Scalable multilayer diffractive neural network with all-optical nonlinear activation

Bohan Zhang; Futai Hu; Junye Zou; Kunpeng Chen; Ruiqi Liang; Sheng Liu; Wenhe Jia; Xiaokai Li; Yiying Dong; Yuanmu Yang

REVIEW 4 major objections 6 minor 1 references

Scalable multilayer diffractive neural network with all-optical nonlinear activation

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a folded, reconfigurable diffractive neural network using one spatial light modulator and a mirror-coated silicon wafer achieves all-optical nonlinear activation, with accuracy gains that grow with depth and task…

desk verdict The folded single-SLM DNN with Si activation is a smart integration, but the experimental evidence that the Kerr/TPA nonlinearity drives the accuracy gain is not yet quantitatively supported. read the letter →

arxiv 2504.13518 v1 pith:ACPIJD4Q submitted 2025-04-18 physics.optics

Yiying Dong , Bohan Zhang , Ruiqi Liang , Wenhe Jia , Kunpeng Chen , Junye Zou , Futai Hu , Sheng Liu

show 2 more authors

Xiaokai Li Yuanmu Yang

This is my paper · ORCID

classification physics.optics

keywords diffractiveneuralnetworkall-opticalnonlinearactivationKerrnonlinearityspatiallightmodulatorfoldedopticalsystemsiliconwaferQuickDraw

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

The paper sets out to show that a folded, reconfigurable diffractive neural network can perform multilayer all-optical computation with genuine nonlinear activation, using just one spatial light modulator and a mirror-coated silicon wafer. The silicon's third-order optical nonlinearity gives a near-instantaneous, intensity-dependent phase and reflectance response that is applied between phase-modulation layers, so no electro-optic conversion is needed for activation. Across handwritten-digit, fashion-product, and QuickDraw classification tasks, the authors report that adding this nonlinear activation raises accuracy, with the simulated gain on the 25-class QuickDraw benchmark growing to 26.33 percentage points at nine layers. If the claim holds, it addresses two known bottlenecks of all-optical neural networks - static weights and missing interlayer nonlinearity - while preserving the parallelism and speed of free-space optics.

What carries the argument

The load-bearing object is the folded optical path paired with a differentiable nonlinear activation model. A single phase-only SLM (1272×1024 pixels, with three 286×286-pixel modulation blocks) is combined with a 5-mm mirror-coated silicon wafer; each pass off the wafer applies the activation $f(U)$, so one SLM provides multiple reconfigurable phase masks and the wafer provides interlayer nonlinearity. The model uses literature values $n_2=4.5\times10^{-18}$ m²/W and the two-photon absorption coefficient $\beta$ for Si at 1550 nm, together with the measured 90.44% linear reflectance, to define $f(U)$, and the phase values on the SLM are trained by backpropagation through this forward model. The property that carries the argument is that $f(U)$ is instantaneous, intensity-dependent, and applied between every phase-modulation stage, so deeper networks cannot be reduced to a single linear layer.

What would settle it

Measure the nonlinear phase shift and reflectance change of the mirror-coated silicon wafer under the actual spatial-light-modulator-shaped field at the classification intensity of 6.33 GW/cm², without the beam splitter and without camera attenuation. If the local phase shift stays below the roughly $2\pi$ range used in training, or if classification accuracy no longer improves when the laser intensity is raised, the central claim is refuted.

Watch

Extended reading notes

Core claim

The central discovery is that the third-order ($\chi^{(3)}$) nonlinearity of a mirror-coated silicon wafer can serve as an all-optical activation function in a multilayer diffractive neural network, and that including it does what nonlinearity does in digital networks: it prevents hidden layers from collapsing into one linear transform and yields accuracy gains that widen with depth and task difficulty. In the forward model, propagation is $T=PM_NPfP\dots(M_1Pf(P(U)))$, with activation $f(U)=\sqrt{R_{\rm NL}}\,U e^{j\varphi_{\rm NL}}$, where the nonlinear phase is $\varphi_{\rm NL}=kn_2|U|^2L$ and the reflectance change is $R_{\rm NL}=R_0e^{-\beta|U|^2L}$. The folded geometry routes the beam through three phase-modulation blocks on one SLM and three reflections off the wafer, giving about $10^5$ programmable parameters. Experimentally, raising the peak intensity from 0.133 GW/cm² to 6.33 GW/cm² improved accuracy on all three tested tasks, and simulations show the nonlinear network beats its linear counterpart and a comparable linear electronic multilayer perceptron on 25-class QuickDraw. The authors note that the Gaussian-beam calibration did not reach a full $2\pi$ phase shift; they attribute the larger experimental effect to higher local intensity of the SLM-modulated light and to greater robustness of the nonlinear system to model errors.

