acceptodds
Under review as a conference paper at ICLR 2027

Basis-Adaptive Learned Asymptotic Nonlocal Inversion Neural Operator (BALANI-NO) for PDE Surrogates and Gravitational Lensing

Abstract

Spectral neural operators parameterize each layer by a learned Fourier multiplier, or symbol, which is usually a lookup table over the lowest modes. For many elliptic operators the symbol is instead a rational function of frequency. We introduce BALANI-NO, a neural operator whose symbol is the sum of a learned pole (rational) branch in the isotropic magnitude and a small low-mode table. The pole branch specifies the symbol at all frequencies from a few coefficients, favors no coordinate axis, and does not depend on the grid. We compare BALANI-NO with ten baselines at matched parameter budgets, with seven seeds each, on Darcy flow, Navier–Stokes vorticity, and gravitational-lensing deflection fields, where models trained on warm dark matter are tested on cold and axion dark matter. BALANI-NO has the lowest zero-shot error on Darcy ( vs. for FNO), on Navier–Stokes, and on every zero-shot lensing split. It also has the lowest in-distribution error on Navier–Stokes and subhalo lensing, and ties with U-FNO on full-field lensing. Two smaller-scale additional experiments, on real (non-synthetic) cosmological density maps and real geomagnetic survey data, support the same pattern of best or near-best zero-shot transfer.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, Fourier neural operator, spectral methods, operator learning, partial differential equations, gravitational lensing, dark matter, zero-shot generalization, resolution generalization, rational spectral parameterization

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

This paper introduces a novel spectral neural operator—BALANI-NO (Basis-Adaptive Learned Asymptotic Nonlocal Inversion Neural Operator)—designed to tackle operator learning problems for elliptic partial differential equations (PDEs) surrogates and strong gravitational lensing.

Core Architectural Innovation: Traditional spectral operators (such as FNO) typically parameterize the Fourier multiplier (symbol W(k)W(k)) in each layer using a learned lookup table over the lowest modes. However, for many elliptic operators, the symbol is instead a rational function of frequency. BALANI-NO innovatively parameterizes the symbol as the sum of two components: 1.A learned pole (rational) branch in the isotropic magnitude ∣k∣\vert{}k\vert{} (specifying the symbol at all frequencies from a few coefficients, favoring no coordinate axis and remaining independent of the grid); 2.A small low-mode traditional Fourier table (providing a complementary low-frequency correction). Experiments and Evaluation: Under matched parameter budgets and across 7 different random seeds, the authors extensively compared BALANI-NO against 10 mainstream baseline models (including FNO, ALNO, LNO, PDNO, various U-shaped operators, CNNs, and DeepONets).

Critical strengths

1.Strong Physics-Informed Inductive Bias: Tailored for elliptic PDEs and inverse problems with physical backgrounds (such as the lensing operator corresponding to the rational form ik∣k∣−2ik\vert{}k\vert{}^{-2}), the paper bypasses pure black-box lookup tables by incorporating an isotropic rational pole branch. This design naturally aligns with the asymptotic behavior of many physical operators. 2.Exceptional Zero-Shot Cross-Resolution and Cross-Distribution Generalization: The standout highlight of this paper is its remarkable zero-shot transfer performance. Whether dealing with resolution changes (Darcy, Navier-Stokes) or shifts in physical dark matter models (WDM →\rightarrow CDM/Axion), BALANI-NO exhibits extremely low degradation, outperforming most flexible baselines that tend to break down under grid variations.

Critical weaknesses

1.Training-Resolution Performance Is Not Always Optimal: As the authors honestly note in the discussion and Table 7, while BALANI-NO demonstrates dominant performance in zero-shot settings, it is not always the winner at the training resolution (e.g., losing to CNNs and PDNO on Darcy flow). This occurs because its constrained global symbol parameterization sacrifices some single-grid fitting flexibility.

2.Limitations on Non-Input-Invariant / Variable-Coefficient Operators: When the underlying physical operator varies spatially with the input (such as the spatially varying diffusion coefficient a(x)a(x) in Darcy flow), a single global symbol does not exist. Consequently, BALANI-NO's advantages in such tasks lie primarily in generalization rather than training-set fitting.

3.Computational and Implementation Complexity: Although parameter budgets are matched, it remains unclear whether introducing softplus constraints and polynomial combinations for the rational pole branch significantly increases training time or introduces numerical instability (such as higher-order poles collapsing or weak identifiability, as mentioned in the appendix regarding the second pole's larger uncertainty). The paper lacks a direct wall-clock training time comparison.

