acceptodds
Under review as a conference paper at ICLR 2027

Structure-Aware Epistemic Uncertainty Quantification for Neural Operator PDE Surrogates

Abstract

Neural operators (NOs) provide fast, resolution-invariant surrogates for mapping input fields to PDE solution fields, but their predictions can exhibit significant epistemic uncertainty due to finite data, imperfect optimization, and distribution shift. For practical deployment in scientific computing, uncertainty quantification (UQ) must be both computationally efficient and spatially faithful, i.e., uncertainty bands should align with the localized residual structures that matter for downstream risk management. We propose a structure-aware epistemic UQ scheme that exploits the modular anatomy common to modern NOs (lifting–propagation–recovering). Instead of applying unstructured weight perturbations (e.g., naive dropout) across the entire network, we restrict Monte Carlo sampling to a module-aligned subspace by injecting stochasticity only into the lifting module, and treat the learned solver dynamics (propagation and recovery) as deterministic. We instantiate this principle with two lightweight lifting-level perturbations, including channel-wise multiplicative feature dropout and a Gaussian feature perturbation with matched variance, followed by standard calibration to construct uncertainty bands. Experiments on challenging PDE benchmarks (including discontinuous-coefficient Darcy flow and geometry-shifted 3D car CFD surrogates) demonstrate that the proposed structure-aware design yields more reliable coverage, tighter bands, and improved residual–uncertainty alignment compared with common baselines, while remaining practical in runtime.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operator, Uncertainty Quantification, Partial Differential Equations

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_1bCAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes an UQ-scheme that focuses on the lifting/encoding module rather than the full network. After training a network, at inference time, feature channels coming out of the first layer are randomly dropped or rescaled, injecting noise in the pipeline, which when repeated many times, produces mean and std. The latter can be used for determining an error band in uncertainty quantization. On two benchmark/model combinations this approach indicates better coverage and narrower bands than alternative existing methods.

Critical strengths

I believe it is a clever find to split the analysis/noise injection over different modules, realizing that the type of uncertainty that follows from the encoding/lifting phase can be worth focusing on more instead of a more generic randomness injection on the full network.

The paper is well written and the method clearly explained.

Critical weaknesses

I am slightly uncomfortable with the general direction of the storyline, and related to this, the tone and statements made in the paper. While I believe the core idea of differentiating between different lifting/encoding and propagation/recovering modules of a network when analyzing sensitivity is a valuable insight, I believe the authors aim too much for framing this as a generic new/better UQ-method. In my view the paper could benefit a lot from changing the focus to a more modest research/ablation type that looks into a wider variety of datasets/models/input settings (i.e. varying input resolution, uncertain inputs etc).

The reason for this discomfort is that I am still convinced that in principle the other parts of the network (propagation,recovering) are still perfectly capable of carrying epistemic uncertainty in general. Fundamentally, nothing changed: these parts are a learned imperfect model on finite data, and as such one may be interested in and wanting to analyze its uncertainty. I think what the experiments show is not "this method is a better UQ method", but that on these specific model/datasets we have certain properties like: (1) the main bottleneck in accurate modeling just happens to be the input and its encoding and (2) the epistemic uncertainty of the middle and output layers happens to be correlated to their input-sensitivity.

I don't believe either (1) or (2) needs to hold in general, but it would be interesting to see research focusing on how prevalent this behavior is (in particular (2)). While I don't want to claim exactly to know what is going on and my analysis may be crude/wrong, I would still generally be more interested in an analysis focusing on such questions, instead of trying to plainly show the author's UQ method is better than other methods.

More model/dataset combinations are of course still very welcome, simply because I do believe that it is hard to untangle in general for the results if they say something about the UQ method or about the model/dataset at hand, and if one wants to truly claim a generic better UQ method, id say a lot of very different model/dataset combinations are needed to make that stick.

Questions for authors

I wonder if the authors considered learning the uncertainty/error bands with an auxiliary model, I believe there is a line of works covering this as well that is worth looking into, even if only for comparison purposes.

Reviewer_pN14OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a uncertainty quantification method that requires no retraining. It decomposes a neural operator into three stages: lifting, propagation, and recovering. During inference, channel-wise noise, implemented as either random zeroing or multiplication by Gaussian coefficients, is applied to the output of the lifting stage. By performing T forward passes, the method uses the mean as the prediction and the standard deviation (scaled by a calibration coefficient) to define the uncertainty interval. The authors claim that this approach offers a superior trade-off between coverage and interval width compared to MC Dropout, last-layer Laplace approximation, and input perturbation methods, while achieving performance on par with deep ensembles at a significantly lower computational cost.