Load-bearing premise

The load-bearing premise is that the mirror-coated silicon wafer's nonlinear response, measured with a smooth Gaussian beam, is also what the spatially structured, higher-intensity field from the spatial light modulator experiences during classification.

Editorial extensions

If this is right

Nonlinear activation prevents the hidden layers of the diffractive network from collapsing into a single linear transformation, so the network depth becomes functionally meaningful.
The folded geometry realizes a three-layer reconfigurable network with about $10^5$ programmable parameters using a single SLM, avoiding the component-count explosion of conventional multilayer reconfigurable diffractive networks.
On the 25-class QuickDraw benchmark, the simulated 9-layer nonlinear DNN reaches 64.53% accuracy, surpassing a linear DNN (38.20%) and a comparable linear electronic MLP (55.23%); a digital MLP with ReLU reaches 74.10%.
Testing loss follows a power law in parameter count, matching digital MLP scaling behavior, so increasing SLM pixel density should keep improving accuracy predictably.
A 3-layer nonlinear DNN with a beam-splitter residual connection reaches 77.20% accuracy on Fashion-MNIST, up from 75.46% without the residual block.

Reading between the lines

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

Extension: because activation strength scales with local intensity, the folded design could be pushed toward lower total laser power by concentrating light into higher local intensities on the wafer, subject to the SLM damage threshold; this design trade-off is not quantified in the paper.
Extension: if the reported robustness to misalignment comes from the nonlinear activation itself, sweeps of laser intensity in error-sensitivity simulations should show monotonically smaller accuracy drops as activation strengthens; this is testable without new hardware.
Extension: the single beam-splitter residual connection suggests a family of multi-branch optical networks in which multiple splitters create trainable shortcut paths, adding representational capacity without adding phase-modulation layers.
Extension: a spatially resolved measurement of the wafer's nonlinear phase and reflectance under the actual SLM-modulated field would settle whether the experimental gain is the instantaneous $\chi^{(3)}$ effect assumed in training or a slower cumulative effect; the paper does not report such a measurement.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The genuinely new bit is the folded geometry: a single SLM provides multiple phase-modulation layers in a round-trip scheme, while a mirror-coated Si wafer supplies an all-optical nonlinear activation via χ(3). That combination directly tackles two known DNN bottlenecks — reconfigurability and nonlinearity — and the implementation looks careful, with a standard forward model (Eqs. 1–4) and a training pipeline that includes measured nonlinear parameters. The code and data are available, and the depth-scaling simulations (Fig. 4) are clean and informative. Feeding the measured nonlinear response back into the model is calibration, not circularity.

The soft spot is quantitative. The paper never reports absolute nonlinear accuracies or error bars for the experimental confusion matrices; it says only that improvement was 'significant.' More tellingly, Supplementary Section 6 concedes that with the measured n2 and β, the model predicts a 'relatively small' gain — the (0.9,1) region in Fig. 2b, where TPA-induced reflectance change offsets most of the phase benefit. Yet the main text claims the experimental nonlinear DNN beat even the ideal nonlinear simulation, and attributes that to robustness against system errors. That explains error tolerance, not the size of the gain. The only bridge is the unmeasured claim that SLM-modulated light has higher local intensity, inducing a larger nonlinear phase shift. With β≈1.5 cm/GW and L≈1 cm, the transmittance at the highest-intensity spots is on the order of e^{-9.5}, so the very regions that would produce a 2π phase shift are largely absorbed. The measured-parameter model therefore predicts a small effect; the large experimental improvement is not quantitatively tied to the Kerr/TPA mechanism.

That is a real weakness, but not a reason to dismiss the paper. The folded reconfigurable DNN works, and the experiments do show improved accuracy at high power. The causal story is incomplete, not obviously wrong. A spatially resolved measurement of the nonlinear phase and reflectance in the actual DNN geometry, plus the raw accuracy numbers, would settle it. The scalability plots are numerical and idealized, so they support the model, not the hardware.

