acceptodds
Under review as a conference paper at ICLR 2027

APEX: Amplitude Anchors and Phase Priors for Target-Scarce Higher-Frequency Wave Prediction

Abstract

Learning-based surrogates have become increasingly effective for wave-field prediction, and neural operators in particular have shown strong performance within observed frequency regimes. However, higher-frequency prediction under scarce target supervision remains comparatively underexplored, especially in wave problems where higher-frequency data are substantially more expensive to simulate or measure than lower-frequency data. A central difficulty is that cross-frequency transfer is inherently asymmetric: coarse amplitude structure remains relatively stable across frequencies, whereas phase-sensitive oscillatory structure deteriorates much more rapidly as frequency increases. Motivated by this asymmetry, we propose , mplitude-anchored and hase-prior-guided nhancement from etrapolated coarse predictions, a framework for target-scarce higher-frequency wave-field prediction. A lower-frequency neural operator first provides a coarse prediction in the target-frequency regime, from which we retain only the amplitude as a transferable structural anchor. A conditional flow-matching enhancer then reconstructs the target higher-frequency field under the guidance of a Green's-function-inspired phase prior. Experiments on SimpleWave, Helmholtz, and Maxwell benchmarks show that APEX consistently outperforms direct lower-to-higher extrapolation, target-adapted operator, and joint generative baselines under limited target-frequency supervision. Our results suggest that reliable higher-frequency prediction of oscillatory wave fields should not rely on direct end-to-end transfer of the full complex field, but instead on explicitly reusing transferable coarse structure while separately recovering the missing oscillatory detail.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cross-Frequency Generalization, Higher-Frequency Wave-Field Prediction, Neural Operators

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

The paper studies higher-frequency wave-field prediction when only lower-frequency supervision is abundant and higher-frequency (target) supervision is scarce. The authors first show empirically (Fig. 1) that cross-frequency transfer is asymmetric: amplitude similarity across frequencies decays much more slowly than phase coherence, both in ground truth and in an FNO trained on lower frequencies and extrapolated to higher ones. This is formalized with a regional complex-field error decomposition (Proposition 1) and a frequency-sensitive upper bound (Corollary 2) that separate the error into an amplitude-mismatch term and a phase-mismatch term whose argument scales with frequency.

Motivated by this, APEX (1) freezes a lower-frequency-trained FNO and uses its log-amplitude output at the target frequency as a "coarse amplitude anchor"; (2) builds a lightweight, geometry-aware "phase prior" from a truncated sum of plane-wave-like terms with a domain-adaptive reference phase coefficient (Eq. 13), intended as a Green's-function-inspired surrogate for the oscillatory phase; and (3) uses both as conditioning signals for a U-Net-based conditional flow-matching model that generates the target complex field (log-amplitude + sin/cos phase representation). The pipeline is evaluated on three synthetic benchmarks (SimpleWave, Helmholtz, Maxwell) at three extrapolation gaps (HF20/50/100), against LF-extrapolation, fine-tuned, and jointly-trained FNO/HNO/CFM baselines, plus an ablation removing each conditioning signal and a sensitivity study over the HF supervision ratio.

Critical strengths

1. The transfer-asymmetry observation is a useful empirical framing, and Proposition 1 cleanly isolates where the two field components enter the error.

Separating amplitude mismatch (direct term) from phase mismatch (an oscillatory term with frequency-scaled argument) in Eq. 10 is a clear way to state why phase, not amplitude, should be the harder quantity to extrapolate, and it motivates the architecture directly rather than being decorative.

2. The method is simple and the components are well-matched to the stated problem. Freezing a cheap, already-trained LF model as a structural anchor, rather than trying to make a single network transfer the whole complex field, is a reasonable and backbone-agnostic design choice (Sec. 3.2 states this explicitly), and the ablation in Table 2 does show both conditioning signals contribute on the one benchmark where the ablation is reported.

Critical weaknesses

W1. The theoretical results do not explain the observed asymmetry; part of it is close to tautological given how phase similarity is measured.

