acceptodds
Under review as a conference paper at ICLR 2027

BROC: Block-Recursive Output Correction for Frozen Neural PDE Surrogates

Abstract

Simulation-trained neural PDE surrogates often suffer systematic performance degradation when deployed on measured physical streams or subjected to shifted target dynamics. We introduce Block-Recursive Output Correction (BROC), a residual observer that calibrates a trained and frozen surrogate using matured sparse measurements. Unlike conventional test-time adaptation (TTA) methods that rely on iterative gradient steps, BROC updates a spatially shared temporal component error map via a recursive block, requiring neither backpropagation nor neural parameter updates. Theoretically, we prove the exact optimality of the recursive update and strict temporal non-anticipation. We also establish a finite-sample certificate for full-field improvement when the shared residual signal dominates delay drift, probe error, and measurement noise. Across 20 system–surrogate–checkpoint cells in RealPDEBench, BROC achieves the best accuracy in every cell, reducing MSE by 32.56% relative to the Frozen surrogate and outperforming recent adaptation baselines by 23.84%–29.84%. Under controlled shifts of physical parameters across 12 PDEBench cells, it removes 38.31%–80.47% of Frozen error and outperforms recent baselines by 25.05%–65.87%. Moreover, its analytic calibration core executes in 5.317 ms on RealPDEBench and 4.348 ms on PDEBench, yielding respective 5.02×–11.97× and 7.74×–17.08× speedups over iterative output adapters. Compared with the baseline requiring surrogate re-execution, BROC achieves speedups of 74.90× and 26.75×. Comprehensive stress tests show that BROC remains effective under sparse, dynamic, delayed, and noisy sensing, maintains substantial gains in autoregressive rollouts, supports protective gating, and improves physical fidelity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural PDE Surrogates, Scientific Machine Learning, Online Calibration, Test-Time Adaptation, Sparse Observations

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_AwSBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper propose BROC, which learns a shared affine output correction for frozen neural PDE surrogates using only observations available before each prediction. It updates the correction through block-recursive ridge regression and provides theoretical analyses of temporal causality and risk improvement, along with experiments across multiple systems, models, and sensing conditions.

Critical strengths
  • BROC provides an efficient way to correct frozen PDE surrogates without backpropagation. Results on RealPDEBench and PDEBench show consistent improvements over the frozen models, with additional experiments examining sparse sensing, noise, observation delays, and autoregressive rollouts.
  • The theoretical analysis is closely tied to the observation protocol. The appendix also evaluates the risk certificate on 960960 states, finding a positive lower bound in 321321 cases. This provides useful empirical support for the analysis, although the audit is necessarily post hoc.
Critical weaknesses
  • The contribution relative to standard RLS is not sufficiently isolated. The paper does not clearly establish whether RLS and BROC use comparable input features, joint output representations, and parameter-sharing schemes. Consequently, the performance gap may reflect differences in representation and coupling rather than the recursive update itself. A controlled comparison would help clarify the source of the gains.
  • Some experimental comparisons leave important questions open. The PDEBench results use baseline configurations selected on the RealPDEBench development set, rather than configurations tuned specifically for PDEBench. This evaluates cross-benchmark transfer but does not establish how the methods compare after benchmark-specific tuning. Similarly, the sparse-sensing experiments mainly compare BROC with the frozen model, leaving its relative performance against other adaptation methods under sparse feedback unclear.
Questions for authors
  • Do BROC and RLS use the same input features, output dimensions, and parameter-sharing rules? How much of the performance gap remains under a matched comparison?
  • Could the authors provide PDEBench-specific tuning results and compare BROC with a small number of strong adaptation baselines under sparse sensing, using independent evaluation trajectories? If not, could the conclusions be qualified accordingly?
Reviewer_dUghOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces BROC an iterative, least squares regression update without neural parameter modification to improve a frozen neural PDE surrogate. The experiments across PDEBench and RealPDEBench show that BROC has low latency and higher MSE reduction than other methods including classical baselines and more complex/costly methods such as ADAPT-Z. The authors also show theoretical results proving optimality and a finite sample certificate for their method. The validation of their method is broad across experiments in sparse sensing, dynamic sensing, noise, and more.

