Pith. sign in

REVIEW 3 major objections 6 minor 12 references

Experimental Demonstration of an Optical Neural PDE Solver via On-Chip PINN Training

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A photonic micro-ring weight bank can be trained on-chip, without backpropagation, to solve the one-dimensional heat equation.

desk verdict 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. read the letter →

arxiv 2501.00742 v1 pith:AKVB6VM6 submitted 2025-01-01 cs.LG cs.ARphysics.optics

Yequan Zhao , Xian Xiao , Antoine Descos , Yuan Yuan , Xinling Yu , Geza Kurczveil , Marco Fiorentino , Zheng Zhang
show 1 more author
Raymond G. Beausoleil
This is my paper · ORCID
classification cs.LGcs.ARphysics.optics
keywords opticalneuralnetworksphysics-informedzeroth-orderoptimizationbackpropagation-freetrainingmicro-ringresonatorphotonicchipheatequationPDEsolving
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

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.

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 (3)
  1. [Section 3, Hardware Demo paragraph] 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.
  2. [Section 3, Simulation Results and Hardware Demo] 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.
  3. [Section 2, Principle and Architecture; Section 3, Hardware Demo] 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.
minor comments (6)
  1. [Section 2, first paragraph] There is a typo: 'repersent' should be 'represent'.
  2. [Section 3, Simulation Results] The sentence 'The results indicate the affect of bit accuracy' uses 'affect' where 'effect' is intended.
  3. [Figure 2(c)] The weight curves lack axis labels and units; the voltage axis and transmission/weight axis should be labeled for reproducibility.
  4. [Throughout] The notation 'ℓ2' is used inconsistently (also 'l2' in the abstract). Please use a consistent math formatting for the norm.
  5. [References and related work] 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].
  6. [Section 3, Experimental Demonstration] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the trained PINN is evaluated against the analytic heat-equation solution using a physics-based loss, and the self-cited ZO algorithm is re-specified in the paper itself.

full rationale

The paper's derivation chain is self-contained and does not reduce its predictions to its inputs. The PINN loss L = Lr + L0 + Lb is constructed from the PDE residual and initial/boundary conditions (Section 2), and the reported 5E-3 l2 error is measured against the analytic heat-equation solution on a hold-out test set, so the outcome is not encoded in the training objective. The back-propagation-free training algorithm is attributed to the authors' prior work [7], but Section 2 and Fig. 1(d) re-specify the zeroth-order gradient estimator (two forward calls with ± perturbations and finite-difference combination), so the self-citation is not load-bearing. The experimentally measured voltage-weight look-up table used in the numerical simulation is device calibration, not a fit to the PDE solution. One caveat is that the Hardware Demo paragraph reports only spectral measurements and weight readouts and does not explicitly state that optical forward passes were performed; if the forward pass was software-computed, the 'optical neural PDE solver' claim would be unsupported, but that is a completeness/correctness issue rather than a circular derivation, and the quoted text does not exhibit any equation or fitted parameter that makes the claimed result true by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim relies on unstated training hyperparameters, on the validity of finite-difference approximations inside the loss, and on the assumption that the measured MRR weights faithfully implement the network during training. No new physical entities are introduced.

free parameters (5)
  • learning_rate
    The paper does not report the learning rate or schedule used in the 1000 update steps; convergence and final error depend on it.
  • perturbation_step_mu
    The ZO gradient estimate uses a plus-minus mu perturbation; the value is not stated, and it trades bias against noise.
  • loss_weights
    The loss is defined as L = Lr + L0 + Lb, but the relative weighting of the PDE residual, initial condition, and boundary condition terms is not specified.
  • initial_weights
    Initial network parameters are not described; initialization can materially affect the convergence path on noisy analog hardware.
  • collocation_points
    The number and placement of the collocation points used to evaluate the loss are not given.
assumptions (4)
  • domain assumption A 2-hidden-layer network with width 4 can approximate the heat equation solution closely enough for 5e-3 error.
    No approximation-error analysis is given; the network is very small and success depends on the problem being smooth and simple.
  • standard math Central finite differences for du/dt and d2u/dx2 are accurate at the chosen collocation points.
    Central differences are standard but require an adequate grid spacing; the paper does not report the spacing or convergence checks.
  • domain assumption Measured MRR weights at the four resonances equal the effective weights during network evaluation.
    The hardware demo reads weights from spectra; this assumes the readout is faithful to the implemented multiplication and stable over time.
  • domain assumption Non-differentiable optical hardware can be treated as a black-box for zeroth-order optimization.
    ZO methods tolerate black-box functions, but physical noise and drift can bias finite-difference estimates; no analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Experimental Demonstration of an Optical Neural PDE Solver via On-Chip PINN Training." pith.science (2026). https://pith.science/paper/AKVB6VM6

@misc{pith2026250100742,
  author       = {Pith},
  title        = {Pith review of: Experimental Demonstration of an Optical Neural PDE Solver via On-Chip PINN Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AKVB6VM6}},
  note         = {Machine review of arXiv:2501.00742}
}
read the original abstract

Partial differential equation (PDE) is an important math tool in science and engineering. This paper experimentally demonstrates an optical neural PDE solver by leveraging the back-propagation-free on-photonic-chip training of physics-informed neural networks.