Proposition 1 is a geometric identity (law of cosines applied to the amplitude-phase decomposition) and is correct, but it does not establish why phase mismatch should grow faster than amplitude mismatch with frequency — it only shows that if Δτ(r)\Delta\tau(\mathbf r) grows with ν\nu, the phase term is weighted by ν2\nu^2 inside a sin⁡2\sin^2. Corollary 2 is explicitly flagged by the authors as "explanatory rather than tight" and assumes bounded amplitudes, so it adds little beyond restating the same point. Neither result shows that Δτ\Delta\tau itself grows with the extrapolation gap; that claim rests entirely on the empirical similarity measures in Fig. 1, which is a different kind of evidence from the "formalize this asymmetry" framing in the text (p.4).

Moreover, the phase-similarity metric itself (Eq. 6, phase-factor coherence) is close to guaranteed to decay faster than the cosine-based amplitude similarity (Eq. 5) once travel-time differences Δτ\Delta\tau reach O(1/ν)O(1/\nu), simply because phase is 2π2\pi-periodic and enters through exp⁡(iνΔτ)\exp(i\nu\Delta\tau): any approximately linear growth of timing error with distance will wrap many times at higher ν\nu and decorrelate the phase factor, while the same timing error leaves amplitude envelopes comparatively undisturbed because amplitude does not wrap. This is a property of the chosen similarity measures and the oscillatory representation, not a newly discovered property of wave fields. The paper would be stronger if it showed the asymmetry survives under a metric that is not inherently biased toward this conclusion (e.g., unwrapped phase error, or an amplitude metric equally sensitive to small structural shifts).

W2. The phase prior is an ungrounded heuristic, and this is never tested against simpler alternatives.

Eq. 13 defines the phase prior through MM "dominant propagation paths" with path lengths LmL_m and fixed coefficients ama_m, and a "domain-adaptive reference phase coefficient" κref(e,ν)\kappa_\mathrm{ref}(e,\nu). The paper states these are "concrete instantiations... in Appendix D" but the main text gives no indication of how MM, ama_m, or κref\kappa_\mathrm{ref} are chosen, nor whether they are hand-tuned per benchmark or genuinely domain-adaptive. Without this, "Green's-function-inspired" is a label rather than a demonstrated connection to Eq. 2's actual Green's function.

More importantly, no ablation isolates whether the specific geometric content of the phase prior matters at all, as opposed to merely providing the flow-matching model with some smoothly-varying, frequency-dependent conditioning signal. Candidate controls that are absent: (a) a random or scrambled phase prior with the same statistics, (b) a naive linear-in-ν\nu phase ramp with no path/geometry structure, (c) a learned phase prior instead of the fixed analytic surrogate. Table 2's "w/o Phase Prior" shows removing it hurts, which only shows conditioning on some phase-like signal helps, not that the Green's-function construction is the right one or better than a trivial baseline.

W3. Evaluation is confined to small synthetic grids with no real or more complex data, and only one benchmark gets the ablation and supervision-sensitivity study.

All three benchmarks are synthetic, on fixed small grids (64×6464\times64 or 80×9280\times92), with scalar or single-polarization fields. There is no evaluation on a real-world or higher-resolution/more complex-geometry dataset, which matters because the central practical motivation (Sec. 1) is that "higher-frequency samples... require finer discretization, stricter control, and substantially higher simulation or measurement cost" — a claim about real deployment settings that the experiments never probe. In addition, Table 2 (ablation) and Table 3 (supervision-ratio sensitivity) are reported only for Helmholtz; the paper does not say whether the same ablation trend holds on SimpleWave or Maxwell, so the generality of W2's "both components help" conclusion within the paper's own evidence is unconfirmed on 2/3 benchmarks.

W4. No variance/seed information, and no training/inference cost accounting.

Table 1–3 report single numbers with no standard deviation or seed count. Given that the margins over the best baseline are sometimes modest relative to the baseline spread (e.g., Maxwell HF100: APEX H1=0.929H^1=0.929 vs. HNO-Joint 0.9480.948 vs. FNO-Joint 0.9590.959 — all three within about 3% of each other), it is not possible to assess whether the ranking is stable. The paper also never reports the added training/inference cost of running a frozen FNO plus a U-Net CFM enhancer relative to the single-model baselines (FNO-FT, HNO-FT), which matters for a paper whose motivation is a cost-sensitive data regime.