Questions for authors

1.Clarification Request: In the rational pole branch defined in Eq. (4), pole locations are determined by sj2=softplus(s^j)>0s_j^2 = \text{softplus}(\hat{s}_j) > 0, with a separate learned offset DD introduced at k=0k=0. During training, how sensitive is the choice of initial pole locations (linearly spaced in [−6,3][-6, 3] in the code) to converging to specific physical screening lengths when dealing with different physical tasks (such as the diffusion term in Navier-Stokes versus the rational operator in gravitational lensing)? Do poles ever get stuck in local minima or degenerate (e.g., multiple poles collapsing to the same value)?

2.Clarification Request: Regarding computational efficiency, while BALANI-NO maintains a comparable parameter count to the baselines, how do its memory footprint and computational throughput (during training and inference) compare to FNO or LNO, given that it computes full-frequency rational functions rather than FNO's truncated m×mm \times m window?

3.Suggestion for Future Work: The authors observed in Darcy flow and complex real-world data (like EMAG2) that locally grid-tied models (such as CNN/PDNO) fit better at the training resolution but degrade during cross-resolution transfer. Could future work explore an "adaptive hybrid weight mechanism" that permits more flexible local corrections near low resolutions/training grids, while smoothly transitioning to BALANI-NO's global rational pole-dominant mode in high-resolution/zero-shot regions?

4 comments
AuthorsOpenReview ↗

Response to Reviewer wTJe (1/2): training-resolution trade-off, cost and stability

We thank the reviewer for the positive assessment of the physics-informed design and of the zero-shot transfer results, and for the detailed questions. Please see our general response for a summary of all revisions; changes in the revised PDF are marked in red.

W1. Training-resolution performance is not always optimal.

We agree. BALANI-NO is the training-resolution winner on Navier–Stokes and subhalo lensing, ties U-FNO on full-field lensing, and is 3rd on Darcy, 2nd on CAMELS and 4th on EMAG2 (Table 7). On those three tasks, CNN or PDNO wins at the training grid but degrades by 7.37.3–41.6×41.6\times after a resolution change, while BALANI-NO has the lowest zero-shot error on every task. Section 5 and the Limitations discuss this trade-off. The revised Discussion also gives the reason for each task: Darcy has no global symbol, the CAMELS input is subsampled and does not determine the target, and the EMAG2 target pipeline includes a taper that is not shift-invariant (Appendix C).

W2. Operators that vary with the input (no single global symbol).

We agree, and the Limitations now state this explicitly for Darcy. The new Appendix A tests it in a controlled setting. On Allen–Cahn we vary only the strength of the nonlinearity ε\varepsilon, from pure diffusion (which has an exact symbol) to a strongly nonlinear map (which has none). For ε=0\varepsilon = 0 the pole branch alone matches the full model (+0.0014±0.0051+0.0014 \pm 0.0051). For ε≥1\varepsilon \geq 1 the table gives a significant gain, and the 95% paired intervals exclude zero (Table 9). The pole branch is needed at every ε\varepsilon: removing it raises the error by 2.32.3–3.6×3.6\times (Table 8).

W3 / Q2. Training time, memory and numerical stability.

The new Appendix L measures training and inference cost for all eleven architectures (Tesla T4, batch 16, Navier–Stokes configurations; 20 timed repetitions after 3 warm-up runs). Selected rows from Table 23:

ModelTrain ms/stepTrain MiBInference @64 (ms)Inference @256 (ms)
BALANI-NO13.96460.34 (0.18 cached)3.60 (3.44 cached)
FNO7.92330.131.74
LNO6.28540.223.33
PDNO8.10440.213.06
U-FNO14.20310.281.54
UNO12.211640.6812.31
  • A training step costs 1.76×1.76\times FNO, similar to U-FNO, HC-UNO and UNO. Peak training memory (46 MiB) is above FNO but below ALNO, LNO, CNN, UNO and HC-UNO.
  • The extra cost comes from evaluating the rational symbol at every retained frequency. The symbol depends only on the grid size and the trained weights, so at inference it can be computed once per resolution and reused. With this cache, inference at 64×6464\times64 takes 1.38×1.38\times FNO and is faster than PDNO, LNO, CNN, U-FNO, ALNO, UNO and HC-UNO.
  • Stability: we saw no divergence or NaN values in any run. The softplus keeps every sj2>0s_j^2 > 0, so the denominator ∣k∣2+sj2|k|^2 + s_j^2 never vanishes for k≠0k \neq 0, and k=0k = 0 uses the separate offset DD.
AuthorsOpenReview ↗

Response to Reviewer wTJe (2/2): pole initialization and dynamics, future work

Q1. Sensitivity to initial pole locations; stuck or collapsing poles.

