- 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:
-
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?
-
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?