment joint (predicted input, true target), with

dTV(Qk,Pk)≤dTV(input marginals)+2Lk E∥y^−y∥d_{\mathrm{TV}}(Q_k, P_k) \le d_{\mathrm{TV}}(\text{input marginals}) + 2L_k\,\mathbb{E}\|\hat{y}-y\|

under an explicit TV-Lipschitz condition on the downstream conditional mechanism (dTV(PY∣a,PY∣b)≤Lk∥a−b∥d_{\mathrm{TV}}(P_{Y|a}, P_{Y|b}) \le L_k\|a-b\|). Full proof in Appendix F, including the triangle-through-intermediate-law step and an explicit statement of why the submitted input-marginal-TV formulation (and the density-boundedness route of Remark 2) fails, with both counterexamples written out. The old Remark 2 is withdrawn; Remark 1 in the revision covers only the absolute/relative error conversion.

Theorem 2 (vLkv W1, GzgB Q2b). The appendix now claims necessity only (the equivariant class excludes the target operator); the universal-approximation sentence is removed and sufficiency is stated as empirical (1.2–63.4×, five scenarios).

Theorem 1 (vLkv Q1, GzgB Q2). Appendix E now gives a three-step chain: boundary deviation (trace inequality), equivariance pinning FF to the constants, and the interior lower bound via an elliptic Poincaré-type estimate; the constant C=cell(d,Ω,D)C = c_{\mathrm{ell}}(d, \Omega, \mathcal{D}) depends on operator and domain, and the boundary-data dependence enters linearly through ∥g−gˉ∥L2(∂Ω)\|g-\bar{g}\|_{L^2(\partial\Omega)}. The statement is per-input and not uniform across BC families; the main text says so.

Cascade experiments (X94g W2, Q3; GzgB Q3). Confirmed: the CP-cascade script injected error-scaled noise. We audited both scripts and disclose the full picture: only the thermal→structural stage of the coupling-error experiment genuinely passes predicted fields downstream (Table 3: structural coupled 6.67% vs standalone 2.41%); the EM→thermal stage is injection in both experiments, and in the data generator the thermal source QQ is independent of the EM field (the coupling is a modeling narrative the generator does not realize). Sec 4.6, Table 3's caption, and the limitations paragraph now state this. Cascade CP has been rerun with a corrected protocol (run_genuine.py, anonymous repository): stage 2 is an explicit no-coupling control (zero degradation by construction, measured as 0.000, validating the pipeline), and stage 3 passes predicted fields downstream with unit-correct normalization. Genuine passing degrades stage-3 coverage from 84.9% to 26.7% (−0.58), far more than the injected stress test's 79.5%; both protocols are now reported side by side (Appendix N), and the submitted 0.306 bound line is withdrawn with the theorem.

Padding (GzgB Q1). Implementation note added in Sec 4.2: FNO2d benchmarks use zero-padding (8 cells), which breaks exact equivariance; the 3D cascade FNO3d models use padding=0 and are strictly equivariant, so the theory applies to them as stated.

Additional main-text change. The industrial fiber-optic-gyroscope validation is restored to the main text (Sec 4.9) using the additional-page allowance; limitations and conclusion are aligned with the corrected theory.

All numbers in the revision trace to archived experiment outputs (results_v100_genuine.json, results_v100.json) in the anonymous repository.

Dear authors and reviewers,

The author-reviewer discussion period closes on November 18 (AoE). Please engage/continue to engage during this period so that we have our issues discussed thoroughly.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.