We thank the reviewer for this question. We now study it directly in the new Appendix B.2, on a screened-Poisson task with a known true pole (s=4s = 4) and on Allen–Cahn (ε=4\varepsilon = 4), which has none.

  • Default training. The poles stay within 9% of their initial values and remain well separated (minimum gap 0.17). The softplus is the cause: at the smallest initial value, s^j=−6\hat s_j = -6, its derivative is σ(−6)≈0.0025\sigma(-6) \approx 0.0025, so these poles receive gradients about 400 times weaker than an unconstrained parameter. The optimizer therefore reduces the loss mainly through the residues, and the pole branch acts as a fixed multi-scale rational basis with learned residues. The main-task models behave the same way: the block-0 poles of the Navier–Stokes and full-field lensing models move by at most 0.01 and 0.024 in absolute terms.
  • Convergence to a physical scale. For a single linear pole layer the scale is identifiable: fitting by grid search recovers known screening lengths within 0.11% for one pole, and within 0.42% for the smaller of two poles (Appendix B.1, Table 10). In the full three-block network it is not. Even with a separate pole learning rate of 100×100\times, no block-0 pole reaches the true value 4 (the largest reaches 2.62), because lifting, projection and pointwise nonlinearities between blocks mean that no single block needs to hold the true pole.
  • Collapse. Under the default optimizer we see no collapse. With a 100×100\times pole learning rate, two poles in block 1 of the Allen–Cahn model do collapse (2.7384 and 2.7372), so freely moving poles can degenerate.
  • Initialization. The initial range matters. On screened Poisson the error ranges from 0.0049 to 0.0098 across five initializations (Table 12). Starting from larger poles, closer to the true scale, reduces the error by 36% relative to the default, and a separate pole learning rate reduces it by up to 21% (Table 11).

On the second pole's larger uncertainty that the reviewer mentions: in Table 10 it comes from overlapping contributions. Once the first pole captures the dominant low-frequency behavior, the second is less well determined, while the test error stays below 1.5×10−31.5 \times 10^{-3}.

Following this analysis, we have revised contribution (i) and the Limitations to describe the pole branch as a multi-scale rational basis with learned residues and trainable pole locations. To keep the comparison fair, the main tables use the same default initialization and optimizer on every task, and none of the improved settings above.

Q3. Adaptive hybrid weighting between local corrections and the global pole branch.

We thank the reviewer for this suggestion and have added it to the Conclusion as future work: a learned, resolution-aware weighting between the global pole branch and local corrections. The local part would need to be defined independently of the grid, since the grid-tied flexibility of CNN and PDNO is exactly what makes them degrade under a resolution change.

We hope these additions address the reviewer's concerns and would be glad to clarify any remaining points.

The revision is responsive, the method is simple, cheap and reproducible, and the empirical protocol — eleven architectures at matched parameter budgets, seven seeds, six tasks including two on real data — is more thorough than is typical at this venue. The new appendices add scientific content rather than filler, and the honest reporting of negative results is a genuine service to the community. My remaining objections concern claim calibration, one missing control, and one self-imposed handicap in the protocol; none requires new theory or a new dataset, and all can be handled in a camera-ready.

Three changes would move me from weak accept to accept: (i) the M2 reframing in the abstract and Section 4.3, so that the reader learns from the abstract, not from a footnote, that the pole branch does not beat a matched table on full-field lensing; (ii) the M3 re-run under the better pole setting, or an explicit statement that the adaptivity of the basis is nominal; (iii) the m1 paired tests on the headline claims, with ties labelled as ties.

AuthorsOpenReview ↗

Response to Reviewer wTJe: all three requested changes made

We thank the reviewer for the detailed follow-up. We have made all three changes in the revised PDF. This round's changes are marked in blue, and changes from the first revision remain in red.

(i) Full-field lensing in the abstract and the ablation. The abstract now states that the pole branch is needed on Navier–Stokes and subhalo lensing, while on full-field lensing a parameter-matched table-only model performs equally well. The ablation (Section 4.4) reports this in the main text with the paired differences (−0.0002±0.0003-0.0002 \pm 0.0003, 0.0000±0.00040.0000 \pm 0.0004 and 0.0000±0.00050.0000 \pm 0.0005 on WDM, CDM and Axion). The full-field convergence is dominated by the smooth main lens, so most of the target energy lies in the low modes that a wider table covers.

