acceptodds
Withdrawn from ICLR 2027

Partition-of-Unity Adapters for Efficient Local Refinement of Implicit PDE Representations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pretrained implicit PDE surrogates and neural operators often produce imperfect predictions on new problem instances, while dense interior target solutions are unavailable for adaptation. Full fine-tuning updates the entire model, and weight-space adapters reduce trainable parameters but do not explicitly localize corrections in the physical domain. We introduce Partition-of-Unity Adapters, a coordinate-space adaptation layer for frozen PDE representations. The method attaches smooth overlapping windows to the domain and equips each window with a low-dimensional Chebyshev correction packet, so every trainable coefficient is tied to a known spatial support and local basis mode. Residual or physical- change signals select local correction packets, with coordinated relaxation or dense activation available for broader or coupled responses. The adapter is fit from PDE residuals, initial and boundary conditions, interface constraints, and optional sparse measurements, without dense interior solution labels during deployment adaptation. Across frozen Fourier-feature MLPs, FNO backbones, localized numerical repair problems, and material and interface changes with nonlocal responses, POU adapters often improve predictions over frozen models and matched global coordinate corrections. Compared with full fine-tuning, they use substantially fewer trainable parameters and shorter adaptation times under the evaluated optimization budgets. Selected local corrections leave the frozen field unchanged outside their supports by construction. The results also identify clear scope boundaries: transport-dominated errors, singularities, and strongly coupled responses can require PDE- aware supports or bases, coordinated relaxation, dense capacity, or full model adaptation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
partition of unity, parameter-efficient adaptation, PDE surrogates, neural operators, physics-informed learning, local refinement

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pWNLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose partition-of-unity (POU) adapters, which are coordinate-space corrections for frozen PDE fields. The method corrects the initial approximations using a superposition of windows of smooth Chebyshev expansions with trainable coefficients that define those expansions. The coefficients are obtained by fitting to the PDE residuals, IC, BC, and sparse measurements. They test their approach in three settings:

a) frozen FNOs pretrained on three PDEs

b) localized and singular errors

c) manufactured material and interface changes

Critical strengths
  • The construction of their approach is simple and backbone-agnostic. It only needs the values and derivatives of the frozen surrogate fields. It does not need access to the actual weights.
  • The correction coefficients are interpretable and we can understand where the approximation was corrected.
  • The method is fast and easy to fit
  • The authors openly discuss negative results.
Critical weaknesses
  • No fully trained backbone is used in any of the FNO experiments. The authors trained their FNO models for only 100 steps in Table 2 and 350 steps in Appendix C.1. I believe that the pretrained models must be trained until convergence before any analysis is performed, while the results for the undertrained models should be presented as ablation studies. It is unclear from the paper whether the pretrained FNOs converged, as the authors do not show the training or validation loss curves. Table 14 indicates that the accuracy advantage of the adapters largely disappears with slightly longer pretraining (350 steps). On Allen–Cahn without sensors, POU even degrades the frozen prediction. Moreover, POU is less accurate than FT in 4/6 cases and equally accurate in the other 2/6.
  • The current set of benchmarks is too simple and insufficient. I suggest that the authors test their approach on more challenging, multiscale benchmarks, such as the Navier–Stokes or Euler datasets in [1, 2, 3].
  • In the Allen–Cahn case with 32 sensors (Figure 3), we observe that the FT error still goes down, so picking the checkpoint at 5600 steps is not fair (the model requires longer training). From Figure 3, it seems to me that the training parameters used for the FT models, such as the learning-rate scheduler, are not optimal (especially for the Allen–Cahn case, where the gain of the adapter over FT is largest).
  • There are too few baselines. I suggest that the authors compare their approach to FT methods for neural operators, such as F-Adapter [4]. Moreover, I would like to see a comparison between LoRA and the proposed method in the main text. I suggest that the authors add LoRA fine-tuning and last-layer fine-tuning of the FNO as two additional baselines in Table 2.
  • In the singular-repair results in Table 22, we observe that the PDE residuals barely change for POU-S, while the sensor RMSE drops significantly. It is unclear whether the physics term in the equation contributes to the fit at all. We could answer this question by running the singular-repair experiments with m = 0.
  • Experiments in sections 4.2 and 4.3 do not use real pretrained neural surrogates. In my understanding of the paper, the goal is to make corrections on pretrained neural operators/networks, not coarse approximations obtained by a numerical solver or hand-crafted solution fields. The authors should include experiments with real neural surrogates in those cases.

[1] Ohana et al., The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning, NeurIPS 2024 (Datasets and Benchmarks).

