{"id":"400fe330-62fd-4ed2-b6bc-0f4394eead0c","arxiv_id":"2607.29505","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Replacing scalar Fraunhofer/Debye propagation with a differentiable Richards-Wolf model during phase hologram optimization yields nearly perfect flat-top and tweezer uniformity at NA=0.9 and enables direct shaping of polarization-dependent optical dipole potentials.","lead":"This computational paper shows that optimizing computer-generated holograms with a vectorial Richards-Wolf light-propagation model, rather than scalar approximations, yields nearly perfect intensity patterns at high numerical aperture. The same differentiable framework can shape polarization-dependent optical trap potentials, not just intensity, which matters for high-NA neutral-atom experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline uniformity values are self-consistent within the ideal RW model; the key open question is whether they survive realistic SLM/objective perturbations, so an independent sensitivity test is needed.","rationale":"The reader correctly identifies the same-model evaluation as the weakest assumption. My concern is a sharper version: the reported 99.97%/99.98% values are not merely unvalidated experimentally; they are the outcome of optimizing with the exact same differentiable operator used for evaluation, so they may reflect model-specific fitting rather than physical robustness. The paper's internal checks—low-NA limit, component energies, dense-Fourier cross-check, zero-padding convergence—are commendable and rule out numerical implementation errors, but they do not address model mismatch. The authors are transparent about the ideal assumptions, so this is not a fraud or overreach. Still, the central claim's practical significance depends on whether the optimized holograms tolerate realistic deviations from those assumptions. The proposed perturbation test is a concrete, computationally feasible way to decide whether the concern lands. It does not require a full experiment immediately, though an experimental measurement would be the definitive test. Since the reader's verdict is already CONDITIONAL, my analysis supports keeping that verdict rather than upgrading or rejecting. My agreement is partial because the reader's weak assumption is broader (experimental validation) while I focus on a specific sensitivity test that could be performed in silico before any experiment.","tokens_in":19445,"tokens_out":6368,"duration_ms":73820,"concrete_test":"Take the NA=0.9 RW-optimized phases from the paper (or regenerate them using the described pipeline) and evaluate them with an independent vectorial Debye-Wolf/RW implementation that adds realistic perturbations: (1) 0.05λ RMS low-order Zernike aberrations; (2) Gaussian pupil illumination with 1/e radius at 80% of the pupil radius plus 10% amplitude clipping; (3) 8-bit phase quantization with simple nearest-neighbor SLM pixel crosstalk. Recompute u_flat, PV_flat, u_tweezer, and ε. If uniformity stays within 0.5 percentage points and ε remains below 1.02, the central claim is robust. If uniformity drops by more than 1 point or ε exceeds 1.05, the headline metrics are model-specific, and the paper should be explicitly conditional on experimental validation or on demonstrating robustness to such perturbations.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central numerical claim—RW optimization reaches 99.97% flat-top and 99.98% tweezer uniformity at NA=0.9—is established by optimizing and evaluating with the same differentiable Richards-Wolf model. This makes the result a self-consistent model-fitting statement rather than a direct statement about physical focusing. The paper's validation checks are strong but all test internal numerical consistency: low-NA scalar limit, phase-independent component energies, dense-Fourier vs FFT, and zero-padding convergence. None of these tests probes whether the optimized phase remains high-fidelity when the physical system deviates from the ideal assumptions stated in the paper—uniform pupil amplitude, ideal phase-only SLM, n=1, and aplanatic objective obeying the sine condition. Because the optimizer has ~1.2e5 phase degrees of freedom and the loss is computed with the same operator used for evaluation, the reported metrics can exploit model-specific degrees of freedom (e.g., exact apodization, PSF shape, or high-frequency pupil phase) that may be fragile to SLM phase quantization, pixel crosstalk, aberrations, pupil illumination nonuniformity, or polarization non-idealities. The authors explicitly acknowledge this limitation in the conclusions, so this is not an internal inconsistency, but it is the load-bearing uncertainty for the paper's broader claim that RW-in-the-loop 'reframes high-NA CGH as optimization of physically relevant electric-field functionals.' If the optimized phases are not robust to small model perturbations, the headline uniformities are numerical artifacts of the ideal RW model rather than predictions for a real high-NA system.