(ii) Adaptivity of the basis. We chose the explicit statement. Section 3 and the Limitations (now Appendix A) state that under the protocol used for all main results the pole locations move by at most 9%9\% from their initialization, so the adaptivity of the basis is nominal: the pole branch acts as a fixed multi-scale rational basis with learned residues. A separate pole learning rate and a task-informed initialization reduce the error on the screened-Poisson task in Appendix D.2, but we do not use them in the main tables, so that all architectures share the same training protocol.

(iii) Paired tests on the headline claims. A new appendix (Appendix B, Table 8) reports paired 95%95\% intervals (tt-test over 7 seeds, df=6\mathrm{df}=6) between BALANI-NO and the strongest baselines, with ties and losses labelled, and Section 4.3 points to it. All zero-shot headline claims hold. The difference (baseline minus BALANI-NO) against the closest baseline is:

TaskSplitClosest baselineDifference
Darcy32×3232\times32 (zero-shot)FNO+0.0106±0.0050+0.0106 \pm 0.0050 (win)
Navier–Stokes64×6464\times64 (zero-shot)FNO+0.0220±0.0030+0.0220 \pm 0.0030 (win)
Full-field lensingCDM (zero-shot)FNO+0.0025±0.0003+0.0025 \pm 0.0003 (win)
Subhalo lensingCDM (zero-shot)PDNO+0.0061±0.0025+0.0061 \pm 0.0025 (win)
Darcy16×1616\times16HC-UNO+0.0024±0.0034+0.0024 \pm 0.0034 (tie)
Darcy16×1616\times16PDNO−0.0040±0.0027-0.0040 \pm 0.0027 (loss)
Darcy16×1616\times16CNN−0.0125±0.0022-0.0125 \pm 0.0022 (loss)

The ties and losses are the in-distribution Darcy comparisons already discussed in Section 4.

We believe these changes address the three points listed in your comment, and we thank you again for the careful reading.

Reviewer_4QksOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

BALANI-NO splits the spectral symbol into two parts: a pole (rational) branch in the isotropic frequency magnitude, which is defined at every frequency and can be evaluated on any FFT grid, plus a small Fourier table T(k) on the lowest modes in the FNO style. The rest of the block — coordinate channels, 1x1 lift, L spectral blocks, 1x1 head — is shared with FNO/ALNO/LNO/PDNO, so the eleven compared architectures differ only in how W(k) is parameterised (Fig. 1). The evaluation covers Darcy 16->32, Navier–Stokes 32->64, gravitational-lensing κ→α\kappa\rightarrow \alpha trained on WDM and tested on CDM and axion, plus two additional tasks (CAMELS gravitational potential, EMAG2 magnetic upward continuation), with seven seeds and matched parameter budgets, together with branch ablations and a pole-recovery study.

Critical strengths

The empirical protocol is the strongest part of the paper and is better than what one usually sees in this literature: eleven architectures inside a shared template, seven seeds, parameter budgets matched to within a few percent, an explicit parameter-matched ablation to separate "combining branches" from "adding capacity", two additional tasks including real measured data, and error bars reported consistently. The design idea is also well motivated. For elliptic operators the symbol is an analytic function of |k|, so parameterising it with a few poles that are defined at every frequency and then correcting the low modes is a sensible response to the chief weakness of a truncated lookup table. I appreciate in particular that the authors do not hide the results that go against them: on Darcy the model ranks third at the training resolution, and section 4 explains why. The limitations section, and the explicit downgrading of the CAMELS experiment to a "fixed-cosmology, single-suite pilot" (App. E), are exactly the kind of honesty that makes a paper usable.

Critical weaknesses

no analytic-oracle baseline on tasks whose operator is known in closed form. Lensing, CAMELS and EMAG2 all admit exact FFT solutions, and the authors say as much: the EMAG2 targets are generated by applying the known kernel, so "the reported errors therefore measure approximation to the analytic operator" (App. F). Yet every comparison is model against model. Without the exact operator's error on the same discretisation, one cannot tell whether BALANI-NO's 0.01130.0113 at 64×6464\times64 and 0.03990.0399 at 128×128128\times128 are near the attainable floor or an order of magnitude above it — and therefore whether the pole branch has learned the operator or merely ranked first among approximations. This also weakens the interpretation of the winning-task pattern: the advantage is measured only where an exact answer exists.

Questions for authors

What is the relative L2L^2 error of the exact analytic operator (FFT Poisson solve for CAMELS, ik ∣k∣−2i\mathbf{k}\,|\mathbf{k}|^{-2} for lensing, exp⁡(−∣k∣Δz)\exp(-|\mathbf{k}|\Delta z) for EMAG2) on the same test splits, with the same windowing and padding? If the network remains well above that parameter-free baseline, what is the intended use case, and would you restate the contribution as a parameter-efficiency and transfer result rather than an accuracy result?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 4Qks (1/2): analytic-oracle baselines

We thank the reviewer for the positive assessment of our empirical protocol and design, and for the precise suggestion of analytic-oracle baselines, which has made the paper considerably stronger. Please see our general response for a summary of all revisions; changes in the revised PDF are marked in red. After the revision, the CAMELS and EMAG2 appendices cited in the review (App. E and F) are now Appendices H and I.

