{"id":"83f7a5f6-728a-4bc2-8063-6b14e41065b6","arxiv_id":"2607.18012","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A PINN trained on Monge-Ampère equations designs phase-only optics for beam shaping, outperforming Gerchberg-Saxton in a far-field test.","lead":"A neural network was trained to design phase masks that reshape laser beams into target patterns by solving Monge-Ampère equations from geometric optics. It offers a new computational tool for flat-optics design that could replace iterative algorithms like Gerchberg-Saxton, with potential applications in laser machining, displays, and metasurfaces.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing transport boundary condition in Monge–Ampère PINN may allow non-onto ray maps; admissibility checks don't guarantee global energy conservation or target coverage.","rationale":"This paper presents a plausible new application of PINNs to flat-optics beam shaping, with a standard derivation and independent diffraction validation. The strongest claim is that the PINN solves the Monge–Ampère ray-mapping equation and outperforms Gerchberg–Saxton. After reading carefully, the most load-bearing soft spot is the absence of a boundary condition in the Monge–Ampère formulation. A solution to the interior PDE is not guaranteed to map the source aperture onto the entire target domain; the transport boundary condition is essential. The paper's §3.2 acknowledges that admissibility is only checked post-hoc, which cannot enforce global energy conservation or target coverage. This is not an internal inconsistency but a structural gap in the method: it may work for the reported smooth examples, but the central claim of a general solver is not fully supported. The Gerchberg–Saxton comparison using only 60 iterations is also a weakness, but it is secondary because if the PINN's ray map is not the true transport map, the comparison is undermined regardless of GS's iteration count. The reader identified exactly this issue as the weakest assumption. A concrete test—adding the missing boundary condition and measuring the effect—is the natural next step and would settle whether the concern actually lands. Since the reader's verdict is already CONDITIONAL and this test is precisely what the condition should require, the verdict need not change; the condition should be made explicit. No other concern appears more load-bearing. There is no evidence of dishonesty; the critique is purely technical.","tokens_in":6161,"tokens_out":3541,"duration_ms":40908,"concrete_test":"Retrain the far-field DT-logo network with an explicit boundary-condition loss: for boundary collocation points on the aperture edge, penalize deviations of the mapped direction (Φ_x, Φ_y) from the target-support boundary, and add a global flux-matching term |∫I_source dxdy − ∫I_target(Φ) dαdβ|. Then compare the diffraction-evaluated RRMS and target-support coverage (e.g., fraction of target pixels above 50% of desired intensity) with the current model. If the boundary-conditioned model yields materially lower RRMS or better coverage, the missing BC is the cause; if not, the admissibility checks are empirically sufficient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper solves the Monge–Ampère equation (Eqs. 29, 37) without imposing any boundary condition; §3.2 explicitly states that no Dirichlet condition is imposed and only post-hoc checks are used (mapped points remain in the target domain, K real). In optimal transport, the boundary condition — mapping the source aperture boundary to the target domain boundary — is required for the solution to be a genuine transport map that conserves global energy and illuminates the full target support. Without it, a network can minimize the pointwise residual in Eq. (41) while producing a ray map that is not onto the target domain: some target regions may be uncovered, or source power may be delivered multiple times to the same target region. The post-hoc checks do not detect such deficiencies because they only verify local admissibility, not global flux balance. The reported energy efficiency ~99.98% and diffraction agreement suggest that, for the tested smooth targets, the network happened to find a boundary-satisfying solution implicitly, but there is no mechanism forcing this. This is a structural gap in the PDE formulation and undermines the general claim that the method solves the beam-shaping problem, not merely the specific examples.