{"id":"aa3fcbc5-57ff-4d03-a961-f2242d433875","arxiv_id":"2501.00742","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"This paper reports a hardware demo in which a 1x4 microring weight bank is trained with zeroth-order optimization to solve a 1D heat equation to 5e-3 error.","lead":"Researchers trained a tiny neural network to solve the 1D heat equation by tuning the weights of a photonic micro-ring chip with a backpropagation-free method. They report a relative error of 0.005 after 1000 updates, suggesting on-chip optical training as a possible route to fast edge PDE solvers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The hardware demo only reports reading MRR weights from spectra; it never states that optical forward passes were performed, so the claimed experimental demonstration of an optical neural PDE solver is unsupported.","rationale":"The reader identified the same load-bearing gap: the hardware demo does not establish that the neural network forward pass was performed optically. This is the decisive issue because the paper's headline claim is an 'experimental demonstration of an optical neural PDE solver.' Without direct evidence of optical forward passes, the experiment reduces to using the photonic chip as a controllable array of resistors whose weights are read out and then used in a purely digital PINN training loop. The abstract's language ('experimentally demonstrates an optical neural PDE solver by... on-chip training') is therefore unsupported by the described methods. The paper does contain plausible simulation results and a plausible architecture for how such a demo could be done, but the hardware evidence as written is insufficient. No other concern is as foundational: even if the ZO algorithm and the weight-tuning loop work as claimed, they only constitute on-chip parameter updates, not optical inference. A direct test—checking for photodetector recordings of forward passes—would settle the question. Thus the reader's REJECT verdict stands unchanged.","tokens_in":5987,"tokens_out":3105,"duration_ms":31538,"concrete_test":"Inspect the experimental logs or supplementary methods for any photodetector voltage recordings that correspond to forward passes with encoded inputs (e.g., readings for each collocation point and each perturbed input used in the ZO gradient estimate). If the only optical recordings are transmission spectra used to extract weights, then the forward pass was computed in software and the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the MRR weight bank was used as a photonic matrix-vector multiplier during training: input vectors encoded at the four wavelengths, photodetector outputs taken as network predictions, and these predictions used to compute the PINN loss and ZO gradient estimates. The Hardware Demo paragraph (Section 3) describes only a tunable laser and photodetector measuring the spectrum of the weight bank at different tuning voltages, with 'weight values read out at the four resonances.' Nothing in that paragraph says inputs were encoded at λ1–λ4 or that the photodetector output was used as the network output. The earlier architecture description (Section 3, first paragraph) explains how the device could work optically, but that is a claim about design, not a report of what was actually measured. If the forward pass was computed in software from the read-out weights, then the experiment demonstrates only that MRR weights can be tuned by a ZO training loop; it does not demonstrate an optical neural PDE solver. The learned solution and ℓ2 error curves in Fig. 3(b–d) could then have been produced entirely by a software simulation using measured weight values, and the phrase 'on-chip training' would be misleading.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a back-propagation-free, zeroth-order (ZO) training scheme for physics-informed neural networks (PINNs) on a photonic microring-resonator (MRR) weight bank, and claims an experimental demonstration of an optical neural PDE solver for the 1D heat equation with a reported 5E-3 l2 error after 1000 iterations. The ZO method estimates gradients from two forward passes with perturbed parameters, avoiding the need for a differentiable photonic model. The simulation section uses an experimentally measured voltage-weight lookup table and reports bit-accuracy effects. The hardware section describes tuning MRR voltages and reading the weight values from measured spectra. The central claim of an experimental optical neural PDE solver rests on whether the neural-network forward passes were actually performed optically on the chip, a point the manuscript does not clearly establish.","tokens_in":6181,"tokens_out":4863,"duration_ms":47307,"significance":"If substantiated, a calibration-free, back-propagation-free optical PINN training demonstration would be a notable advance toward real-time, edge-deployable PDE solvers and would strengthen the case for in-situ training of photonic neural networks. The ZO training idea is sound and the simulation with a measured lookup table is a plausible proof-of-concept. The physics-based loss is not circular because the reported error is measured against the analytic solution on hold-out points. However, the experimental evidence as described is insufficient to support the headline claim of an optical neural PDE solver, which significantly lowers the paper's contribution as written.","major_comments":[{"comment":"The described experimental procedure only reports measuring the spectrum of the MRR weight bank with a tunable laser and photodetector and reading out weight values at the four