Critical strengths
  1. Simple and inexpensive: No retraining is required only T additional forward passes (T = 20–200) are needed, and the method can be applied directly to any network with a three-stage architecture.

  2. Table 3 compares bandwidth at a consistent 99.9% coverage level, which is a sound methodological approach. In 7 out of 8 cases, the bandwidth is no wider than that of a 5-member deep ensemble, yet the training cost is only about one-fifth as high (though the point prediction error remains lower for the ensemble). This serves as the strongest piece of evidence in the entire paper.

Critical weaknesses
  1. The main text claims that the bandwidth is "significantly" narrower compared to a deep ensemble; however, Table 2(A) shows values ​​of 2.73 for Ours(A), 2.03 for Ours(B), and 1.73 for the ensemble. Similarly, in the comparison of coverage rates in Table 3, both entries for FNO-Darcy are 5.54. Neither table supports the stated claim.

  2. The characterization of "epistemic uncertainty / posterior" lacks sufficient substantiation. The paper does not truly approximate p(θS∣D)p(\theta_S | D): the noise intensity is manually tuned, no noise is added during training, and the scale is effectively absorbed by the calibration coefficients. Under a first-order approximation, the bandwidth merely reflects the sensitivity of the output to the scaling of individual channels in the lifting layer. There is no verification of whether the bandwidth narrows as the amount of data increases. Table 4 focuses solely on alignment, with each scenario undergoing independent recalibration. The "geometry-shifted" test set for ShapeNet-Car appears to consist of a standard hold-out set of 100 samples; while the text mentions "altering the geometric features," no specific description is provided.

  3. Some of the underlying assumptions for the provided explanations are invalid. The paper states that the "recovering" step is typically a linear layer, yet the original FNO uses a two-layer MLP for its readout. Furthermore, the equivalence established in Eq. 14 requires the lifting operation to be an affine mapping, whereas the FNO and Transolver implementations in the neuraloperator library utilize MLPs for lifting. While this does not invalidate the method itself, it undermines the proposed explanation for why the method is effective.

Questions for authors
  1. Is the MC dropout baseline trained using the same dropout rate applied during evaluation, or is dropout only introduced at test time?
  2. How does the ShapeNet, Car test set represent a "geometric shift"? The paper mentions that the input has been modified.
Reviewer_EsEkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a structure-aware epistemic uncertainty quantification method for neural operators. The authors decompose neural operators into lifting, propagation, and recovering modules, and argue that epistemic uncertainty should be injected only into the lifting subspace while keeping propagation and recovering deterministic. Based on this idea, the authors design two perturbation strategies: channel-wise multiplicative dropout-style noise and variance-matched Gaussian feature perturbation. Predictive means and uncertainty bands are obtained through Monte Carlo forward passes without retraining.

The experiments are conducted on discontinuous 2D Darcy flow and geometry-shifted 3D ShapeNet car CFD, using FNO and Transolver as the main backbones. Additional results with UNO, MWT, and Airfoil are provided in the appendix, together with several ablation and stress-test experiments.

Critical strengths

The method has a clear intuition and is simple, plug-and-play, and computationally practical. Requiring no retraining and only stochastic forward passes at inference is a meaningful practical advantage.

Critical weaknesses
  1. The novelty is somewhat limited, since the two implementations are essentially feature-space dropout and Gaussian noise, with the main contribution lying in the choice of where to inject stochasticity. The evaluation could also better separate the proposed method from other structured or feature-space UQ approaches. In addition, the use of a very high calibration target, (\tau=0.999), makes it less clear whether the reported advantage would persist under more typical deployment settings.
  2. The justification for lifting-only sampling remains insufficient. The paper argues that lifting prepares initial feature fields while propagation and recovering encode solver dynamics, but epistemic error may also arise from the latter modules. The experiments support lifting-only perturbation empirically, but do not establish when or why restricting uncertainty to this small subspace is theoretically appropriate.
Questions for authors
  1. Could the authors further clarify the connection between feature perturbations and epistemic posterior uncertainty? The paper is motivated from a Bayesian posterior perspective, but the actual stochastic variables are manually chosen dropout or Gaussian perturbations of lifted features. It is therefore unclear whether the resulting variance should be interpreted as epistemic uncertainty or more simply as a useful spatial error indicator.
  2. How sensitive are the conclusions to the calibration scale and the target coverage (τ\tau=0.999)? Would the same ranking hold under stronger distribution shift where the calibration split is no longer representative?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.