W1 / Q1. Error of the exact analytic operator on the same splits.

We have added Appendix C, which applies the exact operator to the same test inputs as the models, with the same resampling, windowing, padding and error metric as the main tables. When targets are defined only up to a constant, we fit one global scale on the training set. We also report the best linear filter: the per-frequency least-squares fit W(k)=∑ix^i(k)‾y^i(k)/∑i∣x^i(k)∣2W(k) = \sum_i \overline{\hat x_i(k)} \hat y_i(k)/ \sum_i |\hat x_i(k)|^2 on the training set, which is the lowest training error any fixed Fourier multiplier can reach.

Lensing (Tables 13–14). The targets are finite-difference gradients of the resampled potential, so the exact operator is a periodic Poisson solve followed by the same finite-difference gradient. We also report the spectral symbol 2ik/∣k∣22ik/|k|^2 that the reviewer suggests.

Relative L2L^2Full WDMFull CDMFull AxionSub. WDMSub. CDMSub. Axion
Exact operator0.00390.00410.00390.00240.00410.0020
Spectral symbol 2ik/∣k∣22ik/\lvert k\rvert^20.01410.01430.01370.02030.02180.0165
Best linear filter0.00570.00730.00640.00760.00990.0106
BALANI-NO0.00380.00600.00640.00580.01090.0124
FNO0.00460.00840.00860.01730.03870.0310
  • The exact operator's residual error (0.2–0.4%) comes from resampling the inputs. The spectral symbol is 3–8 times worse because it uses a different derivative from the one that defines the targets.
  • BALANI-NO reaches the exact operator's error on full-field WDM (0.0038 vs. 0.0039). On the zero-shot splits it is 1.5–1.6 times above it on full field and 2.7–6.2 times above it on subhalo, and it is the closest of all eleven learned models on every zero-shot split (Table 14).
  • The exact operator does not degrade from WDM to CDM or axion, so the degradation in the main tables comes from learning.
  • The best linear filter, fitted on the WDM training grid, is below BALANI-NO on the zero-shot subhalo splits (0.0099 vs. 0.0109 on CDM, 0.0106 vs. 0.0124 on axion) and ties it on full-field axion. This is expected for a linear, nearly shift-invariant map evaluated at one fixed resolution. Such a filter exists only on the grid on which it is fitted, so it has no counterpart under a resolution change, and it fails when the map is not a fixed multiplier (0.34 on EMAG2, 0.69 on CAMELS; see below).
AuthorsOpenReview ↗

Response to Reviewer 4Qks (2/2): EMAG2, CAMELS and intended use

EMAG2 (Table 15). The targets are produced by a Tukey taper, 50% zero-padding, the kernel exp⁡(−∣k∣Δz)\exp(-|k|\Delta z) with each patch's latitude-dependent east–west spacing, and cropping. The exact pipeline reproduces the targets to machine precision. The models do not see the latitude, so we also apply the pipeline with one mean spacing for all patches: this costs 0.27%, the lowest error any single fixed operator can reach.

Relative L2L^264×6464\times64128×128128\times128 (zero-shot)
Exact pipeline, mean spacing0.00270.0027
Kernel without taper or padding0.45080.4717
Best linear filter0.3418–
BALANI-NO0.01130.0399
FNO0.01270.0402

BALANI-NO is 4 times above this floor at 64×6464\times64 and 15 times above it at 128×128128\times128, where it has the lowest error of all learned models. The kernel alone and the best linear filter are far worse (0.34–0.47) because the taper is not shift-invariant; the networks represent it through their coordinate channels.

CAMELS (Table 16). We undo the standardization and logarithm of the model input and solve the Poisson equation at the input resolution.

Relative L2L^232×3232\times3264×6464\times64 (zero-shot)
Poisson solve on model input0.40900.2011
Poisson solve on resized linear density0.44930.2001
Best linear filter0.6901–
BALANI-NO0.23180.1769

BALANI-NO has lower error than the closed-form solve at both resolutions, but we do not read this as beating the exact operator. The 256×256256\times256 density is bilinearly resized by a factor of 8 without anti-aliasing, so much of the mass in each coarse cell is missing from the input while the smooth potential survives. The input therefore does not determine the target, and learned models compensate using training statistics. Solving on the linear density gives the same error, so the log transform is not the cause, and the best linear filter reaches only 0.69, so this compensation needs a nonlinear map. CAMELS therefore tests transfer under incomplete inputs, not recovery of a clean operator; the Discussion now states this.

