acceptodds
Under review as a conference paper at ICLR 2027

PHYSICS-QUALIFIED SAFEGUARDS FOR RELIABLE ITERATIVE NEURAL OPERATORS

Abstract

Iterative neural operators can reach a learned fixed point that remains physically invalid, and fast fixed-point acceleration can exacerbate this failure under distribution shift. We introduce PHYSSAFE-IRNO, a reliability wrapper that aligns learned stationarity with physical validity during training and physics-qualifies accelerated proposal at inference. During training, stationarity-conditioned physical-risk supervision discourages physically invalid learned fixed points, while averagedness-oriented regularization shapes the refinement map for safeguarded fixed-point updates. At inference, a physics-qualified SuperMann solver screens numerically admissible fast proposals using a learned–physical bi-residual filter and falls back to separating-halfspace or Krasnosel’skii–Mann updates when needed. Under the explicit assumption that the learned refinement map is exactly averaged, the wrapper inherits SuperMann convergence, under a calibrated residual-envelope condition, sufficiently small learned residual implies finite-tolerance physical validity. Prediction-only temporal residuals avoid dependence on future ground-truth states, and a disjoint physical audit evaluates returned states using separately implemented diagnostics. Across four PDE systems and six held-out shifts, PHYSSAFE-IRNO reduces aggregate error from 0.087 for physics-aware Anderson acceleration and 0.092 for SuperMann to 0.068, cuts the false-fixed-point rate to 1.3%, and raises joint success to 96.8%. The gains over fixed IRNO transfer across FNO, TFNO and CNO at 1.16× the wall-clock cost of fixed-step IRNO. The implementation code is provided in the Supplementary Material.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operators, Physics-Informed Learning, Fixed-Point Iteration, Distribution Shift, Safeguarded Acceleration, Reliable Scientific Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_euPkOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper defines false fixed points in iterative neural-operator refinement, and propose a framework that moves the refiner’s fixed points towards physically valid states during training, and at inference lets physics veto (but never authorise) accelerated steps while keeping SuperMann’s safeguarded convergence path.

Neural operators are used as fast surrogates for PDE solvers. Because of spectral bias, a single forward pass tends to under-resolve fine-scale, high-frequency structure. Iterative refinement methods address this by repeatedly correcting the predicted solution uu. The paper's starting point is IRNO (Liu et al., 2026), in which a frozen pre-trained operator gives a first guess of the solution, followed by iterative refinement that uses a learnt refiner T(u)T(u) in a fixed-point iteration. Iteration is stopped once the update becomes small, i.e. the (learnt) residual Rθ,a(u)=u−T(u)R_{\theta, a}(u)=u-T(u) approaches to zero.

Nevertheless a learnt residual of zero does not mean that the solution in which the iterative schema has converged is physically valid, i.e. it satisfies all the appropriate constraints such as boundary conditions, conservation laws, interior PDE operator constraints. A uu for which the learnt residual is 0 but violates the physics constraints is a false fixed point. The paper distinguishes thus between two types of residuals, the learnt residual Rθ,a(u)R{\theta, a}(u), and a physics residual (operational physical score) that quantifies physics violations/errors. The paper's objective is joint convergence, i.e. the learnt residual and the physical residual should become small together.

The paper trains the refiner to approach in physically consistent solutions, and adjusts the iterative process to check for physical consistence. The iterative process is applied only at inference time.

Training: The refiner is parametrised as Tθ,a(u)=(1−β) u+β Nθ,a(u),β∈(0,1)T_{\theta, a}(u) = (1-\beta)\,u + \beta\, N_{\theta, a}(u), \qquad \beta \in (0, 1); if $ N_{\theta, a}(u)$ is non-expansive then Tθ,a(u)T_{\theta, a}(u) is β\beta -averaged and compatible with the safeguard theory of Krasnosel’skii–Mann. The paper use regularisation (spectral layer norm, Jacobian-vector penalty) to "encourage" the non-expansiveness of $ N_{\theta, a}(u)$. The paper trains the refiner by enforcing physics consistency (through the physics residual) close to stationary points, i.e. where false-fixed points would appear, through the stationarity-conditioned physical-risk loss, LSCPR\mathcal L_{\text{SCPR}}. The end result of this objective is to approach stationary points that are physics compliant. The final loss is a weigthed sum of the original IRNO loss, the SCPR loss, the non-expansive Jacobian regulariser, and a normalized physical score along the trajectory.

