acceptodds
Withdrawn from ICLR 2027

ConsensusCover: Adaptive Input Selection for Neural Operators under Distribution Shift

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Training neural operators requires deciding which candidate partial differential equation (PDE) inputs merit expensive solution labels. Under distribution shift, the best selection geometry depends on both the equation and the operator architecture: target-distribution coverage is stronger in five of six DeepONet/FNO settings, whereas field information is stronger for Darcy/FNO. We introduce ConsensusCover, an adaptive acquisition rule that resolves this conflict without target solutions or equation-specific switching. At each round, ConsensusCover ranks candidate inputs using two independent views: Fourier field modes and the current operator's readout Jacobian. Rank-consistent overlap beyond chance enters the next batch, while the remaining selections preserve target-distribution coverage through local probability transfer. We evaluate Darcy flow, Allen–Cahn and viscous Burgers with DeepONet and FNO. In a separate confirmation with 20 paired trials per setting and 32–256 labelled simulations, ConsensusCover cuts the worst-case gap in out-of-distribution learning-curve error from 55.44% for target coverage and 30.92% for field information to 6.62%. The paired-bootstrap 95% upper bound for this gap is 8.55%, and the geometric-mean gap across all six settings is 2.07%. The agreement signal activates most strongly on Darcy/FNO, showing how information consensus adapts PDE input selection across neural operators.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, scientific machine learning, active learning, input selection, distribution shift, partial differential equations, Fourier neural operator, DeepONet

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 4 reviews, above 0% of the others.
Reviewer_kc4POpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper tries to develop active selection methods for neural operator learning, targeting the setting in which there is significant distribution shift. The paper proposes the ConsensusCover method, in which unsolved candidate inputs are ranked based on how much that input would reduce the variance in a linear weighted Fourier model, and a kernel based on the Jacobian of the current neural operator. It counts how many agree in a top-k for each batch, and incorporates that into the selection. It works well on 5 out of 6 benchmarks, except Darcy / FNO.

Critical strengths
  1. The paper does a reasonable job of explaining the motivation of the problem, in which PDE solvers are costly, so its a good idea to train neural operators with as little labeled data as possible via active learning. Also it does a good job of defining two fairly sensible strategies of picking source inputs that are close to the deployable inputs, and pick inputs that are most informative for learning, and figures out a way of combining the strengths of both.
Critical weaknesses
  1. There is a severe lack of related work, with only 10 citations throughout the entire paper. The authors need to add related work, of which there is plenty (Multi-Resolution Active Learning of Fourier Neural Operators, Active learning for neural PDE solvers, Active learning with selective time-step acquisition for PDEs, On the benefits of active data collection in operator learning, PICore: Physics-informed unsupervised coreset selection for data efficient neural operator training, etc.) to be able to show the positioning and novelty of the work in relation to the prior work.
  2. As a consequence of the severe lack of the related work, the baselines are quite weak.
Questions for authors
  1. How does prior work that leverages the geometry of the input space and the physics-informed nature of the problem to do active selection compare to ConsensusCover?
  2. What about standard active learning approaches, or even a random baseline?
  3. Can the authors explain the drop in performance for Darcy/FNO
Reviewer_YAkNOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper is too badly written to be understood. My best guess is that the authors are proposing an algorithm called ConsensusCover, an adaptive data-acquisition strategy for neural operators under distribution shift. It combines target-distribution coverage with agreement between Fourier-field information and the trained operator.

Critical strengths

The paper identifies a genuine puzzle in neural operators about their usability in OOD scenarios.

Critical weaknesses

The draft is in-general just not comprehensible.

What is the "ConsensusCover" algorithm, exactly? An unambiguous pseudocode is needed and its not available.

The draft is making arguments about "adaptive acquision" : a concept that is neither native to the neural operator training pipeline nor has it been defined.

Questions for authors

The draft needs to be thoroughly rewritten to make a bunch of things clear,

  • what exactly is the pseudocode of the "consensuscover" algorithm?