Critical strengths
  • The paper compares to a large set of baselines across many method classes and shows strong experimental results across many of them.
  • In addition to the simplicity and speed of BROC, the effectiveness on dynamic sensing and noise experiments are promising for potential real world experiments.
Critical weaknesses
  • They compare BROC to RLS but it's unclear if RLS has the same features, regularization, and observation sequence as BROC. It is important to compare to RLS in a fair way.
  • The theoretical results are primarily known from standard closed-form ridge-regression theory. Theorem 7 is more novel, but it requires residual statistics that are often unavailable at inference time.
Questions for authors

Why does BROC outperform RLS? Since BROC is essentially doing recursive ridge regression, shouldn't RLS have similar MSE reduction performance? I may have missed this in the paper, but are the inputs to the RLS baseline different than BROC? If it's the case that RLS has different inputs than BROC, then I think a feature matched RLS experiment is necessary.

Reviewer_cM9DOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BROC calibrates frozen neural PDE surrogates using a spatially shared affine map from local forecast-block values to residual corrections. Delayed sparse measurements update exponentially weighted sufficient statistics, followed by a regularized linear solve. The correction couples forecast leads and physical components without neural-parameter updates. The paper proves equivalence to its ridge objective and temporal non-anticipation, and derives a conditional full-field improvement bound. Experiments cover measured RealPDEBench systems and controlled PDEBench shifts, with sensing, delay, noise, rollout, physical-fidelity, and latency analyses.

Critical strengths
  • The paper separates Development and Evaluation trajectories, resets the adaptation state between physical trajectories, issues predictions before mature observations update the state, and reports broad stress tests.
  • The measured-system results and trajectory-level paired bootstrap are stronger evidence than a simulation-only demonstration.
  • A compact corrector delivers substantial practical gains with a transparent objective.
Critical weaknesses
  • The methodological contribution must be distinguished from classical ridge regression. Equations 4–6 implement exponentially weighted multivariate ridge regression on an intercept and forecast-block values. Theorem 1 gives the normal-equation solution, while temporal non-anticipation follows from the stated update schedule. The distinctive contribution therefore needs to be the spatially shared block feature structure and its deployment, rather than a new recursive optimization rule. This is a substantive novelty concern that requires precise positioning.

  • The RLS comparison is not interpretable without its feature map and objective. The RLS baseline's lead/component coupling, intercept, normalization, ridge convention, and feature map are not specified; Table 8 lists tuning parameters only. If it solves the same objective with the same features and settings, its predictions should match BROC up to numerical error. If its model is restricted, the comparison changes the model class as well as the update method. This is the most consequential unresolved comparison issue. Provide the exact objective and explain the source of the reported gap.

  • The improvement bound is not shown to be an online sparse-observation certificate. Theorem 7 and Corollary 3 involve complete clean-field moments, a full-field fitted correction, future drift, and future residual quantities. Appendix A.5 can audit these terms post hoc using clean fields, but that does not establish a certificate computable from deployment probes alone. This is a limitation of the guarantee's operational scope, not evidence that the conditional mathematics is false. The distinction should be prominent in the main contribution.

  • The main sparse-sensing experiments leave the surrogate's input availability unclear. The sequential protocol does not sufficiently specify whether the frozen surrogate receives refreshed dense target states between decisions or how such inputs are obtained under sparse sensing. Measurability relative to a declared input does not establish that the input is available in deployment. Appendix B.5's observation-free rollout is valuable, but does not resolve the information contract of the main comparisons. Explicitly identify every dynamic input and distinguish sparse calibration from fully sparse state acquisition.

Questions for authors
  • Please give the exact RLS feature map and objective, including lead/component coupling, intercept, block weighting, ridge, and forgetting conventions. Is it algebraically equivalent to BROC under matched settings? If not, which modeling restriction explains the difference?
  • What enters the frozen surrogate at each sequential decision under 1% sensing? Are dense measured target states refreshed between blocks, or are all dynamic inputs constructed from prior predictions and available probes?
  • Which terms in Theorem 7 or Corollary 3 can be bounded using deployment observations alone? Is the reported positive-bound count exclusively a post-hoc clean-field audit?
  • Appendix B.1 selects baseline configurations on RealPDEBench and transfers them to PDEBench. Why is the separate PDEBench Development set not used for benchmark-specific baseline selection, and what interpretation is intended for these transferred-configuration comparisons?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.