Questions for authors
  1. Isolating the asymmetry claim from the metric. Can you show the amplitude/phase asymmetry in Fig. 1 holds under an unwrapped or otherwise non-periodic phase-error metric, to rule out that the effect is an artifact of phase being circular while amplitude is not?
  2. Phase prior ablation. How sensitive are the Table 1/2 results to the specific choice of MM, ama_m, and κref\kappa_\mathrm{ref} in Eq. 13? Does a random-phase or linear-ramp control with matched statistics perform comparably to the Green's-function-inspired prior?
  3. Generality of the ablation. Do the Table 2 and Table 3 trends (both components help; APEX > CFM-Joint at every supervision ratio) hold on SimpleWave and Maxwell as well as Helmholtz?
  4. Variance. How many seeds were used for Tables 1–3? Please report mean ± std, particularly for the closest-margin cells (e.g., Maxwell HF100).
  5. Cost. What is the added training and inference cost of the frozen-FNO + CFM pipeline relative to FNO-FT/HNO-FT?
3 comments

We thank the reviewer for the detailed comments and constructive suggestions. We have revised the manuscript to clarify the theoretical interpretation and phase-prior construction, and added experiments on phase-error metrics, ablations, training-seed stability, and computational cost. All modifications are highlighted in blue in the revised manuscript.

Weakness 1 / Question 1

We have clarified the role of the theoretical analysis in Sec. 2.3. Proposition 1 provides an exact decomposition of regional complex-field error into amplitude and phase-sensitive contributions. Corollary 2 further characterizes the explicit frequency dependence of the phase-related error bound through travel-time mismatch. These results reveal the different mechanisms through which amplitude and phase errors contribute to higher-frequency prediction, providing a complementary interpretation of the empirical asymmetry observed in Fig. 1.

To further examine the influence of phase periodicity, we have added a non-periodic phase-error evaluation in Appendix E.1 (Table 7). Specifically, we use PUMA-based phase unwrapping to compute phase RMSE on the full test sets of SimpleWave, Helmholtz, and Maxwell. Across all three benchmarks, unwrapped phase error increases substantially beyond the reference frequency. These results further support the observed amplitude–phase asymmetry under a non-periodic phase-error metric.

Weakness 2 / Question 2

We have expanded Sec. 3.3 to clarify the Green's-function-inspired construction. Based on the geometric propagation interpretation of the Green's-function representation, Eq. (13) constructs a lightweight multipath surrogate using dominant propagation paths and effective phase coefficients. Its phase provides source- and geometry-aware oscillatory cues for the generative enhancer, complementing the structural information supplied by the coarse amplitude anchor. The propagation paths and their weights are specified in advance, while the reference phase coefficient is computed from the target spectral variable and sample-dependent medium properties. The benchmark-specific configurations are provided in Appendix D (Table 6).

We have further evaluated the sensitivity to the number of paths MM, path weights ama_m, and reference phase coefficient κref\kappa_{\mathrm{ref}} on Maxwell in Appendix E.3 (Table 9). Here, α\alpha denotes the scaling factor applied to κref\kappa_{\mathrm{ref}}.

ConfigurationHF20 H¹↓HF20 AWPC↑HF50 H¹↓HF50 AWPC↑HF100 H¹↓HF100 AWPC↑Overall H¹↓Overall AWPC↑
Default0.6970.8120.7900.7250.9290.6240.8650.703
M=1M=10.6760.7680.8240.6281.1280.4231.0030.574
a1:a2=1:0.5a_1:a_2=1:0.50.6490.8090.6990.7450.9140.5950.8300.696
a1:a2=1:2a_1:a_2=1:20.7070.8470.7550.7760.9400.6800.8650.752
α=0.8\alpha=0.80.6790.8760.7810.7840.8940.7190.8390.779
α=1.2\alpha=1.20.7980.8080.8950.6551.2140.4451.0880.602

Removing the reflected path degrades overall reconstruction performance, particularly at larger frequency gaps. The model is relatively robust to changes in path weights, while the reference phase scale has a more pronounced influence on prediction accuracy. The default parameters serve as a representative configuration, leaving room for further improvement.

