acceptodds
Withdrawn from ICLR 2027

HoloFNO: A Physics-Supervised Neural Operator for Low-Latency Holographic Phase Synthesis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Fast synthesis of phase-only holograms is critical for programmable optical trapping, for which classic iterative phase-retrieval algorithms need repeated forward-backward propagation and constraint-projection steps. We formulate holographic phase synthesis as a physics-supervised inverse-design problem: the network predicts a spatial light modulator (or SLM) phase mask and is optimized through a differentiable Fourier-optics model against the reconstructed target-plane intensity, without requiring Gerchberg-Saxton (GS) generated phase labels. We investigate this problem at a fixed native discretization resolution of pixels across both the SLM aperture and the focal plane, and introduce HoloFNO, a Fourier neural operator augmented with a harmonic-oscillator-inspired multi-frequency sinusoidal bypass. To uncover the internal representation mechanics and validate that the operator learns genuine diffraction physics rather than superficial heuristics, we perform fixed-resolution mechanistic evaluations consisting of branch-scrubbing interventions, low-order Zernike-residual decomposition, and mechanistic diagnostics across 3 independent training runs and 5 shared evaluation draws per run. Diagnostic probes demonstrate that the selected topology exhibits a reproducible phase content. At its 60-iteration quality-matched point, HoloFNO leads GS on reconstruction fidelity within the region of interest, quantified as the relative root-mean-square error (Rel-RMSE) between the reconstructed and ideal target intensity patterns ( vs. for HoloFNO), while HoloFNO's fixed one-shot cost eliminates the per-sample runtime variability inherent to GS's convergence-dependent iteration count.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operators, Fourier Optics, Physics-Supervised Learning, Learned Representation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_HjUyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper aims to estimate phase of a spatial light modulator (SLM) for atom trapping.

Critical strengths

None

Critical weaknesses

The organization of the paper requires substantial restructure and rewrite. The current version of the paper seems highly AI generated without human proof-reading and modification. I will point out some items that I think the authors can work on for the future version (if they plan to submit again).

  1. The first paragraph of the introduction should clarify the motivation and application (i.e., atom trapping) and provide context of why this is important (e.g., for quantum simulation). The current version is like spaghetti --- each sentence follows grammatical rules while the cohesiveness is extremely low. Because the first paragraph does not serve its role in the current version, the second and onward paragraphs make me lost.

  2. Additionally, the computation paragraphs in the introduction can be considered to move to related work. Introduction and related work should serve distinct roles!

  3. There are LOTS of figures and tables (e.g., Figs. 1-3 and Tabs. 2-3) are not referenced, which make readers hard to following the paper arguments.

  4. I understand that page limit stops the authors including details of the loss function in matin text, but it is highly recommended to provide the function and/or motivation for each of the loss terms in main text.

Questions for authors
  1. In line 140, what is E(r)\mathcal{E}(\textbf{r})? What is different between it and E(u,v)\mathcal{E}(u,v)?
Reviewer_f8ujOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper uses a Fourier neural network to generate phase masks for holograms. It trains the network by comparing the simulated light pattern with the target, rather than copying phase masks produced by GS. The authors add a sinusoidal branch and test its effect on reconstruction quality and runtime.

Critical strengths

Training on the resulting light pattern makes sense because different phase masks can produce similar patterns. The paper evaluates both reconstruction error and simulated trapping quality.

The authors test what happens when the sinusoidal branch is removed and repeat training for the two main variants. They also discuss the limits of these tests and the lack of hardware experiments.

Critical weaknesses

It is unclear whether the sinusoidal branch helps. The simpler model in Table 1 has a lower reconstruction error than HoloFNO. that means if the settings are the same, this weakens the case for the added branch. Removing it after training shows that the network uses it, but does not show that it is better than simpler alternatives or that it learns the claimed physics.

The runtime results do not show a clear advantage over GS. GS can also run for a fixed number of iterations, so predictable runtime is not unique to HoloFNO. The tests cover a narrow set of simulated targets, with only four OOD examples and no optical experiment.

The writing needs substantial editing. There are too many lists, repeated explanations, and qualifications, which make the main contribution harder to follow.

Questions for authors

Were Tables 1 and 5 evaluated under the same settings? If so, why use the sinusoidal branch when the simpler model has lower error?