Inference: The paper adjusts the iterative schema used at inference time to veto refinment steps that are physically invalid. Each step is screened with SuperMann’s acceptance tests, that look at the learnt residual R; i) blind test: accept the step without evaluating the new point, as long as the current residual has fallen enough relative to a reference value, ii) educated test: evaluate the candidate and accept it if its residual is sufficiently smaller than the current one. A step that passes either test is checked for physical validity. A step that fails both goes to the safeguarded route: backtracking, a halfspace projection, or a KM step. These rules let steps through only while the residual is clearly shrinking, so they can never block convergence and SuperMann’s guarantees still hold.

Evaluation: The paper evaluates the proposed method of four PDEs and compares it against a number of baseline. The vanilla IRNO, adaptive damping (no details are given here); Anderson & Broyden (these two are quasi Newton acceleration methods, applied to the iterative refinement setting ot the paper which can have convergence problems away from good basins); PA-AA, this one is the closest to the method presented in this paper, it changes the accelerator by adding the physics residual in it acceleration least squares problem so that the fast step is chose to reduce both the learnt residual and the physical residual; SuperMan the acceleration method that is also used by PhySafe-IRNO, essentially the PhySafe-IRNO but without the physics compliance check. It is not clear whether these baselines use the retrained refiner (that leads to physics complaint fixed points) or they do not retrain it, I might have missed that in the text.

PhysSafe-IRNO has the lowest error in every PDE, whether in distribution or out of distribution (table 1). In terms of physics realiability it outperforms the baselines (table 2), though the exact definitions of the metrics used there are never given and I have to guess to what exactly they correspond to in terms of the quantities discussed elsewhere in the paper. PhysSafe exhibits also better physical convergence. Experiments also explore how structural properties of the method are empirically satisfied, e.g. violations of the Jacobian constraints.

Critical strengths
  • The method addresses a relevant gap in iterative refinement of PDE solutions, i.e. the false fixed points, fixed points that do not satisfy the physics,
  • The method shows clear empirical gains with respect to the competing baselines.
Critical weaknesses
  • the paper is written very elliptically, a lot of context is assumed known, I cannot say that it has been a pleasure to read. Unless someone is really in the domain it is rather difficult to follow, e.g. many of the constituents of PhysSafe are taken for granted, I was at a loss with different technical terms such as averagedness, halfspace step, SuperMann’s blind and educated conditions, and secant updates to mention just a few; metrics used for evaluation not been (clearly) defined. Sentences are overly compact and hard to decipher.
  • The method that the paper proposes achieves good empirical results, the novelty though is somehow questionable since it brings together a number of basic building blocks from accelaration of iterative refinement methods and couples them with physics constraints satisfiability.
Questions for authors
  • PhysSafe-IRNO uses a retrained refiner trained using physics losses, SCPR, averagedness penalty. Do the baselines use the vanilla IRNO refiner?
  • Some relevant ablations/baselines in which we use a physics aware IRNO variant would be quite relevant:
    • retraining the refiner as $\mathcal L_{\text{IRNO}} + \mathcal L_{\text{phy}} $, to isolate the effect of the physics loss, then
    • LIRNO+Lphy+LSCPR\mathcal L_{\text{IRNO}} + \mathcal L_{\text{phy}} + \mathcal L_{\text{SCPR}}, followed by all losses. these will allow to disect the contribution of the shaping of the refinement operator, from the physics informed filtering at the inference/iterative refinement stage.
  • On the comparison of PhysSafe-IRNO & PA-AA. These two differ in a number of dimensions:
    • in PA-AA the physics steers the fast step computation while PhysSafe-IRNO does not change the step but filters based on physics
    • PhysSafe-IRNO uses the SuperMann safeguard which PA-AA lacks
    • PA-AA does not have a retrained refiner.
    • thus a number of ablations could be done to isolate the effect of each component, e.g.
      • PA-AA + numerical safeguards (gate/halfspace/KM)
      • PA-AA + numerical safeguards (gate/halfspace/KM) + physics screening
Reviewer_BUp1OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces PHYSSAFE-IRNO, a reliability framework for iterative neural operators that addresses false fixed points, where learned convergence does not necessarily imply physical validity, particularly under distribution shifts. The proposed method combines stationarity-conditioned physical-risk supervision, and averagedness-oriented regularization during training with a physics-qualified SuperMann solver that filters accelerated updates using learned and physical residuals while preserving safeguarded fallback mechanisms.

