- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
Paper proposes SPFNO, which combines frozen Fourier features with constrained coefficient solves.
Critical strengths
-
Enforcing physical constraints directly through the representation is useful and avoids need to tune penalty weights.
-
Experiments cover several PDEs.
Critical weaknesses
-
Novelty is not clear enough.
-
Well dataset results use a separate trained model based on the Flowers architecture. It is unclear how these results validate the frozen solver or its theoretical guarantees.
-
Proposed learned model uses eight backbone passes, whereas the baselines use one. It makes it difficult to tell how much improvement comes from the structure-preserving components and how much comes from additional compute.
-
Pressure filtering improves conditioning but makes one velocity component substantially worse in Table 4C. I am not convinced about its practical benefit.
-
Well dataset experiments use only one training seed, so the gains are unclear. Table 7 also shows that WaveLiT-9.5M-PT obtains lower median errors with fewer parameters and only one backbone pass.
Questions for authors
-
Could you explain the main methodological novelty??
-
Which parts of frozen solver are actually used in the learned hybrid? Can you separate the gains from physical projection, refinement, and symmetry ensembling while keeping the backbone and training setup fixed?
-
How would baselines perform with the same eight-pass symmetry ensemble?? Consider reporting inference time along with accuracy under a similar experimental setup.
-
Can the pressure-filtering threshold improve stability without substantially hurting velocity accuracy?
-
Can you show tighter error bounds on more realistic problems? How would these bounds help decide when to refine the solution or stop?
-
Could you report results with multiple training seeds?
-
I found the paper difficult to follow in several places, and some important details needs more explanation. I would lean towards rejecting the paper in its current form.
2 comments
I thank the authors for their rebuttal. However, I found both the paper and the rebuttal difficult to follow in several places, and some important details still require clearer explanation. My concerns remain unresolved, so I will retain my original score and continue to lean toward rejection in its current form.
Note for the authors: Please make the rebuttal a little more detailed, update the manuscript accordingly, and address the reviewers’ concerns clearly.
Novelty, gains at equal compute, filtering, bounds, seeds, and presentation
Thank you for the review.
Q1. Null-space solves, Schur complements, divergence-free spaces, and functional majorants are established, and we cite them. We make them work, with stated guarantees, for frozen random Fourier features, which satisfy no boundary or divergence condition and pair velocity with pressure at random. An enveloped stream function makes divergence and trace exact in the continuum while keeping integrals analytic. A whitened filter gives the retained pressure space βh=σr and κ(S)≤τ−2, independent of the coupling scale. Analytic assembly reduces the majorant to one linear solve, and only the conforming representation satisfies its assumptions. In Table R5, with identical features and five seeds, the conforming representation reaches L2(u)=1.45×10−8 at F=256, against 4.82×10−8 for SVD null-space elimination and 1.92×10−5 for the best penalty weight.
Q2. The learned model shares only the projection onto a constraint kernel, applied to the predicted field, so The Well results say nothing about Propositions 1 to 4; the revision gives it its own name and section. Table R1 adds one component at a time to a fixed backbone and recipe, with three seeds. The plain backbone gives one-step VRMSE 0.03182±0.00064. Training with projection gives 0.02974±0.00038 and lowers divergence RMS from 0.2841 to 1.5×10−7 for 0.26 ms per sample. Refinement gives 0.02745 and eight-pass D4 averaging 0.01679±0.00021 (0.01846 without projection). Averaging is the largest gain and the most expensive, 31.85 ms per sample against 4.38 ms.
Q3. Agreed, since any backbone can be averaged over D4. Table R2 gives FNO, CNextU-Net, scOT, Flower, and WaveLiT-9.5M-PT the same eight-pass average plus Πphys and times every row. Their one-step errors fall by 12% to 30% at about 7.8 times the latency. At equal cost we still lead Flower (0.01668 against 0.0174), but at one pass Flower is better (0.0249 against 0.02960). WaveLiT-9.5M-PT beats our ensemble already at one pass (0.0156, 3.15 ms). Table 2 will separate single-pass and ensemble rows, and we withdraw the claim of best mean VRMSE on all five datasets.
Q4. The velocity loss in Table 4C comes from the test basis. With 8 or 16 features and no boundary conformity, the velocity errors are 11.50 and 14.45 before filtering (Table 9), and releasing one divergence condition can move such a velocity far. The threshold is not delicate: every τ from 10−4 to 10−2 gives the same system (Table 18A), and a cut at the largest spectral gap needs no τ. Table R4 uses a boundary-conforming discretization with five seeds. At F=256, filtering at τ=10−4 or at the gap lowers the pressure error from 1.65 (940 with the weak mode) to about 5.5×10−6 and the velocity error from 3.8×10−6 to 2.1×10−6. Only over-filtering at τ=10−2 hurts the velocity.
Q5. The Poisson effectivity of 4.158×103 comes from a deliberately poor split parameter; chosen analytically it gives 4.889. Table R6 adds multiscale and localized Poisson, the Stokes functional majorant for a non-polynomial solution, and Maxwell at κ=10. All 40 checks satisfy M≥E, with effectivity 1.67 to 3.48. These are still manufactured problems on boxes, and general domains need new constants and quadrature. The bound needs no reference solution, and M≤tol implies E≤tol, which a residual cannot give (Table 23B: residual 8.1×10−13, error 0.447). In Table R7, doubling F until M≤tol meets the tolerance in 40 of 40 runs, while stopping on the residual halts one doubling earlier and meets it in 8 of 40.
Q6. Table 2 uses one training seed, and the eight passes share weights. Tables R1 and R3 use three seeds. The one-step standard deviation is at most 0.00064 on active matter and 0.0028 elsewhere. At equal passes we are below Flower on all five datasets, by more than two standard deviations on four; on TRL 2D the gap is 1.6 standard deviations at one pass and 1.4 at eight. Your reading of Table 7 is correct: WaveLiT-9.5M-PT is better in every window with fewer parameters and one pass, and we claim no per-parameter or per-pass advantage.
Q7. We agree. Much of the difficulty comes from two methods sharing one name with interleaved results. The revision separates them, collects every guarantee and its assumptions in one table, states the question and setting at the start of each experiment, marks single-seed results, and adds a notation table. If particular passages were hard to read, please tell us which ones.