Does the proposed branch outperform a generic sinusoidal branch or the standard FNO pointwise branch when all models are trained from scratch?

What quality does GS achieve if it gets the same runtime as HoloFNO?

What shows that the model learns diffraction physics rather than patterns specific to the training data? Removing a branch does not answer this question.

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

The paper studies fast phase-only hologram synthesis for blue-detuned optical traps, with targets built from superpositions of 2D quantum-harmonic-oscillator eigenstates. A Fourier neural operator with a multi-frequency sinusoidal bypass is trained through a differentiable Fourier-optics model, without GS phase labels, at a fixed 128×128 resolution. Branch-scrubbing and Zernike-residual diagnostics are used to analyze the bypass and select the final topology. HoloFNO reaches an RMSE_ROI of 0.0883, compared with 0.0912 for the strongest U-Net baseline, and runs in about 5.2 ms per mask, compared with 3.37 ms for quality-matched GS at 60 iterations.

Critical strengths

Relevant application. Fast hologram computation is a practical bottleneck in neutral-atom experiments.

Label-free training. Supervising through the forward model avoids fitting one GS solution of a non-unique problem.

Critical weaknesses
  • Positioning relative to learned CGH. Training hologram networks through a differentiable propagation model, without iterative phase labels, has been widely explored in computer-generated holography, e.g., HoloNet (Peng et al., SIGGRAPH Asia 2020), tensor holography (Shi et al., Nature 2021), end-to-end learning of 3D phase-only holograms (Shi et al., Light Sci. Appl. 2022), and more recent learned random-phase holography (Chu et al., ACM TOG 2025). It would help if the authors could discuss these works and clarify what Contribution 1 adds beyond them.

  • Latency claim. At the tested resolution, quality-matched GS appears faster than HoloFNO (3.37 ms vs. 5.22 ms) and reaches a higher κ_geom (0.890 vs. 0.787) (Table 12, Sec. 8). The authors may wish to clarify how the "low-latency" framing in the title and abstract should be read in light of this, given that the resolution trend in Fig. 11 is extrapolated. Contribution 1 reports 5.35 ms while Sec. 8 reports 5.219 ms; could the authors clarify which value is correct? And since the latency is emphasized in title and abstract, the authors may need to move the relevant evidence into the main paper rather keeping it in the appdendix.

  • Clarity of presentation. The paper is somewhat hard to follow, as key components are spread across the main text and appendices. Fig. 1 shows a single block but not the full pipeline, e.g., the lifting and projection layers, the phase bias, the wrapping step, the differentiable forward model, and the loss. A figure illustrating the complete architecture and training loop would make the method much easier to understand.

  • Training objective and baselines. On the same U-Net, NPCC gives a lower RMSE_ROI than the proposed five-term loss (0.0912 vs. 0.1043, Table 5), while HoloFNO is trained only with the five-term loss. Could the authors discuss this, and report HoloFNO trained with NPCC?

  • Reliability of κ_geom. In the loss ablation, κ_geom for the full model is 0.169 ± 0.844, and the authors note it is unstable (Table 7). In Tables 3 and 5, however, κ_geom and Δκ_geom are reported with much smaller deviations (e.g., ±0.005) and used as key evidence. Could the authors report per-sample deviations here and discuss how reliable this metric is? It may also help to comment on why U-Net (Yan) is closest to the target value of 1 (0.971, the thrid line in Table 5).

  • Data and evaluation details. Samples are kept only if GS reconstructs them well (Sec. 4.2), so the test set appears biased toward GS-solvable targets; the effect of this filtering on the GS comparison could be discussed. Sec. 4.2 reports 200 test samples, while the tables use five sets of 100 (500 in total); could the authors clarify how these relate? Some details also seem missing, including the distribution of the superposition coefficients, the threshold and σ_blur of the envelope, and the latent width d.

Questions for authors
  • How does the method compare with a HoloNet- or tensor-holography-style network trained on the same task?
  • Could the authors report the significance of the gap over U-Net (NPCC) using the same type of standard deviation for all models?
  • What are the parameter counts and FLOPs of HoloFNO and each baseline?
  • How does HoloFNO perform when trained with NPCC?
  • How do the results change on targets not filtered by GS?
  • What experimental scenario motivates building the target intensity from the post-measurement density?

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

Related papers

Discussion (0)

Sign in to comment.