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a physics-informed neural network (PINN) approach for designing phase-only flat-optics elements that reshape an input beam into prescribed intensity distributions. The phase profile is represented by a fully connected network and trained by minimizing residuals of Monge–Ampère equations derived from generalized-Snell-law ray mappings, for both finite-distance targets (Eq. 29) and far-field angular targets (Eq. 37). The learned phases are validated with independent scalar diffraction simulations for a Gaussian-to-star finite-distance example and a Gaussian-to-logo far-field example, and the far-field result is compared against 60 iterations of the Gerchberg–Saxton algorithm. The central claims are that this is the first PINN application to flat-optics beam shaping and that it achieves lower error and higher energy efficiency than GS for the tested target.","tokens_in":6443,"tokens_out":4839,"duration_ms":51276,"significance":"If fully supported, the method would be a useful addition to flat-optics design: it replaces iterative phase retrieval with a PDE-constrained neural optimization, handles non-cylindrically symmetric targets, and the authors provide open code and independent diffraction validation. The ray-mapping derivation appears self-consistent, and the normalization of the equations is nontrivial and largely convincing. However, the central claim is currently weakened by a structural gap in the PDE formulation: no boundary condition is imposed on the Monge–Ampère equation. Because the method is demonstrated on only one target per geometry and the GS baseline is not shown to be converged, the generality and superiority claims are not yet established at the level required for publication.","major_comments":[{"comment":"The Monge–Ampère equation is solved without any boundary condition. The manuscript explicitly states that no Dirichlet condition is imposed and that admissibility is only checked after training by verifying that mapped points remain in the target domain and that K is real. This is not sufficient: in optimal transport, the correct boundary condition—typically that the source aperture boundary maps onto the target boundary—is required for the solution to be a genuine transport map that covers the full target and conserves energy globally. A network can minimize the interior residual in Eq. (41) while producing a ray map that is not onto, leaving target regions unilluminated or delivering source power multiple times to the same region. The post-hoc checks are local admissibility checks, not global flux-balance checks. The reported 99.98% energy efficiency comes from wave-optics simulations","section":"§3.2, Eqs. (29), (37)"},{"comment":"The comparison with Gerchberg–Saxton is weakened by the use of only 60 iterations with no reported convergence criterion. GS is known to converge slowly for structured targets such as a two-letter logo, so an underconverged baseline is not a fair reference for the claimed superiority (relative RMS 2.97% vs 22.6%, energy efficiency 99.98% vs 95.90%). Please report the error as a function of iteration count and either run GS until a convergence metric stagnates or state the stopping rule. Without this, the quantitative superiority claim is not decisive.","section":"§4.3"},{"comment":"The paper's general claim—that the PINN approach solves beam-shaping problems in flat optics—rests on only two numerical examples, one per formulation. While the finite-distance star example is smooth and the far-field logo example has internal strokes, a single realization per geometry is not enough to establish that the method reliably produces admissible, boundary-satisfying transport maps, especially given the missing boundary condition. I recommend adding at least one additional target per formulation, or tempering the conclusions to a proof-of-concept demonstration. This is load-bearing for the stated scope of the method.","section":"§4.1–§4.2, §5"}],"minor_comments":[{"comment":"The notation D_near in Eq. (39) is introduced without a formal definition. It is presumably the left-hand side of Eq. (29) before multiplying by K^4; please define it explicitly to avoid ambiguity.","section":"§3.2"},{"comment":"The MSE values (3.59e6 and 2.08e8) are reported without normalization or units. Please report per-pixel or normalized MSE and specify the integration domain, so the comparison is reproducible.","section":"§4.3"},{"comment":"The panels in Fig. 5 are not individually labeled or captioned. Adding labels (a)–(d) or similar would greatly improve readability, especially since the figure is central to the validation claim.","section":"Fig. 5"},{"comment":"The derivation of the Monge–Ampère beam-shaping equation cites the author's own project report [2] and a PhD thesis [3]. For a journal submission, please also cite peer-reviewed sources for the ray-mapping/Monge–Ampère formulation, e.g., standard references on nonimaging optics or phase retrieval.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting and the code release plus independent diffraction validation are strengths. The missing boundary condition is the main technical issue; it is fixable within the scope of the paper by adding a boundary constraint and/or a rigorous energy-coverage diagnostic. The GS comparison also needs more care. I do not see a fundamental flaw that would require rejection, but the current version overclaims generality relative to the evidence presented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a genuinely new application — a PINN solving the Monge–Ampère ray-mapping equation for flat-optics beam shaping, including non-axisymmetric targets, with independent scalar diffraction validation for both examples. The code and data are on GitHub. That puts it ahead of many papers in this area.