To examine the contribution of geometric information, we have added two phase-prior controls on Helmholtz in Sec. 5 (Table 2). Simple Phase Ramp uses frequency-dependent linear phase variation without propagation geometry, while Geometry-Shuffled Prior applies spatial rotations or flips to the original prior to disrupt its geometric alignment. Both controls use the same network architecture and training settings.

MethodHF20 H¹↓HF20 AWPC↑HF50 H¹↓HF50 AWPC↑HF100 H¹↓HF100 AWPC↑Overall H¹↓Overall AWPC↑
w/o Phase Prior0.4830.8800.5800.8090.8120.6890.6250.793
Simple Phase Ramp0.4760.9140.5950.8700.6810.7820.5840.856
Geometry-Shuffled Prior0.5620.8360.6390.7580.8580.5640.6860.720
APEX0.4220.9280.4920.8850.6340.8150.5160.876

The simple phase ramp improves upon removing the phase prior but remains less effective than APEX. The geometry-shuffled prior performs even worse than removing the phase prior. These results indicate that the geometric alignment of the proposed prior contributes to higher-frequency reconstruction beyond providing a generic phase-like conditioning signal.

Weakness 3 / Question 3

We have extended the ablation and target-frequency supervision experiments to SimpleWave and Maxwell, complementing the original Helmholtz results.

The additional ablation results in Appendix E.2 (Table 8) further demonstrate the contributions of both components across the three benchmarks. The coarse amplitude anchor supports overall complex-field reconstruction accuracy, while the phase prior provides additional guidance for recovering phase-sensitive oscillatory structures. The results on SimpleWave and Maxwell support the effectiveness of both components beyond the Helmholtz benchmark.

We have also extended the supervision-sensitivity analysis in Appendix E.4 (Table 10). Across all four HF supervision ratios, APEX achieves lower overall H¹ error and higher overall AWPC than CFM-Joint on both SimpleWave and Maxwell, consistent with the Helmholtz results in Table 3.

Regarding the evaluation scope, we acknowledge that the current experiments are limited to simulation-based benchmarks with fixed spatial grids. We have clarified this limitation in Sec. 6 and Appendix F, with real-world measurements, higher-resolution fields, and more complex propagation geometries identified as directions for future work.

Weakness 4 / Questions 4–5

We have added training-seed experiments on Maxwell in Appendix E.5 (Tables 11–12). Given the limited rebuttal period, we focus on HNO-Joint, which has the closest H¹ performance to APEX at Maxwell HF100 among the baselines in Table 1. Both methods are evaluated using three training seeds (0, 1, and 2). The following results report the mean ± standard deviation across training seeds.

MethodMetricHF20HF50HF100Overall
HNO-JointH¹↓0.863 ± 0.0290.915 ± 0.0220.962 ± 0.0240.937 ± 0.024
HNO-JointAWPC↑0.421 ± 0.0250.335 ± 0.0120.275 ± 0.0130.331 ± 0.016
APEXH¹↓0.710 ± 0.0530.796 ± 0.0350.950 ± 0.0290.882 ± 0.032
APEXAWPC↑0.813 ± 0.0270.724 ± 0.0340.633 ± 0.0090.706 ± 0.020

At Maxwell HF100, APEX consistently achieves lower H¹ error than HNO-Joint across all three training seeds, with a substantially higher AWPC. These results further support the stability of APEX under different training initializations. We have also added 95% bootstrap confidence intervals for the main comparisons across all three benchmarks in Appendix E.7 (Table 13).

We have additionally included a computational cost comparison in Appendix E.8 (Table 14), reporting total training time, per-sample inference latency, FLOPs, and peak allocated GPU memory on the Helmholtz benchmark. All measurements are conducted on an NVIDIA RTX 4090 GPU.