[2] Takamoto et al., PDEBench: An Extensive Benchmark for Scientific Machine Learning, NeurIPS 2022 (Datasets and Benchmarks).

[3] Herde et al., Poseidon: Efficient Foundation Models for PDEs, NeurIPS 2024.

[4] Zhang et al., F-Adapter: Frequency-Adaptive Parameter-Efficient Fine-Tuning in Scientific Machine Learning, NeurIPS 2025.

Questions for authors

On top of the concerns and questions from above:

  1. Global Cheb. correction in table 2 is more accurate than POU-S in every single case. Where do the authors believe that the local corrections can play a significant role?

  2. The adapter require knowledge of the PDE, IC, BC of the new instance. How does it compare (accuracy and runtime) with directly solving the PDE on the same grid, or with some classical solver initialized with frozen prediction?

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

This paper introduces Partition-of-Unity Adapters, a coordinate-space adaptation framework for improving predictions from frozen PDE surrogates without updating their internal parameters. The method constructs correction fields using smooth, overlapping spatial windows and low-dimensional Chebyshev polynomial bases. Each correction coefficient is associated with an explicit spatial support, enabling localized adaptation and preservation of the original prediction outside selected supports.

The authors investigate three operating modes: POU-S, which activates selected local packets; POU-C, which introduces coordination for broader corrections; and POU-D, which activates the full packet scaffold. The adapters are fitted using PDE residuals, boundary conditions, and optional sparse measurements, without dense target solutions during deployment.

Experiments cover frozen neural operators, localized singular errors, and interface changes. The results demonstrate improvements over frozen predictions and provide analyses of parameter efficiency, spatial preservation, correction capacity, and sensitivity to problem structure.

Critical strengths
  1. interpretable and parameter-efficient adaptation of frozen PDE models.
  2. adaptation using limited observations and physics-based constraints.
Critical weaknesses
  1. The proposed local adapters do not consistently outperform matched global coordinate corrections.
  2. Performance depends strongly on problem specific support geometry, basis selection, and physical coupling.
Questions for authors
  1. identify deployment conditions under which local partitioning provides a clear advantage over a matched global correction
  2. characterize when localized corrections or coordination are likely to worsen solution accuracy
Reviewer_nYbsOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses the problem that corrections to a frozen PDE surrogate on new instances cannot be localized and may degrade regions that were already accurate. It proposes attaching, at the backbone's output, a layer of local Chebyshev correction packets confined by compactly supported partition-of-unity windows, so that each trainable coefficient is bound a priori to a known physical region, is fitted only from physics residuals and sparse measurements, and leaves the field exactly unchanged outside its support. For this claim, the paper provides only constructive derivations (linear parameterization, normalization, preservation implied by compact support, linearized least-squares fitting), with no error bounds or convergence analysis. The experiments cover three settings—broad errors, localized singular errors, and nonlocal material responses—against a budget-matched global correction and validation-selected full fine-tuning, supplemented by several ablations. The evidence is solid in the localized-singular setting, shows the method is less accurate than the baselines in the broad-error setting, and its main value lies in clearly delineating the method's scope of applicability.

Critical strengths
  1. The paper is clearly structured and easy to follow.
  2. The paper supports its claims with a large body of hierarchically organized experiments and analyses, making a significant contribution to the development of the field.
Critical weaknesses
  1. The paper should provide a comparative analysis against partial fine-tuning methods, at least covering how the trade-off between cost (complexity, compute time, etc.) and interpretability is handled.

  2. The main experiments are misaligned with the central claim. The "local refinement" in the title is never actually tested in the main experiments; the only evidence that genuinely supports it is the three Poisson singularity problems in §4.2 (each with 4 resolutions × 3 seeds) and the support-radius ablation in C.6 (where, at rp=hr_p = h, the method still outperforms the frozen model and the global control but falls short of FT).

Questions for authors

Based on the evidence presented, in the broad-error regime the method trades accuracy for a form of "localizability" that carries almost no practical information (in Table 2 the selected supports cover ≥82% of the domain, and accuracy is uniformly below the global correction). The regime where the method does show clear value—errors known to be local, with a requirement that the field remain strictly unchanged elsewhere (singularity repair, local material defects)—is precisely the traditional strength of adaptive FEM / PUFEM, yet the paper explicitly declines to compare against them. Could the authors clarify what the intended application regime is? Within that regime, what is the advantage over adaptive FEM / PUFEM? If the advantage lies in compatibility with learned surrogates, is there an experiment demonstrating that this compatibility yields a benefit the classical methods cannot provide?

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

Related papers

Discussion (0)

Sign in to comment.