Q2. Intended use case, and restating the contribution.

We agree with the reviewer's framing. Where the operator is known exactly, linear and shift-invariant, and the input determines the target, an exact solver is the right tool, and even a fixed Fourier multiplier fitted on the training grid is competitive at that grid. We use such tasks as controlled tests where the floor is known: on lensing they show that BALANI-NO comes close to that floor, closer than any other learned model on every zero-shot split.

The intended use is operators with no closed form or fixed symbol, or with incomplete inputs, such as Darcy, Navier–Stokes, nonlinear reaction–diffusion (new Appendix A) and CAMELS. We have revised the Limitations accordingly: our contribution is accuracy and resolution transfer at matched parameter budgets on such operators, not accuracy beyond a known solver.

We hope these additions address the reviewer's concerns and would be glad to clarify any remaining points.

AuthorsOpenReview ↗

Follow-up: analytic-oracle baselines and revised framing

Dear Reviewer 4Qks,

Thank you again for suggesting the analytic-oracle baselines. They are now in Appendix C for lensing, EMAG2 and CAMELS, computed on the same splits and preprocessing as the models, and we have restated the contribution as a transfer and parameter-efficiency result as you proposed. The discussion period closes tomorrow. We would be grateful to know whether these additions address your question, and we are happy to clarify anything further.

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

The authors introduce a neural operator in which the frequency dependent function W(k) is the sum of a pole and table branch. The pole branch is the primary contribution of this work. This pole branch is designed to be a nonlinear function of the frequency k. The authors claim with this choice the neural operator can better learn elliptic functions such as gravitational lensing

Critical strengths

The paper aims to improve the interpretability and utility of neural operators

Critical weaknesses
  • My opinion is that this paper is poorly written/organized. While it is readable (i.e., the grammar and sentence construction is fine) the content doesn't perform the expected functions. For example instead of motivating the problem, communicating the gap in the literature, summarizing the method, results, and contributions, the introduction mostly lists related work.
  • The font, and headings do not follow the ICLR template.
  • It isn’t clear to me that the experiments validate the claim. For example the ablation doesn't seem to show that both branches are needed for strong performance, in particular for most experiments the poles only branch performs comparably with the full model.
  • It is not clear to me that the experiments are performed on conventional datasets
Questions for authors

None

3 comments
AuthorsOpenReview ↗

Response to Reviewer WGyp (1/2): introduction, formatting and branch ablation

We thank the reviewer for the careful reading and for recognizing that the paper aims to improve the interpretability and utility of neural operators. Please see our general response for a summary of all revisions; changes in the revised PDF are marked in red.

W1. The introduction mostly lists related work.

We agree and have rewritten the introduction. It now (i) motivates the problem: lookup-table symbols such as FNO's say nothing outside the retained modes, and variants with grid-dependent components degrade under a resolution change; (ii) states the gap: existing pole-based operators (LNO, ALNO) place their poles along a single frequency axis and trail FNO at matched parameter budgets; (iii) summarizes BALANI-NO and its main results; and (iv) ends with an explicit list of contributions. The survey of prior work now sits in Section 2.

W2. Font and headings do not follow the ICLR template.

We apologize for this oversight. The paper had been compiled with XeLaTeX, which cannot load the Times font required by the ICLR style file, so the default font was used and the small-caps headings were lost. The revised paper is compiled with pdfLaTeX using the unmodified ICLR 2027 style file, and all headings follow the template.

W3. The ablation does not show that both branches are needed.

We thank the reviewer for this important point. We have made two additions.

First, the ablation (Section 4.4) now includes parameter-matched poles-only and table-only variants, enlarged to roughly BALANI-NO's parameter count, which separates the benefit of combining branches from that of added capacity. Relative L2L^2 error (Tables 5–6):

VariantDarcy @32NS @64Full CDMSub. WDMSub. CDMSub. Axion
Poles only0.11720.15190.00750.00730.01180.0143
Poles only, matched0.12000.15160.00830.00760.01380.0198
Table only0.12810.17640.00700.01500.03200.0275
Table only, matched0.13230.14470.00600.01170.02820.0216
Both (BALANI-NO)0.12070.12240.00600.00570.01120.0132

The ablation was trained as a separate set of runs, so the full-model values differ slightly from Tables 1–4.