This is a paper for ONN experimentalists and anyone working on nonlinear activation. It deserves a serious referee — the idea is valuable and the gaps are addressable. I would not put the experimental numbers in my own work yet. The reading group would enjoy picking apart the evidence chain.

Referee Report

4 major / 6 minor

Summary. The paper proposes a folded, reconfigurable diffractive neural network (DNN) that uses a single spatial light modulator (SLM) for multiple phase-modulation layers and a mirror-coated silicon wafer as a χ(3) nonlinear activation medium. The forward model (Eqs. 1-4) combines Rayleigh-Sommerfeld propagation with a Kerr-type nonlinear phase shift and two-photon-absorption-induced reflectance change. The authors train the SLM phase masks via backpropagation, implement three-layer systems experimentally for MNIST subsets and Fashion-MNIST, and report accuracy improvements when the laser intensity is increased to activate the nonlinearity. They also present simulations showing larger gains for deeper networks and harder tasks (e.g., 26.33% improvement on 25-class QuickDraw with a 9-layer nonlinear DNN), compare against electronic MLPs, and demonstrate a residual-connection variant.

Significance. If the central claim is fully substantiated, the work would be a valuable step toward scalable all-optical neural networks with reconfigurable multilayer nonlinear processing, a known bottleneck for diffractive approaches. The folded geometry using a single SLM is an elegant way to multiply phase-modulation layers, and the inclusion of a nonlinear activation function is conceptually important. Strengths include a clearly specified forward model, a public code repository (GitHub), and a systematic simulation study of depth and task-complexity scaling. The experimental demonstration, however, is the key load-bearing part, and it currently lacks the quantitative and mechanistic support needed to validate the nonlinear-activation claim.

major comments (4)

[Experimental results (Fig. 3) and Supplementary Section 6] The experimental attribution of the accuracy improvement to the χ(3) nonlinearity of the mirror-coated Si wafer is not supported by the paper's own calibration data. Supplementary Section 6 states that when the measured n2 and β are inserted into Eqs. (2)-(4), the predicted accuracy improvement is only "relatively small" (the point (0.9,1) in Fig. 2b). The main text's only bridge is the unmeasured assertion that SLM-modulated light has higher local intensity and thus a larger nonlinear phase shift. With the measured β ≈ 1.5 cm/GW and effective interaction length L ≈ 1 cm, the intensity reflectance at the stated peak intensity of 6.33 GW/cm² is R0 exp(-βIL) ≈ 0.9×exp(-9.5) ≈ 6.8×10⁻⁵, meaning the very hot spots that would produce 2π phase shifts are almost completely absorbed. The robustness argument (Supplementary Section 7, Fig. S5) explains error tolerance, not the magnitude of the activation benefit. Please provide spatially resolved measurements of the nonlinear phase and reflectance change under the actual SLM-modulated illumination, or otherwise quantify the local intensity distribution, and report the expected improvement from the measured-parameter model side by side with the experimental results.
[Experimental results (Fig. 3c-h)] The paper does not report the numerical classification accuracies for the nonlinear DNN. The text only says that accuracy "improved significantly" and shows confusion matrices, without giving the accuracy percentages for the nonlinear cases. Given that the experimental test set is only 100 samples per category (Methods), the binomial standard error is roughly 4-5 percentage points, so a quantitative statement with confidence intervals or repeated trials is essential to establish that the improvement is statistically significant and indeed "surpassed simulations." Please include the accuracy values for all three tasks and the linear/nonlinear comparison in the text or figure.
[Introduction and Experimental results (calibration paragraph)] The Introduction states that the mirror-coated Si wafer "provides a full 2π phase shift," but the paper's own characterization (right panel of Fig. 3a; Supplementary Section 5) shows that the measured maximum nonlinear phase shift with a Gaussian beam does not reach 2π. The Results section later acknowledges this and invokes an unmeasured local-intensity enhancement. Please revise the abstract and Introduction to state that the 2π phase shift is an idealized design target, and clearly distinguish the measured nonlinear response from the extrapolated one.
[Nonlinearity-assisted scalable DNN (Fig. 4)] The scalability results (Fig. 4a, including the 26.33% improvement for 25-class QuickDraw) are obtained from the idealized model with 2π phase shift and negligible TPA, not from the measured-parameter model. The paper should explicitly state that these are idealized simulations and that the experimentally validated regime (Supplementary Section 6) shows only a small improvement. Otherwise, the reader may conflate the experimental demonstration with the idealized scalability projection.