Critical strengths
  1. The paper is well written and easy to follow.

  2. The paper identifies an important reliability issue in iterative neural operators: convergence of the learned refinement map does not necessarily imply physical correctness. This distinction is especially relevant under distribution shifts.

  3. The proposed SCPR loss specifically penalizes physical violations near learned stationary points, rather than applying uniform physics supervision throughout the refinement trajectory. The design distinguishes reshaping the learned fixed-point landscape during training from controlling accelerated trajectories during inference, making the contributions conceptually interpretable.

Critical weaknesses
  1. Theorem 1 requires the learned refinement map to be exactly averaged, but the implemented Jacobian-based regularization only encourages this property without guaranteeing it. Consequently, the theoretical convergence guarantee may not apply to the actual trained models.

  2. The framework combines several established ingredients, including physics-informed losses, Jacobian regularization, SuperMann safeguarding, and multi-objective nondominance filtering. Although the integration is useful, the paper could more clearly explain which component constitutes a fundamental methodological advance beyond existing techniques.

Questions for authors

Theorem 1 requires the refinement operator to be exactly averaged, while the proposed training procedure only approximately enforces nonexpansiveness through spectral normalization and sampled Jacobian penalties. Can the authors provide stronger evidence or certification that the averagedness assumption holds for the trained models? If this assumption is violated, to what extent do the claimed convergence guarantees remain meaningful?

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

This paper addresses the discrepancy between convergence of a learned iterative neural operator and convergence toward a physically valid solution. In particular, a small learned fixed-point residual does not necessarily imply that the predicted state satisfies the governing PDE or other physical constraints.

The authors introduce PHYSSAFE-IRNO, a framework combining reliability-oriented training with safeguarded fixed-point acceleration. During training, a stationarity-conditioned physical-risk loss discourages physically invalid learned fixed points, while Jacobian regularization encourages averagedness of the refinement operator. During inference, accelerated proposals generated using Broyden or Anderson-type directions are screened through numerical safeguards and a learned–physical bi-residual filter, with SuperMann-type fallback updates.

The theoretical analysis provides conditional convergence guarantees under exact averagedness and relates learned stationarity to physical validity under an additional residual-envelope assumption.

The method is evaluated on four PDE systems under six distribution shifts, with comparisons against iterative neural operators, acceleration methods, and physics-aware Anderson acceleration. The experiments demonstrate improvements in prediction accuracy, false-fixed-point rates, and joint convergence success. An independent physical audit and component ablations further support the empirical findings.

Critical strengths
  1. A well-motivated distinction between numerical stationarity and physical validity.

The paper identifies an important limitation of learned fixed-point iterations: convergence to an equilibrium of the learned refinement map does not necessarily imply that this equilibrium satisfies the governing PDE. This distinction is particularly relevant under distribution shift, where learned correction maps may extrapolate poorly.

The proposed separation between reliability-oriented training and safeguarded inference is conceptually interesting. The stationarity-conditioned physical-risk loss aims to improve the physical validity of learned fixed points, while the bi-residual filter prevents accelerated proposals from degrading physical reliability without introducing numerically inadmissible updates.

  1. Convincing experimental validation, including independent physical auditing and meaningful ablations.

The evaluation covers several PDE systems, multiple distribution shifts, different neural operator backbones, and competitive fixed-point acceleration baselines. The reported improvements are substantial: compared with physics-aware Anderson acceleration, the proposed method reduces aggregate relative error from 0.087 to 0.068, decreases the false-fixed-point rate from 4.8% to 1.3%, and increases joint success from 88.7% to 96.8%.