resonances under different tuning voltages. It never states that input vectors were encoded at the four wavelengths, that the photodetector output was used as the network prediction, or that the ZO gradient estimates and the PINN loss were computed from optical forward passes. Without an explicit statement that the forward propagation was executed optically, the experiment demonstrates only voltage-controlled tuning of MRR weights, not an optical neural PDE solver. The earlier architecture description in Section 3's first paragraph explains how the device could be used optically, but that is a design description, not a report of the performed measurement. If the forward pass was instead computed in software from the read-out weights, the central claim of an optical neural PDE solver and of 'on-chip training' is not supported.","section":"Section 3, Hardware Demo paragraph"},{"comment":"The quantitative claim of '5E-3 l2 error after 1000 iterations' is presented without error bars, the number of repeated experimental runs, or any statistical variability. Zeroth-order optimization is stochastic, and hardware noise is expected to produce run-to-run variation; a single learning curve cannot establish the claimed accuracy. Additionally, the reported comparison in Fig. 3(d) between the measured hardware curve and the simulated bit-accuracy curves is made without confidence intervals, so it is not possible to assess whether the observed performance is consistent with the simulation. The paper should report repeated trials, hyperparameters (learning rate, perturbation magnitude μ, loss weights), and a clear definition of the test set and collocation points used in the hardware experiment.","section":"Section 3, Simulation Results and Hardware Demo"},{"comment":"The claim that the approach is 'calibration-free' is contradicted by the experimental procedure as described. The hardware demo reads out the weight values at the MRR resonances from measured spectra, which provides a voltage-to-weight mapping; this is a calibration step, even if used only for monitoring. If the forward pass is performed in software using these measured weights, then the training depends critically on this calibration, and the 'calibration-free' claim is misleading. The authors should clarify what exactly is calibration-free in the demonstrated experiment and how the training updates are computed from optical measurements alone.","section":"Section 2, Principle and Architecture; Section 3, Hardware Demo"}],"minor_comments":[{"comment":"There is a typo: 'repersent' should be 'represent'.","section":"Section 2, first paragraph"},{"comment":"The sentence 'The results indicate the affect of bit accuracy' uses 'affect' where 'effect' is intended.","section":"Section 3, Simulation Results"},{"comment":"The weight curves lack axis labels and units; the voltage axis and transmission/weight axis should be labeled for reproducibility.","section":"Figure 2(c)"},{"comment":"The notation 'ℓ2' is used inconsistently (also 'l2' in the abstract). Please use a consistent math formatting for the norm.","section":"Throughout"},{"comment":"Reference [7] is the authors' own prior work on back-propagation-free optical PINN training. The present paper should explicitly state the incremental contribution over [7] (e.g., is this the first experimental implementation of that algorithm?) and what new knowledge is provided by the hardware demonstration relative to the simulation in [7].","section":"References and related work"},{"comment":"Only the ℓ2 error curves are shown, not the physics-informed loss (composed of PDE residual, initial condition, and boundary condition terms). Showing the loss convergence would strengthen the claim that the trained network satisfies the PDE, not just matches a held-out solution set.","section":"Section 3, Experimental Demonstration"}],"recommendation":"reject","confidential_remarks":"The central issue is that the described hardware experiment does not support the claimed optical forward pass; the setup as described (tunable laser and photodetector measuring spectra) is consistent only with reading out weights, not with performing matrix-vector multiplications. If the authors actually performed optical forward passes, they have omitted the essential details, but as written the manuscript's central claim is unsupported. This is a load-bearing flaw that cannot be remedied without either new experimental evidence or a substantial re-scoping of the claim to 'electronic training of photonic weights with simulated optical inference.' I also note that the paper appears to be a short conference-style manuscript; a longer version with complete experimental details could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The key thing to know: this is a real experimental step toward chip-in-the-loop training of a photonic weight bank, but the headline claim that it demonstrates an optical neural PDE solver is not backed by the evidence in the hardware section.\n\nWhat is actually new: the authors take a 1x4 MRR weight bank and use a zeroth-order (ZO) optimization loop to tune its voltages against a physics-informed loss. That is a sensible extension of their earlier simulation work (Ref. 7), and the paper gives a clear description of why ZO is attractive for non-differentiable photonic hardware. The simulation part is plausible, and the authors honestly include the 8-bit precision limitation. The comparison of experimental error curves to simulations in Fig. 3(d) is a reasonable way to bracket the effective bit accuracy, though it is reverse inference, not direct measurement.