\n\nWhat it does well: the equations are derived carefully from the generalized Snell law, the normalization is sensible, and the network is trained entirely on the PDE residual, so no circularity from data labels. The far-field DT logo is compared against Gerchberg–Saxton, and the PINN phase indeed produces lower RMS error (2.97% vs 22.6%) and visibly less speckle. The independent diffraction check is the right validation and it passes.\n\nThe soft spot is exactly what the stress-test note flags. The Monge–Ampère formulation is an optimal transport problem, and without a boundary condition mapping the source aperture boundary onto the target boundary, the solution can be non-onto — some target regions unilluminated, some source power double-counted. The paper explicitly imposes no Dirichlet condition, only post-hoc checks: mapped points inside the target, K real. Those checks are local; they do not enforce global flux balance or target coverage. So the PDE formulation is structurally incomplete. The diffraction simulations show the method worked for the two tested targets, which suggests the network implicitly found a boundary-satisfying solution. But \"happened to\" is not a method, and the general claim that this solves the beam-shaping problem needs qualification.\n\nSecond point, minor: the GS comparison runs only 60 iterations with no convergence check. That makes the quantitative superiority claim suggestive, not conclusive. The reported energy efficiency of 99.98% also looks too clean, though it is not central.\n\nCitations: the paper cites the author's own project report [2] for the standard ray-mapping equation. That is not ideal — a proper optics reference would be better — but it is not what drives the result, and the independent diffraction validation breaks any circularity concern.\n\nBottom line: honest work, real examples, a correctable but real technical gap. The missing boundary condition should be fixed before the method is claimed to solve general beam shaping, but for the demonstrated targets the claim holds. I would send this to peer review and ask for the boundary-condition fix and a stronger GS baseline.","headline":"Legitimate new PINN application with real diffraction validation, but the Monge–Ampère solver omits the boundary condition that makes the transport map onto, and the GS comparison is not convergence-checked.","tokens_in":6875,"tokens_out":1656,"would_cite":false,"duration_ms":20127,"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":"A physics-informed neural network solves the Monge-Ampère beam-shaping equation for flat optics, producing phase masks that outperform conventional phase retrieval on a structured far-field target.","keywords":["physics-informed neural networks","beam shaping","flat optics","phase masks","Monge-Ampère equation","ray mapping","scalar diffraction","metasurfaces"],"falsifier":"A direct test would be to design a target distribution concentrated near the edge of the aperture or with a sharp cutoff, and measure via wave propagation whether the PINN's phase produces the correct boundary coverage and total energy. A failure would show missing or extra energy at the edges, or an efficiency below the reported 99.98%.","tokens_in":6080,"feed_emoji":"🔦","tokens_out":3458,"duration_ms":41010,"temperature":0.7,"pith_summary":"The paper introduces a physics-informed neural network (PINN) that designs phase-only optical elements for beam shaping by solving the Monge-Ampère ray-mapping equation. The trained network outputs a smooth phase profile that, when validated by scalar diffraction, reproduces a prescribed target intensity with a relative RMS error of about 2.97% for a far-field logo pattern, compared with 22.6% for a standard phase-retrieval algorithm. This matters because beam shaping is used in laser machining, displays, and flat optics, and the PINN avoids the speckle and irregular phases typical of Fourier-based methods. The approach extends previous cylindrically symmetric metasurface designs to arbitrary two-dimensional targets without relying on data-driven training.","feed_headline":"Physics-trained neural net shapes laser beams 8x more accurately","feed_subtitle":"Solving the Monge-Ampère equation directly replaces speckle-prone phase retrieval for flat-optics targets.","key_machinery":"The key machinery is the Monge-Ampère equation derived from the generalized Snell law and power conservation. For far-field targets, the equation reduces to a standard Hessian-determinant form in which the phase gradient directly gives the direction cosines of the