The component ablations help distinguish the contributions of reliability-oriented training, physical qualification, and numerical recovery. Furthermore, the comparison with a matched scalarized physical-merit line search provides evidence that the proposed nondominance filter contributes beyond simply incorporating a physical residual into the acceptance criterion.

The independently implemented physical audit is particularly valuable, as it reduces the risk of assessing physical reliability solely through the same score used during training and inference.

Critical weaknesses
  1. The theoretical guarantees remain conditional on assumptions that are not established for the implemented neural operator.

The main numerical convergence result (Theorem 1) assumes that the refinement map is exactly averaged. However, the proposed implementation only encourages averagedness through spectral normalization, residual scaling, and sampled Jacobian regularization. These mechanisms do not certify global nonexpansiveness of the learned correction map.

Consequently, the main numerical convergence guarantee does not directly apply to the trained neural operator. Moreover, the proof largely relies on existing SuperMann convergence theory, with the additional observation that rejecting numerically admissible accelerated proposals preserves the available safeguarded update classes. While useful, this limits the novelty and practical scope of the theoretical result.

The physical-validity guarantee is similarly conditional. In particular, Assumption 3 requires a residual-envelope relation of the form

Ma(u)≤ca∣Rθ,a(u)∣+εa,δ,M_a(u)\leq c_a|R_{\theta,a}(u)|+\varepsilon_{a,\delta},

together with a PDE stability estimate. Although the authors calibrate this relationship empirically and report favorable coverage under distribution shift, the analysis does not establish that the envelope holds for the actual learned operator across the considered trajectories.

In addition, Proposition 1 requires

caτr+εa,δ≤τmc_a\tau_r+\varepsilon_{a,\delta}\leq\tau_m

to exclude false fixed points at the declared tolerances. It is not clear whether this sufficient condition is satisfied by the experimental configurations reported in Tables 7–8, particularly given the different residual normalizations used for filtering and stopping. Clarifying this point would help determine whether the proposition provides a nonvacuous guarantee for the reported experiments.

Overall, the theoretical results provide useful sufficient conditions, but their applicability to the implemented method remains insufficiently established.

  1. The benefit of safeguarded learned refinement over direct physics-based correction is not sufficiently investigated.

The proposed inference algorithm repeatedly evaluates a physical residual to qualify accelerated steps. However, this information is used primarily to accept or reject proposals generated from the learned fixed-point residual, rather than to directly construct updates that reduce the PDE residual.

It would be interesting to establish whether the proposed mechanism remains advantageous compared with approaches that directly use physical-residual information to correct the predicted solution, for example through residual-based optimization or Gauss–Newton-type updates.

The paper already includes physics-aware Anderson acceleration and a matched physical-merit line search, which are valuable comparisons. However, neither directly investigates the alternative of computing correction directions from the physical residual itself.

Such a comparison, under equivalent computational budgets and accounting for the cost of differentiation, would help clarify the specific advantage of the proposed learned fixed-point refinement framework.

Questions for authors
  1. [Main concern: theoretical applicability] Is it possible to enforce averagedness by construction, or otherwise certify the assumptions required by Theorem 1 for the trained refinement map? If not, what convergence guarantees remain for the implemented algorithm? Could the authors also clarify whether the numerical convergence argument requires additional conditions on the relaxation parameters beyond the ranges stated in Appendix A?

  2. [Main concern: physical-validity guarantee] Could the authors clarify whether the calibrated residual-envelope coefficients and the stopping thresholds used in the experiments satisfy the sufficient condition of Proposition 1? In particular, how should the residual normalizations in Tables 7–8 be interpreted when evaluating this condition? If it is not satisfied, what practical guarantee does the proposition provide for the reported configurations?

  3. [Main concern: comparison with physics-based correction] Since the method already evaluates physical residuals during inference, could the authors compare their approach with a simpler refinement strategy that directly constructs correction directions from the physical residual, for example through gradient-based or Gauss–Newton-type residual minimization? Such a comparison under matched computational budgets would help establish the specific benefit of the proposed safeguarded learned fixed-point approach.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.