\n\nNow the soft spot, and it is load-bearing. The hardware demo says: \"A tunable laser and photodetector are used to measure the spectrum of the MRR weight bank at different tuning voltages... The weight values are read out at the four resonances of the MRRs.\" It never says that input data were encoded at the four wavelengths and that the photodetector output was used as the network prediction. Without that, the experiment demonstrates only that MRR weights can be tuned by a ZO training loop in software reading measured spectra. The forward pass, the thing that makes it an optical solver, is not shown to be optical. The stress-test note is right: the learned solution and error curves in Fig. 3(b-d) could have been produced by a software simulation using measured weight values.\n\nMinor issues: no error bars on the hardware curves, no code or data release, and the weight curves in Fig. 2(c) show variances but no quantitative calibration details. The PDE is a single linear heat equation, which is fine for a demo but makes the \"solver\" claim thin.\n\nMy take: the paper deserves a serious referee because the gap is fixable in principle—add a forward-pass measurement or soften the claim to \"chip-in-the-loop weight tuning\"—but as it stands the central experimental claim is unsupported. I would send it to review with a request for major revision, not desk-reject it, and I would not cite it as an optical neural PDE solver until the forward pass is actually demonstrated.\n\nBring it to reading group? Maybe—it is a useful case study in how experimental claims can outrun the reported measurement.","headline":"A plausible chip-in-the-loop demo of zeroth-order PINN training, but the paper's central claim of an optical forward pass is not supported by the hardware section as written.","tokens_in":6770,"tokens_out":1684,"would_cite":false,"duration_ms":17690,"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 photonic micro-ring weight bank can be trained on-chip, without backpropagation, to solve the one-dimensional heat equation.","keywords":["optical neural networks","physics-informed neural networks","zeroth-order optimization","backpropagation-free training","micro-ring resonator","photonic chip","heat equation","PDE solving"],"falsifier":"Inspect the Hardware Demo in Section 3: if the four inputs $\\lambda_1,\\dots,\\lambda_4$ were never modulated with the network's input data and the photodetector signal was never used as the network output during training, the forward pass was not optical. A decisive test is to compare, for a fixed set of trained weights, the photodetector measurement against the software-computed prediction on the same input; any mismatch would identify where the optical loop actually stops.","tokens_in":5768,"feed_emoji":"💡","tokens_out":9939,"duration_ms":82581,"temperature":0.7,"pith_summary":"Partial differential equations are usually solved by discretization, which is expensive; physics-informed neural networks (PINNs) are a discretization-free alternative, but training them normally takes hours on a GPU and needs backpropagation through a differentiable model. This paper tries to show that both obstacles can be removed by training a PINN directly on a photonic chip using zeroth-order (ZO) optimization, which estimates gradients from two forward evaluations and therefore needs only a black-box optical network. The experimental demonstration trains a small PINN implemented with a $1\\times4$ micro-ring resonator weight bank to solve a one-dimensional heat equation, reaching a $5\\times10^{-3}$ $\\ell^2$ error after 1000 parameter updates. If the claim holds, real-time, low-power PDE solving on edge devices becomes plausible without pre-calibration or backward-pass hardware.","feed_headline":"Optical chip learns heat equation without backpropagation","feed_subtitle":"A micro-ring weight bank solves a 1D heat equation to 0.5% error in 1,000 on-chip updates","key_machinery":"The machinery is zeroth-order (ZO) optimization applied to a photonic weight bank. Instead of computing $\\partial\\mathcal{L}/\\partial\\theta$ by backpropagation, the algorithm estimates the gradient by finite differences of the physics-informed loss: $\\partial\\mathcal{L}/\\partial\\theta \\approx (\\mathcal{L}(\\theta+\\mu\\delta)-\\mathcal{L}(\\theta-\\mu\\delta))/(2\\mu)$, which requires only two forward evaluations of the optical network per direction. The physical device is a $1\\times4$ micro-ring resonator (MRR) weight bank that performs wavelength-division-multiplexed dot products: inputs encoded at four wavelengths $\\lambda_1,\\dots,\\lambda_4$ multiply the four ring weights and sum at a photodetector. Training updates the thermal tuning voltages of the rings directly, using an experimentally measured voltage-to-weight lookup table in simulation; this is what makes the chip calibration-free and backprop-free.","core_discovery":"The central discovery claimed is that an optical neural network can serve as a physics-informed PDE solver whose weights are trained in situ, on the chip, using only forward passes. The paper's training loop replaces backpropagation with a zeroth-order gradient estimate: for each parameter update it evaluates the physics-informed loss at $\\theta+\\mu\\delta$ and $\\theta-\\mu\\delta$, and uses the difference to move $\\theta$. Because the photonic network is treated as a black box, its nonlinear voltage-to-weight response, fabrication errors, and environmental noise can be compensated during training rather than modeled in advance. Concretely, the authors report that on-chip training of the heat equation $u_t = \\frac{1}{\\pi^2}u_{xx}$ with initial condition $u(x,0)=\\sin(\\pi x)$ and zero boundary conditions reaches a $5\\times10^{-3}$ $\\ell^2$ error on a hold-out test set after 1000 iterations; they interpret the monotonically decreasing error as evidence that the optical chip learns the underlying physics.","pith_inferences":["An ambiguity left implicit is whether the hardware demonstration actually performed the forward pass optically end-to-end. The text says the weight values were read out at the four MRR resonances, but it never states that inputs were encoded at the four wavelengths and that the photodetector output was used as the network prediction; if the forward pass was computed in software from measured weigh","A direct way to settle this is to re-run the same 1000-iteration training with optical inputs and read the photodetector output as $\\hat{u}(x,t)$, then compare the resulting $\\ell^2$ error with the reported $5\\times10^{-3}$; a large discrepancy would show that the current result is not fully optical.","The zeroth-order method trades convergence speed for hardware simplicity: it requires two optical forward passes per parameter per iteration, so its practical advantage on larger PINNs will depend on how cheap and fast an optical forward pass is relative to digital backpropagation.","Since analog noise caps the effective precision at roughly 8-10 bits, a useful extension would be to measure how the learned PDE error scales with bit precision and with the number of ring resonators, to see whether larger optical PINNs can stay below a target error."],"forward_implications":["Photonic PINN training no longer requires a differentiable model of the hardware, eliminating exhaustive calibration of fabrication errors and device nonlinearities.","Because only forward inference calls are needed, the same training loop can run on an optical inference accelerator without adding backward-pass hardware or extra memory for computation graphs.","The reported accuracy ($\\ell^2 \\approx 5\\times10^{-3}$, behaving like between 8- and 10-bit arithmetic) indicates that analog bit precision, not the training algorithm, is the main limit on solution quality.","The paper's proposed route to larger problems is tensor-train decomposed PINNs (TT-PINN) executed on a tensorized optical neural network (TONN), which would extend the demonstration from small weight matrices to real-size networks."],"supporting_citations":[{"why":"Defines physics-informed neural networks and the residual, initial-condition, and boundary-condition loss that the paper trains.","marker":"[4]"},{"why":"Earlier zeroth-order on-chip learning method (FLOPS) that motivates the backpropagation-free training approach.","marker":"[5]"},{"why":"Experimentally realized in situ backpropagation for photonic neural networks; the hardware-overhead comparison the paper argues against.","marker":"[6]"},{"why":"The back-propagation-free, tensorized optical PINN training algorithm that this paper implements and experimentally demonstrates.","marker":"[7]"}],"fun_headline_variants":["Optical chip learns heat equation to 0.5% error","On-chip training solves PDE in 1000 updates","Photonic neural net solves heat equation without backprop","In-situ PINN training on optical chip no backprop"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole demonstration stands on the assumption that the trained network's forward pass was actually executed by light through the micro-ring weight bank; if the experiment measured only the ring weights and ran the network in software, the central optical-solver claim loses its experimental support.","fun_headline_variants_meta":{"raw":{"variants":["Optical chip learns heat equation to 0.5% error","On-chip training solves PDE in 1000 updates","Photonic neural net solves heat equation without backprop","In-situ PINN training on optical chip no backprop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000586,"raw_usage":{"total_tokens":2669,"prompt_tokens":777,"completion_tokens":1892,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":393,"completion_tokens_details":{"reasoning_tokens":1825}},"tokens_in":393,"tokens_out":1892,"duration_ms":15117,"temperature":1.0,"reasoning_tokens":1825,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:43:34.219212+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the Hardware Demo in Section 3: if the four inputs $\\lambda_1,\\dots,\\lambda_4$ were never modulated with the network's input data and the photodetector signal was never used as the network output during training, the forward pass was not optical. A decisive test is to compare, for a fixed set of trained weights, the photodetector measurement against the software-computed prediction on the same input; any mismatch would identify where the optical loop actually stops.","supporting_citations":[{"cited_title":"F., Polimeridis, A","cited_arxiv_id":null,"evidence_quote":"Defines physics-informed neural networks and the residual, initial-condition, and boundary-condition loss that the paper trains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier zeroth-order on-chip learning method (FLOPS) that motivates the backpropagation-free training approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Experimentally realized in situ backpropagation for photonic neural networks; the hardware-overhead comparison the paper argues against."},{"cited_title":"T., and Pan, D","cited_arxiv_id":null,"evidence_quote":"The back-propagation-free, tensorized optical PINN training algorithm that this paper implements and experimentally demonstrates."}],"review_version":1}