  • what is the precise mathematical question one is trying to solve?

  • above equation 2 there is a sudden mention of a 16 dimensional POD. Why? Why this specific number is part of the methodology? and just below that there is a mention of "target coordinates" which too make no sense to me.

  • define clearly the concepts of "acquisition", and "coverage".

  • the draft needs to be weeded out of host of undefined terminology : like before equation 8 there are phrases like "complex linear Fourier channel" and "target channel". In the "agreement and fallback" section there is a phase "field and Jacobian ordered batches".

  • things like the Table-1 have columns labeled "coverage", "Field info.", "consensus cover" and "Ratio". I have no idea what these even mean!

None of these make sense!

In general one needs to read the paper carefully to ensure that either every terminology used is precisely defined or referred to somewhere for definition and that it is not used without definition. Certain basic and preliminary sanity checks seem entirely not done, before we can get to the science of the thing.

Reviewer_PHnAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes an adaptive data acquisition strategy for neural operators under distribution shift. The method combines Fourier-field information and model response information through an agreement mechanism. This design is motivated by the observation that different acquisition strategies perform differently across PDEs and architectures. The paper evaluates the strategy on three PDEs using DeepONet and FNO across six settings, with 20 paired trails for each setting.

Critical strengths

The paper highlights a challenge in neural operator learning, where obtaining PDE solution labels is computationally expensive. The observation that different acquisition strategies perform differently across PDEs and model architectures provides a motivation for adaptive sample selection.

Critical weaknesses
  1. In five of the six experimental settings, Target coverage outperforms field information, while ConsensusCover achieves similar performance in these cases. Furthermore, ConsensusCover requires additional computation for multiple information rankings and model-response Jacobians, resulting in longer acquisition time than the fixed strategies. The current results are insufficient to establish whether the additional complexity of ConsensusCover is justified when compared with coverage strategies.
  2. Although the appendix provides extensive ablation studies for the TargetCover component, it does not separately evaluate the contribution of the agreement mechanism. The method assumes that samples selected by multiple information-based rankings contain a greater amount of information. However, the experiments do not directly demonstrate that samples selected via this agreement mechanism are more effective at reducing prediction error than those selected based on a single ranking method. Consequently, it remains unclear whether the observed robustness primarily comes from the proposed agreement mechanism or the coverage fallback.
Questions for authors
  1. The paper briefly mentions existing active learning approaches for neural operators in the related work. Can you clarify how ConsensusCover differs from those prior works and explain which limitations of existing approaches motivate the proposed design?
  2. The main experiments compare ConsensusCover only against two fixed acquisition strategies. Can you provide comparisons with established active learning strategies for neural operators or PDE surrogate learning to better demonstrate the effectiveness of ConsensusCover?
Reviewer_icT2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors present a method for selecting new inputs to solve, such that a target set can be better learned. The method is a heuristic that looks at Fourier modes of potential new inputs, as well as the current model’s Jacobian. The performance of the method is mixed, and does not substantially lower the rollout error on OOD inputs.

Critical strengths

The paper addresses active learning / new input selection, which is not often considered in PDE studies

Critical weaknesses

The paper is very poorly written and is missing many citations (seems to be primarily written by an LLM). Many concepts are introduced and never explained, which leads to significant amounts of technical jargon that accumulate throughout the paper. The baselines and experimental setup are also not entirely clear. A figure describing the main method would be a good idea as well, as well a qualitative figure comparing samples that the method selects, for example. The method’s performance on selecting heldout, OOD inputs is not significantly better than the baseline method of “Coverage”

Questions for authors

AUC is a metric for binary classification. I am struggling to understand why this is used as the main metric in the study? The standard metrics for evaluating neural PDE surrogates is some sort of rollout error (VRMSE/NRMSE), and when using this, the proposed method seems to be very similar to some fixed coverage baseline.

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

Related papers

Discussion (0)

Sign in to comment.