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a differentiable Richards-Wolf (RW) forward model for gradient-based phase-only CGH optimization. It compares RW, scalar Debye, and Fraunhofer optimization under a common RW evaluation for two targets (extended flat-top and 10×10 tweezer array), showing that at NA=0.9 RW optimization achieves 99.97% flat-top uniformity and 99.98% tweezer uniformity with mean ellipticity ε=1.007, while scalar-optimized phases exhibit target-dependent degradation (flat-top error from aplanatic weighting; tweezer ellipticity from polarization mixing). The paper further optimizes a polarization-dependent optical dipole potential directly from the RW field, reducing the mean normalized residual by about 11× compared with intensity-proxy optimization. The Supplement provides extensive internal validation: low-NA scalar-limit convergence, analytic component-energy fractions, dense Fourier-sum cross-check, zero-padding convergence, and optimizer/normalization sensitivity.","tokens_in":19790,"tokens_out":15540,"duration_ms":156484,"significance":"If the numerical results are robust, this is a useful contribution: it demonstrates a computationally efficient (FFT-based) vectorial propagation model in an automatic-differentiation CGH pipeline and clearly separates two distinct scalar-model errors. The validation is unusually careful for a computational optics paper, including analytic component-energy checks and a dense-matrix cross-validation of the FFT implementation. However, the headline fidelity numbers are self-consistent within the ideal RW model; no perturbation study shows how they degrade under realistic SLM/objective non-idealities. The paper explicitly acknowledges these limitations, so the main uncertainty is the practical reach of the claims, not their internal consistency.","major_comments":[{"comment":"All reported fidelities in Table 1 (99.97% flat-top, 99.98% tweezer uniformity, ε=1.007 at NA=0.9) are computed with the same differentiable RW operator used for optimization. With ~1.2×10^5 phase DOF, the optimizer may fit model-specific degrees of freedom (pupil apodization, exact PSF shape, high-frequency phase). The Supplement's validation checks are internal-consistency tests (low-NA limit, analytic component energies, dense Fourier sum, zero-padding) and do not probe sensitivity to deviations from the ideal assumptions stated in §2: uniform pupil amplitude, ideal phase-only SLM, n=1, and aplanatic objective. Since the paper frames the method as addressing a critical limitation in high-NA optical systems, please add a perturbation study (e.g., low-order aberrations, pupil apodization/clipping, SLM phase quantization, or an independent vectorial propagator) and report the resulting T","section":"§3, Table 1 and Eq. (1)"},{"comment":"The factor-of-11 reduction in mean normalized potential residual (§4) is the key demonstration for polarization-sensitive targets, but the axial-stationarity weight Λ_z=0.4 in Eq. (11) is fixed ad hoc. No sensitivity of the reported residual (2.0×10^-2 to 1.8×10^-3) to Λ_z is shown. If the improvement over the scalar-only RW control is robust only for a narrow Λ_z range, the claim of direct RW-potential optimization is weakened. Please include a Λ_z sweep (or a justification of the chosen value) for the three optimizations compared.","section":"§4, Eq. (11)"}],"minor_comments":[{"comment":"The flat-top uniformity and PV metrics are evaluated only over SRflat, which excludes the steep target edges by construction. A solution with a perfect central plateau but wrong edge profile would score 99.97% uniformity. Although the full target enters the loss in Eq. (5), the reported headline metrics do not measure edge fidelity. Please report a full-region metric (e.g., RMSE over the entire 40 r_Airy mask) alongside the plateau metrics to substantiate the flat-top generation claim.","section":"§2.1, Eqs. (6)-(8)"},{"comment":"Data availability states that data are not publicly available but may be obtained from the authors. For a computational methods paper, releasing the JAX implementation and the optimization scripts would substantially improve reproducibility and allow readers to test the sensitivity concerns raised above.","section":"Data availability"},{"comment":"The legend contains a stray '1' at the bottom right, apparently a LaTeX artifact. Please correct.","section":"Fig. 2"},{"comment":"The vortex-removal restart in Supplement §6 is described only briefly and is not used in the main benchmarks. The heuristic scalar-proxy inverse-FFT initialization works well for the asymmetric split flat-top, but its general applicability should be discussed, or the main-text claims should be limited to the target classes studied.","section":"Supplement 1, §6"},{"comment":"The axial scale z0 = λ/(2 NA^2) is introduced without explanation; it is the Rayleigh-like axial scale of the focal field. Please define it explicitly to aid readability.","section":"Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound and the supplement is exemplary in its internal validation. The main concern is the same-model optimization/evaluation loop, which makes the headline 99.97%/99.98% numbers a statement about the optimizer and the ideal RW model rather than about a physical focusing system. I recommend asking the authors to add a robustness/perturbation study (aberrations, pupil non-uniformity, SLM phase quantization) and a Λ_z sensitivity check. If they prefer to keep the claims purely computational, the practical framing should be softened. I would also encourage code release, as the paper is otherwise highly reproducible in methodology."