MethodTraining (s)Inference (ms)FLOPs/sampleMemory (GiB)
FNO-FT5033.5820.0139 G0.0613
HNO-FT8025.7900.0332 G0.0315
CFM-Joint16,7561,94318.798 T0.207
APEX1,4682,30118.930 T0.247

The training times include LF pretraining and HF fine-tuning for FNO-FT/HNO-FT, and LF pretraining followed by HF enhancement training for APEX. Both APEX and CFM-Joint use 50-step midpoint sampling at inference.

FNO-FT and HNO-FT offer substantially faster inference, but their reconstruction accuracy remains limited under scarce HF supervision. APEX uses additional generative inference computation to achieve higher prediction accuracy, reflecting an accuracy–computation trade-off.

Compared with CFM-Joint, APEX reduces total training time by approximately 11.4× while achieving better reconstruction accuracy. Its two-stage design learns transferable structural information from abundant LF data before training the generative enhancer on limited HF samples. This separation allows the enhancement stage to focus on target-frequency reconstruction, enabling more effective use of scarce HF supervision. Improving the sampling efficiency of the generative enhancer remains a direction for future work, as discussed in Appendix F.

Thank you very much for your constructive review. As the discussion period is nearing its end, we would greatly appreciate any feedback on whether your concerns have been addressed or if further clarification is needed. Thank you for your time and consideration.

Reviewer_Tg4kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is based on the motivation of amplitude-phase asymmetric behaviour, while a solution is proposed to use a phase prior combined with neural operators for representing high-frequency wave fields. The central finding about the asymmetric cross-frequency transfer reveals that the amplitude similarity is preserved across frequencies, while there is a drop in phase similarity. From this motivation, APEX is introduced that creates a phase prior based on the intuition of the Green’s function, which is then combined with the coarse predictions from the Fourier Neural Operator. These are used to condition the flow matching model that reconstructs the wave field. The results from the experiments done for three datasets show that APEX has outperformed the state-of-the-art algorithms.

Critical strengths
  1. The identification of the asymmetric cross-frequency transfer and empirical analysis. The empirical observations would provide valuable insights to understand the characteristics of the wave fields. Hence, it would be useful when designing effective representations for these fields.

  2. The use of coarse amplitude anchors as a pseudo ground truth. Since neural networks have limitations in high-frequency modelling, the research findings are useful to guide the learning process combined with a phase prior (Not specifically the introduced prior in this paper).

Critical weaknesses
  1. Although the phase prior is inspired by the Green’s function, the formulated prior is simplified, which the authors also agree. It should be shown the relationship between the proposed prior and the Green’s function and a justification for the simplification.

  2. The quantitative results for the usage of computational resources in the algorithm should be mentioned. At least the main characteristics, including training time, peak GPU usage, and FLOPs, should have been mentioned.

  3. The empirical result in Figure 01 reveals the amplitude-phase asymmetric trade-off in the ground truth Maxwell dataset. However, to strengthen it further, at least further empirical results should be included for multiple datasets.

  4. Minor weakness, figure 02 should be improved to be more professional. The design guideline box in lines 211-215 should also be consistent with the conference guidelines.

Questions for authors
  1. Provide quantitative results for computational resource usage for APEX and other compared models. At least find the basic metric values stated under weaknesses.
  2. To validate the generalizability of amplitude-phase asymmetric transferability, at least provide empirical evidence for the other datasets used in the paper.
  3. Need further explanations about the relationship between the proposed phase prior and the Green’s function. Need justification for the selection of the simplified prior.
3 comments

We thank the reviewer for recognizing the value of our amplitude–phase transfer analysis and for the constructive suggestions. We have revised the manuscript to clarify the Green's-function-inspired phase prior, added computational cost comparisons and cross-dataset validation, and improved the presentation of Fig. 2 and the Design Guideline. All modifications are highlighted in blue in the revised manuscript.

Weakness 1 / Question 3

We have expanded Sec. 3.3 to clarify the relationship between the proposed phase prior and the Green's function, as well as the motivation for its simplified construction.