  • Navier–Stokes: the full model is best at both resolutions, including against the parameter-matched poles-only model (0.1224 vs. 0.1516 at 64×6464\times64, 19% lower).
  • Subhalo lensing: the full model has the lowest error on all three splits, 22% below the best single-branch variant on WDM and 5–8% below it on the zero-shot splits.
  • Full-field lensing: the full model is 7–24% below both single-branch models of the original size and below the matched poles-only model. It matches the parameter-matched table-only model (paired differences −0.0002±0.0003-0.0002 \pm 0.0003, 0.0000±0.00040.0000 \pm 0.0004 and 0.0000±0.00050.0000 \pm 0.0005 on WDM, CDM and axion). The full-field convergence is dominated by the smooth main lens, so most of the target energy lies in low modes that a wider table covers.
  • Darcy: the poles-only models are the most accurate, so here the pole branch alone suffices.

So the poles-only model is comparable to the full model on Darcy, but not elsewhere: relative to it, the full model's error is 19% lower on Navier–Stokes, 17–24% lower on full-field lensing and 5–22% lower on subhalo lensing.

AuthorsOpenReview ↗

Response to Reviewer WGyp (2/2): controlled study and datasets

Second, the new Appendix A asks when the table is needed in a controlled setting. On Allen–Cahn we vary only the strength of the nonlinearity ε\varepsilon, with everything else fixed. For pure diffusion (ε=0\varepsilon = 0), which has an exact symbol, poles alone match the full model (+0.0014±0.0051+0.0014 \pm 0.0051). For ε≥1\varepsilon \geq 1 the table gives a statistically significant gain (95% paired intervals exclude zero, Table 9), and it is not explained by capacity: the matched poles-only model is worse than the original at every ε\varepsilon. Removing the pole branch instead raises the error by 2.32.3–3.6×3.6\times at every ε\varepsilon. Darcy also lacks a global symbol yet gains nothing from the table, so the revised paper states that the table's role is task-dependent: the pole branch is the main component, and the table adds a significant gain on Navier–Stokes, subhalo lensing and nonlinear reaction–diffusion.

W4. It is not clear that the experiments use conventional datasets.

We agree that the source of each dataset was not stated clearly enough, and we have added a dataset appendix with a summary table (Appendix K, Table 22):

  • Darcy flow: the original FNO benchmark of Li et al. (2021), as distributed with the neuraloperator library. This is the split the library uses in its own examples and for resolution-transfer evaluation.
  • Navier–Stokes: the Re=500Re = 500 vorticity dataset from the same library, taken from its public Zenodo record.
  • CAMELS and EMAG2v3: public scientific datasets (Villaescusa-Navarro et al., 2022; NOAA NCEI), Appendices H and I.
  • Gravitational lensing: to our knowledge no public benchmark exists for learning the convergence-to-deflection operator across dark-matter models, so we generated one with lenstronomy; it will be released with the code.
  • Reaction–diffusion: synthetic by design, used only as a controlled study (Appendix A).

On the same Darcy split, our models are also more accurate than the library's own FNO reference models, with fewer parameters (relative L2L^2 error):

ModelParametersDarcy 16×1616\times16Darcy 32×3232\times32 (zero-shot)
neuraloperator FNO tutorial191,8810.13310.1616
neuraloperator FNO tutorial~1.19M0.11160.1519
Our FNO baseline111,8570.10280.1300
BALANI-NO112,3870.09010.1194

Our results are means over seven seeds after 60 epochs; the library numbers are from its FNO Darcy tutorials (15–20 epochs), so they are reference points rather than tuned baselines.

We also acknowledge that the 16×1616\times16 Darcy and 32×3232\times32 Navier–Stokes training grids are coarser than those usually reported. They follow the splits distributed with the library and let us train eleven architectures with seven seeds each on a single T4 GPU. The Limitations now state that our claims concern resolution transfer at matched parameter budgets, not absolute accuracy at high resolution.

We hope these revisions address the reviewer's concerns and would be glad to clarify any remaining points.

AuthorsOpenReview ↗

Follow-up: have our revisions addressed your concerns?

Dear Reviewer WGyp,

Thank you again for your review. In our responses above we addressed each of your four points: the introduction is rewritten with explicit contributions, the paper is recompiled with the official ICLR style, a new controlled study (Appendix A) shows when each branch of the symbol is needed, and Appendix K documents the provenance of every dataset. The discussion period closes tomorrow. We would be grateful to know whether these revisions address your concerns, and we are happy to clarify anything further.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers: summary of revisions

We thank all three reviewers for their careful reading and constructive feedback. We have uploaded a revised PDF in which all changes are marked in red.