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper is a careful numerical study: the RW implementation is validated against low-NA scalar limits, analytic component-energy fractions, dense Fourier sums, zero-padding convergence, and the optimizer/normalization choice is tested. The genuinely new piece is using the full vectorial field inside the loss to optimize a polarization-dependent optical dipole potential—vector and tensor light shifts—rather than just total intensity. That is a real capability for cold-atom platforms, and the order-of-magnitude residual reduction in that demo is credible.\n\nThe headline uniformities (99.97% flat-top, 99.98% tweezer at NA=0.9) are self-consistency results within an ideal Richards-Wolf model. The same operator that optimizes the phase also evaluates the fidelity. With ~1.2e5 phase degrees of freedom, the optimizer can exploit model-specific degrees of freedom. The paper acknowledges the ideal assumptions at the end—uniform pupil, ideal phase-only SLM, n=1, aplanatic objective—but it does not test what happens under realistic perturbations. A simple robustness check (SLM phase quantization, pupil illumination nonuniformity, residual aberration, or a polarization error) would tell whether the phases survive contact with a real objective. That is the load-bearing uncertainty, and it is not resolved here.\n\nTwo more gaps. There is no comparison with existing vectorial beam-shaping algorithms (generalized projection for high-NA flat-tops, for example), so I cannot tell whether the RW-in-the-loop approach is a meaningful advance or just a different optimizer on the same forward model. And no code or data are released, despite the Data availability statement saying they may be obtained on request; for a numerical method paper, that limits reproducibility.\n\nWho this is for: groups doing high-NA holographic traps, and people working on differentiable optics. It deserves a serious referee. The numerical core is solid and the potential objective is novel enough to warrant referee time, but I would not accept the headline figures as predictions about a physical system until the robustness work is done. If the revision adds a perturbation sensitivity study and one or two comparisons with prior vectorial methods, the conditional concerns largely drop away.","headline":"RW-in-the-loop CGH is a genuinely useful numerical idea, but the headline uniformities are model-self-consistency numbers, and the paper needs perturbation sensitivity before claiming to reframe high-NA CGH.","tokens_in":20323,"tokens_out":3056,"would_cite":false,"duration_ms":32498,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Optimizing holograms with a differentiable Richards-Wolf vector-field model achieves 99.98 percent uniformity at NA=0.9.","keywords":["computer-generated holograms","Richards-Wolf model","high numerical aperture","optical dipole traps","phase-only SLM","vectorial beam shaping","gradient-based optimization","polarization-dependent optical potential"],"falsifier":"Measure the same 10x10 tweezer array and flat-top on a real NA=0.9 aplanatic objective with a phase-only SLM: if RW-optimized holograms do not visibly outperform Fraunhofer-optimized ones in measured uniformity and ellipticity, or if feeding an independently measured pupil amplitude and aberration map into the same RW loss does not preserve the reported fidelities, the central claim is falsified.","tokens_in":19306,"feed_emoji":"🔬","tokens_out":6382,"duration_ms":67748,"temperature":0.7,"pith_summary":"The paper claims that the standard scalar diffraction models used to design computer-generated holograms break down at high numerical aperture in target-specific ways, and that replacing them inside the optimization loop with a differentiable Richards-Wolf vectorial forward model fixes both failure modes. Optimizing the SLM phase with this model yields flat-top and tweezer-array uniformities of 99.97% and 99.98% at NA=0.9, with nearly circular tweezers, whereas Fraunhofer-optimized holograms degrade due to missing aplanatic apodization and Debye-optimized ones due to neglected polarization mixing. Because the model returns all three electric-field components, the same pipeline can optimize not just intensity but the actual optical dipole potential felt by an atom, including vector and tensor light shifts. This matters for high-NA neutral-atom traps and other tightly focused beam-shaping applications, where trap depth, differential shifts, and trap shape depend on local polarization, not on intensity alone.","feed_headline":"Vectorial optimization hits 99.98% hologram uniformity at high NA","feed_subtitle":"Using the full Richards-Wolf vector field in the optimization loop beats scalar models for tweezers and flat-top traps.","key_machinery":"The workhorse is an FFT-based, differentiable implementation of the Richards-Wolf focusing integral, with the aplanatic apodization factor 1/sqrt(cos theta), the pupil-to-spatial-frequency mapping, and the 3x3 polarization transformation matrix that mixes the incident transverse polarization into cross-polarized and longitudinal focal components. Because