target; for finite distances, additional terms involving the ray propagation length and the longitudinal component K appear. The field is represented by a fully connected neural network with smooth tanh activations, and automatic differentiation provides all derivatives needed for the residual. The target intensity is sampled at the phase-dependent mapped coordinates using differentiable bilinear resampling, so the gradients flow t","core_discovery":"The central claim is that a PINN can directly solve the Monge-Ampère equation that governs energy-conserving ray mappings for a phase-only optical element, for both finite-distance and far-field targets. The network represents the phase potential and is trained by minimizing the residual of the PDE, using automatic differentiation to compute the required second derivatives. For a Gaussian beam to a 'DT' logo target in the far field, the learned phase produces an energy efficiency of about 99.98% inside the target support, versus 95.90% for a conventional projection-based algorithm, and suppresses the high-frequency speckle caused by irregular phase retrieval. The ray-based design is confirme","pith_inferences":["The absence of an explicit boundary condition enforcing that the source aperture maps onto the entire target domain is a likely weakness. Without such a condition, the interior residual may leave portions of the target unilluminated or violate global energy conservation, even if post-hoc admissibility checks catch gross errors.","A natural testable extension is to add a boundary loss or use a conservative formulation; if successful, this would make the method robust for targets whose support is strongly clipped by the aperture.","The same residual approach could be combined with a parameterized target family to build a generative model for beam-shaping optics, though such a model would require many training runs and may struggle with very sharp or discontinuous target features.","The method's reliance on smooth tanh activations and a scalar phase potential may limit its ability to represent wavefronts with vortices or other phase singularities, which are outside the Monge-Ampère transport-map description."],"forward_implications":["If the method holds, designing phase masks for complicated two-dimensional beam shapes becomes a PDE-solving task rather than an iterative Fourier projection, potentially yielding smoother and more accurate optic surfaces.","The approach can be extended to non-cylindrically symmetric targets, opening up uses in structured illumination, holography, and display backlighting.","The ray-based PINN solution remains valid under scalar wave propagation for the tested cases, suggesting that the method bridges geometric and wave descriptions of beam shaping.","Because the training signal is the physics residual itself, the method requires no labeled phase data and can be adapted to new source-target pairs by retraining the network for that specific design.","The reported efficiency and error metrics imply that phase-only elements can channel nearly all incident power into a desired pattern, which is valuable for high-power laser applications."],"fun_headline_variants":["Physics-informed net solves Monge-Ampère for beam shaping","Neural net beats Gerchberg-Saxton in flat-optics design","PINN achieves 99.98% efficiency shaping laser beams","Flat optics: AI designs phase profiles without speckle","Deep learning solves beam-shaping PDE for flat optics"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method assumes that satisfying the interior Monge-Ampère residual alone is enough to produce a globally energy-conserving ray map, without an explicit boundary condition linking source aperture to target boundary; if this fails, parts of the target may be missing or under-illuminated.","fun_headline_variants_meta":{"raw":{"variants":["Physics-informed net solves Monge-Ampère for beam shaping","Neural net beats Gerchberg-Saxton in flat-optics design","PINN achieves 99.98% efficiency shaping laser beams","Flat optics: AI designs phase profiles without speckle","Deep learning solves beam-shaping PDE for flat optics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000309,"raw_usage":{"total_tokens":1543,"prompt_tokens":625,"completion_tokens":918,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":369,"completion_tokens_details":{"reasoning_tokens":846}},"tokens_in":369,"tokens_out":918,"duration_ms":9638,"temperature":1.0,"reasoning_tokens":846,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:21:08.798296+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to design a target distribution concentrated near the edge of the aperture or with a sharp cutoff, and measure via wave propagation whether the PINN's phase produces the correct boundary coverage and total energy. A failure would show missing or extra energy at the edges, or an efficiency below the reported 99.98%.","supporting_citations":[],"review_version":1}