Scope of the revision

The revision does not change the method, the main benchmark results (Tables 1–4) or the core claims of the paper. The new material answers specific questions raised by the reviewers. It is placed in the appendices so that the main text keeps the same structure, and the main text changes only to clarify the motivation, the scope of the claims and the limitations.

Analyses added in response to reviewer questions

AdditionRaised byQuestion answered
App. A: controlled Allen–Cahn studyWGyp (W3), wTJe (W2)When is each branch of the symbol needed?
App. B.2: pole dynamics and initializationwTJe (Q1)Do poles move, get stuck or collapse, and how sensitive is training to their initial values?
App. C: analytic-oracle baselines4Qks (W1, Q1)How close are the learned models to the exact operator on lensing, CAMELS and EMAG2?
App. K: dataset provenance and splitsWGyp (W4)Where does each dataset come from? (documentation of the existing datasets; no new data)
App. L: training and inference costwTJe (W3, Q2)What do the models cost in time and memory? (measurements of the existing models)

Clarifications in the main text

  • Introduction rewritten to motivate the problem, state the gap, summarize the results and list the contributions (WGyp W1).
  • Paper recompiled with the unmodified ICLR 2027 style file (WGyp W2).
  • Limitations: our contribution is stated as accuracy and resolution transfer at matched parameter budgets, not accuracy beyond a known solver (4Qks Q2), and the pole branch is described as a multi-scale rational basis with learned residues and trainable pole locations (wTJe Q1).
  • Discussion: a task-by-task explanation of where BALANI-NO is not the training-resolution winner (wTJe W1, W2).
  • Conclusion: a resolution-aware weighting between the pole branch and local corrections added as future work (wTJe Q3).

Appendix renumbering. The CAMELS and EMAG2 appendices cited in the reviews (App. E and F) are now Appendices H and I.

We respond to each reviewer's points individually below their review.

AuthorsOpenReview ↗

General response (update): second revision and appendix renumbering

Update (17 Nov 2026). We have uploaded a second revision in response to Reviewer wTJe's follow-up. Changes from this round are marked in blue, and changes from the first revision remain in red. The Limitations section is now Appendix A, and a new Appendix B reports paired significance tests for the main comparisons. As a result, all later appendices have moved by two letters. The references below have been updated to the current PDF. In the main text, the abstract and the ablation (Section 4.4) now state that on full-field lensing a parameter-matched table-only model performs as well as BALANI-NO, Section 3 states that the adaptivity of the pole basis is nominal under our training protocol, and the Discussion was condensed to stay within the page limit.

We thank all three reviewers for their careful reading and constructive feedback. We have uploaded a revised PDF in which all changes are marked in color.

Scope of the revision

The revision does not change the method, the main benchmark results (Tables 1–4) or the core claims of the paper. The new material answers specific questions raised by the reviewers. It is placed in the appendices so that the main text keeps the same structure, and the main text changes only to clarify the motivation, the scope of the claims and the limitations.

Analyses added in response to reviewer questions

AdditionRaised byQuestion answered
App. B: paired significance testswTJeDo the headline comparisons hold under paired tests over seeds?
App. C: controlled reaction–diffusion studyWGyp (W3), wTJe (W2)When is each branch of the symbol needed?
App. D.2: pole dynamics and initializationwTJe (Q1)Do poles move, get stuck or collapse, and how sensitive is training to their initial values?
App. E: analytic-oracle baselines4Qks (W1, Q1)How close are the learned models to the exact operator on lensing, CAMELS and EMAG2?
App. M: dataset provenance and splitsWGyp (W4)Where does each dataset come from? (documentation of the existing datasets; no new data)
App. N: training and inference costwTJe (W3, Q2)What do the models cost in time and memory? (measurements of the existing models)

Clarifications in the paper

  • Introduction rewritten to motivate the problem, state the gap, summarize the results and list the contributions (WGyp W1).
  • Paper recompiled with the unmodified ICLR 2027 style file (WGyp W2).
  • Limitations (now Appendix A): our contribution is stated as accuracy and resolution transfer at matched parameter budgets, not accuracy beyond a known solver (4Qks Q2), and the pole branch is described as a multi-scale rational basis with learned residues and trainable pole locations whose adaptivity is nominal under our training protocol (wTJe Q1).
  • Discussion: a task-by-task explanation of where BALANI-NO is not the training-resolution winner (wTJe W1, W2).
  • Conclusion: a resolution-aware weighting between the pole branch and local corrections added as future work (wTJe Q3).

Appendix renumbering. The CAMELS and EMAG2 appendices cited in the reviews (App. E and F) are now Appendices J and K.

We respond to each reviewer's points individually below their review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.