The prior is motivated by the geometric propagation interpretation of the Green's-function representation in Eq. (2). When a small number of dominant propagation paths govern the main oscillatory behavior, the accumulated phase along each path can be approximated using an effective phase coefficient and the corresponding path length. Following this interpretation, Eq. (13) constructs a lightweight multipath surrogate that retains the dominant geometric phase structure, with its phase extracted through Eq. (14).

The simplified construction follows the role of the phase prior in APEX. Rather than approximating the complete Green's function or reproducing the full wave solution, the prior provides source- and geometry-aware oscillatory cues for the generative enhancer. The coarse amplitude anchor separately supplies transferable structural information, allowing the phase prior to focus on the dominant propagation geometry. This construction is deterministic, requires no trainable parameters, and avoids solving the full wave equation at the target frequency.

Weakness 2 / Question 1

We have added a quantitative computational cost comparison in Appendix E.8, reporting training time, inference latency, FLOPs, and peak allocated GPU memory for APEX and the main baselines on the Helmholtz benchmark. All measurements are conducted on an NVIDIA RTX 4090 GPU.

MethodTraining (s)Inference (ms)FLOPs/sampleMemory (GiB)
FNO-FT5033.5820.0139 G0.0613
HNO-FT8025.7900.0332 G0.0315
CFM-Joint16,7561,94318.798 T0.207
APEX1,4682,30118.930 T0.247

FNO-FT and HNO-FT offer substantially faster inference, but their prediction accuracy remains limited under scarce HF supervision. APEX introduces additional generative inference cost to achieve higher prediction accuracy. Compared with CFM-Joint, APEX achieves higher reconstruction accuracy while reducing total training time by approximately 11.4×. This improvement is supported by its two-stage design: the LF backbone first learns transferable structural information from abundant LF data, allowing the subsequent generative enhancer to focus on reconstructing HF details from limited target supervision. The resulting framework makes more effective use of scarce HF samples while avoiding the substantial training cost of joint generative modeling across frequency regimes. Improving the sampling efficiency of the generative enhancer remains a direction for future work, as discussed in Appendix F.

Weakness 3 / Question 2

We have extended the ground-truth cross-frequency similarity analysis to SimpleWave and Helmholtz in Appendix E.1. Across both additional benchmarks, amplitude similarity remains high across frequencies, while phase similarity decreases substantially with frequency separation, consistent with the observations on Maxwell. Together with the LF-trained FNO extrapolation results in Fig. 1(c), these findings provide empirical evidence of amplitude–phase asymmetric transferability across all three benchmarks.

Weakness 4

We have revised Fig. 2 to adopt a more professional and technical schematic style. The updated figure illustrates the two-stage architecture and the connections between its main components. Specifically, it includes the mathematical definitions of the extrapolation-derived amplitude anchor and the Green's-function-inspired phase prior, and shows how they are incorporated into the conditional flow-matching enhancer.

We have also revised the Design Guideline in Sec. 2.3, replacing the previous boxed presentation with a standard paragraph format consistent with the conference template.

Thank you very much for your constructive review. As the discussion period is nearing its end, we would greatly appreciate any feedback on whether your concerns have been addressed or if further clarification is needed. Thank you for your time and consideration.

Thank you for the extra experiments and clarifications. However, the justification for Q3 lacks a strong theoretical foundation. According to the answer to Q1, APEX overhead is higher than the gain in performance. Hence, I am not convinced enough to increase my score, and it will remain the same.

Reviewer_mxL2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies data-driven higher-frequency wave-field prediction when lower-frequency data are plentiful but target-frequency supervision is scarce. It argues that amplitude structure transfers across frequencies more reliably than phase-sensitive oscillations. APEX therefore retains the extrapolated lower-frequency operator's log-amplitude as an anchor and supplies a simplified Green's-function-inspired phase prior. A conditional flow-matching model combines these conditions to reconstruct the higher-frequency complex field.

Critical strengths

The method is motivated by a clear component-wise diagnosis of cross-frequency failure. Rather than treating higher-frequency degradation as a single scalar-error phenomenon, the paper separately measures amplitude similarity and phase coherence and shows that phase deteriorates much faster across frequency gaps.