the model is differentiable, the squared-error loss between target and simulated intensity (or simulated optical potential) is minimized directly by a quasi-Newton optimizer using automatic differentiation. The same model supplies the full local field E, so the loss can include vector and tensor light-shift terms that depend on Im(E* x E) an","core_discovery":"The central discovery is that gradient-based phase retrieval, when driven by an FFT-accelerated Richards-Wolf forward model rather than scalar Fraunhofer or Debye propagation, produces holograms whose RW-evaluated total-intensity profiles are essentially uniform in the deep non-paraxial regime. At NA=0.9 the flat-top uniformity reaches 99.97% (peak-to-valley variation 0.29%) and the 10x10 tweezer array reaches 99.98% with mean ellipticity 1.007. The same vectorial field enables a previously unavailable objective: optimizing the polarization-dependent optical dipole potential directly, which the paper demonstrates reduces the mean normalized potential-shape residual by about a factor of 11 co","pith_inferences":["If the ideal-aplanatic assumption holds, the same differentiable vectorial pipeline could be dropped into other high-NA beam-shaping tasks—STED, two-photon, or optical trapping—where scalar optimization currently produces focal-spot asymmetry.","The framework's loss is a template: substituting measured pupil amplitude and aberration maps from a real objective would turn the simulation-based benchmarks into a calibration-aware hologram generator, effectively testing whether model mismatch or optimization error dominates in practice.","Because the polarization-overlap factor is quadratic in intensity spatial frequency and anisotropic along the incident polarization, the largest fidelity gain from RW optimization should appear for targets with fine features oriented along the polarization direction; a systematic orientation sweep of an asymmetric target would probe this.","For neutral-atom platforms, the same loss could be re-targeted to minimize differential light shifts between atomic states, which the paper names as the natural next step."],"forward_implications":["At NA=0.9, RW-optimized holograms achieve flat-top uniformity 99.97% with 0.29% peak-to-valley variation and tweezer-array uniformity 99.98% with mean ellipticity 1.007, under RW evaluation.","Fraunhofer-based design loses flat-top fidelity mainly because it omits the aplanatic pupil weighting; scalar Debye restores the flat-top but leaves diffraction-scale tweezers elliptical (about 1.16 at NA=0.9) because it ignores focal polarization mixing.","RW optimization keeps the favorable FFT scaling of the forward propagation; the full vectorial gradient computation costs roughly 2.4 times the scalar per-call time, and the full quasi-Newton optimization about 1.7 times, so the vectorial model does not change the asymptotic complexity.","For polarization-sensitive optical dipole traps, optimizing the full RW potential rather than an intensity proxy reduces the mean normalized residual from 2.0e-2 to 1.8e-3 (roughly 11x) and raises the Pearson correlation from 0.9780 to 0.9998.","The reported fidelities are computational benchmarks for an ideal phase-only SLM, uniform pupil illumination, and an aplanatic objective; experimental use requires calibration of aberrations, pupil amplitude, SLM pixel response, and polarization optics."],"fun_headline_variants":["Vectorial holograms hit 99.98% uniformity at NA=0.9","Richards-Wolf model outperforms scalar in high-NA holography","Phase-only vectorial optimization fixes high-NA fidelity","High-NA holograms: vectorial beats scalar for tweezers and flat-tops","Vectorial field optimization cuts dipole-potential residual 11x"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"All fidelity values are computed with the same Richards-Wolf model that is being optimized, under ideal assumptions—uniform pupil amplitude, phase-only SLM, refractive index 1, and an aplanatic objective obeying the sine condition—so the numbers are model self-consistency, not a measured property of a real optical system.","fun_headline_variants_meta":{"raw":{"variants":["Vectorial holograms hit 99.98% uniformity at NA=0.9","Richards-Wolf model outperforms scalar in high-NA holography","Phase-only vectorial optimization fixes high-NA fidelity","High-NA holograms: vectorial beats scalar for tweezers and flat-tops","Vectorial field optimization cuts dipole-potential residual 11x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1230,"prompt_tokens":752,"completion_tokens":478,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":380}},"tokens_in":496,"tokens_out":478,"duration_ms":5177,"temperature":1.0,"reasoning_tokens":380,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:37:15.681740+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the same 10x10 tweezer array and flat-top on a real NA=0.9 aplanatic objective with a phase-only SLM: if RW-optimized holograms do not visibly outperform Fraunhofer-optimized ones in measured uniformity and ellipticity, or if feeding an independently measured pupil amplitude and aberration map into the same RW loss does not preserve the reported fidelities, the central claim is falsified.","supporting_citations":[],"review_version":1}