Figures

Figures reproduced from arXiv: 2501.00742 by the authors.

Figure 1
Figure 1. (a) PINN framework. (b) Comparison between first-order training and zeroth-order training. (c) First-order training [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) Schematic of the 1×4 MRR weight bank. (b) Microscope image of the weight bank and the schematic of the experimental setup for on-chip training. (c) Weight curves of the four MRRs. implement it on a photonics AI chip (a 1×4 micro-ring resonator (MRR) weight bank). The on-chip training experiment learns a solution for a one-dimensional heat equation with 5E-3 ℓ2 error after 1000 iterations of update, which experim… view at source ↗
Figure 3
Figure 3. Comparison between the ground truth solution and learned solutions of (a) BP-free training simulation and (b) [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 9 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter edition editor eid howpublished institution journal key month note number organization pages publisher school series title type volume year label INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence := #2 'after.sentence := #3 'after.bl...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in " " * FUNCTION format....

  3. [3]

    T., Asheghi, M., and Chandra, R

    Li, P., Pileggi, L. T., Asheghi, M., and Chandra, R. (2004, November). Efficient full-chip thermal modeling and analysis. In IEEE/ACM International Conference on Computer Aided Design, 2004. ICCAD-2004. (pp. 319-326). IEEE

    work page 2004

  4. [4]

    F., Polimeridis, A

    Villena, J. F., Polimeridis, A. G., Wald, L. L., Adalsteinsson, E., White, J. K., and Daniel, L. (2015). MARIE–a MATLAB-based open source software for the fast electromagnetic analysis of MRI systems. In Proceedings of the 23rd Annual Meeting of ISMRM, Toronto, Canada (p. 709)

    work page 2015

  5. [5]

    Bansal, S., and Tomlin, C. J. (2021, May). Deepreach: A deep learning approach to high-dimensional reachability. In 2021 IEEE International Conference on Robotics and Automation (ICRA) (pp. 1817-1824). IEEE

    work page 2021

  6. [6]

    Raissi, M., Perdikaris, P., and Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378, 686-707

    work page 2019

  7. [7]

    T., and Pan, D

    Gu, J., Zhao, Z., Feng, C., Li, W., Chen, R. T., and Pan, D. Z. (2020, July). FLOPS: Efficient on-chip learning for optical neural networks through stochastic zeroth-order optimization. Design Automation Conference (pp. 1-6)

    work page 2020

  8. [8]

    W., Park, T., Bartlett, B., Williamson, I

    Pai, S., Sun, Z., Hughes, T. W., Park, T., Bartlett, B., Williamson, I. A., ... and Miller, D. A. (2023). Experimentally realized in situ backpropagation for deep learning in photonic neural networks. Science, 380(6643), 398-404

    work page 2023

Show all 12 references
  1. [9]

    and Zhang, Z

    Zhao, Y., Xiao, X., Yu, X., Liu, Z., Chen, Z., Kurczveil, G., ... and Zhang, Z. (2023). Real-Time FJ/MAC PDE Solvers via Tensorized, Back-Propagation-Free Optical PINN Training. arXiv preprint arXiv:2401.00413

    2023 arXiv

  2. [10]

    Liu, Z., Yu, X., and Zhang, Z. (2022). TT-PINN: a tensor-compressed neural PDE solver for edge computing. arXiv preprint arXiv:2207.01751

    2022 arXiv

  3. [11]

    Zhao, Y., Yu, X., Chen, Z., Liu, z., Liu, Z., and Zhang, Z. (2023). Tensor-compressed back-propagation-free training for (physics-informed) neural networks. arXiv preprint arXiv:2308.09858

    2023 arXiv

  4. [12]

    B., Van Vaerenbergh, T., Liang, D., Beausoleil, R

    Xiao, X., On, M. B., Van Vaerenbergh, T., Liang, D., Beausoleil, R. G., and Yoo, S. J. (2021). Large-scale and energy-efficient tensorized optical neural networks on III–V-on-silicon MOSCAP platform. Apl Photonics, 6(12)

    2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.