Critical weaknesses

The key weakness of the paper is lack of clarity. The exact model being trained and the loss function needs to be clearly defined and the algorithm needs to be given as a pseudocode. I do not at all believe that someone can read Section 3 and reproduce the results -- which is what ideally is the level of details a methods section of the paper should provide.

Figure 2 is just a cartoon with no technical content in it.

Questions for authors

I believe the above stated weakness is fixable. I totally appreciate the mathematical clarity in Section 2 of the writing. So I totally believe the authors can put in equal effort to make their Section 3 equally clear.

Section 3 needs to be fully re-written keeping in mind a technical audience looking to reproduce the experiment. I can see a lot of concepts get defined in equations 12, 13 and 14, but I cant infer how does everything come together to define the new operator learning algorithm being proposed!

The so-called loss function in equation 18 is so vague : I have no idea what are the actual architectures being recommended for vψv_\psi and why the quantities defined in equations 15 and 16 are the way they are!

2 comments

We thank the reviewer for recognizing the component-wise analysis in Sec. 2 and for the constructive suggestions on improving the technical clarity of our method. We have revised Sec. 3 and Fig. 2 and added a training and inference pseudocode to make the proposed framework more explicit. All changes are highlighted in blue in the revised manuscript submitted alongside this response.

Q1. Complete algorithmic formulation

We have clarified how the individual components form a unified two-stage framework.

In Stage I, an FNO is pretrained on lower-frequency (LF) data and subsequently frozen. At a target higher frequency (HF), its extrapolated complex-field prediction is used to construct the coarse log-amplitude anchor (Eq. 12). In parallel, a deterministic Green's-function-inspired phase prior is constructed from the source and propagation geometry (Eqs. 13–14).

In Stage II, these two components, together with the environment and frequency embeddings, form the conditioning variables for a conditional flow-matching (CFM) enhancer (Eq. 16). The enhancer learns to reconstruct the target HF field, which is decoded into its complex-valued representation at inference (Eq. 21).

The complete training and inference procedure is now summarized in Algorithm 1, explicitly covering LF pretraining, HF enhancement training, and HF inference. The corresponding formulation is provided in Secs. 3.2–3.4.

Q2. Model architecture, representation, and training objective

We have expanded Sec. 3.4 to clarify the target representation, network architecture, and optimization objective.

Eq. (15) represents the target complex field using log-amplitude and sine–cosine phase components. The log transformation compresses the amplitude's dynamic range, while the sine–cosine representation avoids discontinuities at the phase-wrapping boundary. Eq. (16) explicitly defines the conditioning variables, combining the coarse amplitude anchor, phase-prior features, environment information, and frequency embeddings.

The enhancer is parameterized by a residual U-Net with bottleneck attention. The amplitude anchor and phase-prior maps are processed through separate spatial encoders and fused with U-Net features at multiple resolutions. The resulting features are modulated through FiLM conditioning, as explicitly formulated in Eq. (17).

For training, we explicitly define the linear flow interpolation in Eq. (18) and the CFM training objective in Eq. (19), where the network predicts the velocity target between Gaussian noise and the target field. During HF enhancement, the LF operator remains frozen, and only the enhancer parameters are optimized. We further specify conditioning dropout during training and the classifier-free-guided flow ODE used for inference in Eq. (20).

Additional architecture configurations and training settings are provided in Appendix C, with benchmark-specific phase-prior constructions detailed in Appendix D.

Q3. Technical clarity of Figure 2

We have revised Fig. 2 to improve its presentation style. The updated figure adopts a more technical schematic style to illustrate the two-stage architecture and the connections between its main components. Specifically, it includes the mathematical definitions of the extrapolation-derived amplitude anchor and the Green's-function-inspired phase prior, and shows how they are incorporated into the conditional flow-matching enhancer.

Thank you very much for your constructive review. As the discussion period is nearing its end, we would greatly appreciate any feedback on whether your concerns have been addressed or if further clarification is needed. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.