minor comments (6)

[Eq. (4)] The expression f(U) = √RNL×UU*e^{j(φ+φNL)} appears dimensionally inconsistent; it should be f(U) = √RNL U e^{j(φ+φNL)} (or similar). Please clarify the notation.
[Supplementary Section 6] The sentence "although the nonlinear phase shift approaches 2π" contradicts the main text's statement that the measured phase shift does not reach 2π. Please specify which case (idealized vs measured) is being referred to.
[Table S1] The Si thickness is listed as L=1 cm, while the main text describes a 5-mm-thick wafer. Presumably the effective double-pass length is 1 cm; please state this explicitly in the table footnote.
[Experimental setup (Methods)] The text says "Using 100 test samples per category" but the dataset description in Methods gives much larger test sets; please clarify that these are the experimentally tested subsets, not the full test sets.
[Fig. 4e-f and Supplementary Section 8] The residual-connection results are simulations in which the nonlinear refractive index n2 is trained as a free parameter; this is not physically tunable in the experiment. Please state clearly that these are simulation results and that the trainable n2 is a numerical convenience.
[Fig. 4d] The power-law fit uses three free parameters (α, β, L0) to fit a small number of data points; please report the fit uncertainty and the number of points to support the claimed scaling trend.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the nonlinear activation claim rests on an independently parameterized forward model and external benchmark experiments, not on a self-referential fit.

full rationale

The paper's derivation chain is self-contained rather than circular. The forward model in Eqs. (1)-(4) combines standard Rayleigh-Sommerfeld propagation with a Kerr and two-photon-absorption activation model, using n2 and beta values taken from independent literature (Refs. 51, 52) and from the paper's own interferometric and I-scan characterizations; these parameters are not fitted to the classification accuracies that the paper reports. Training the SLM phase masks via backpropagation inside this forward model, including re-training with the measured Si response (Supplementary Section 6), is calibration of a design, not a prediction extracted from the data being predicted. The central experimental evidence is an external comparison: the same classification tasks are run at low power (linear regime) and high power (nonlinear regime) on held-out test samples, and the accuracy difference is observed rather than imposed. The measured-parameter simulation in Supplementary Section 6 predicts only a relatively small improvement, so the paper's experimental claim of a larger improvement is not manufactured by the model; it is instead an unresolved quantitative discrepancy, which is a support/correctness concern, not circularity. The Fig. 4d power-law fit is a descriptive fit to the authors' own simulation data and is not used as evidence for the main accuracy improvement claim. The self-citations present (Refs. 48 and 49, in the materials survey) are not load-bearing: they merely point to examples of epsilon-near-zero and 2D nonlinear materials, and the choice of Si does not depend on those citations. No step in the paper reduces by construction to its own inputs.

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

The central claim rests on a small number of measured/literature optical parameters (n2, β, R0) and on modeling assumptions about diffraction and the SLM phase response. The most fragile is the assumption that the measured Gaussian-beam nonlinear response extrapolates to the SLM-modulated field. No invented physical entities are introduced.

free parameters (4)

n2 (nonlinear refractive index of Si) = 4.5e-18 m^2/W (literature, 1550 nm)
Used in Eq. (2) to compute nonlinear phase shift; value taken from refs. 51,52, not independently measured in this paper.
β (two-photon absorption coefficient of Si) = ~1.5 cm/GW (literature/measured)
Used in Eq. (3) for nonlinear reflectance change; value from literature or calibration in Supplementary Table S1.
R0 (linear reflectance of mirror-coated Si wafer) = 90.44%
Measured linear reflectance used in the activation model; load-bearing for the output intensity.
α, β, L0 (power-law scaling fit) = not reported
Fitted to the model's testing loss vs. parameter count in Fig. 4d to claim scaling behavior; this is a fit to the system's own simulations.

assumptions (5)

standard math Rayleigh-Sommerfeld diffraction integral accurately models free-space propagation in the folded system.
Used in Eq. (1) and Supplementary Eqs. (S1)-(S7).
domain assumption The SLM phase response is independent of incident angle for the small 3.55° angle.
Stated in Results: 'MN is approximated to be independent of the incident angle since the phase response of the SLM is relatively insensitive to small incident angles.'
domain assumption The nonlinear response of the mirror-coated Si wafer is described by instantaneous χ(3) Kerr phase shift and TPA with the parameters used in Eqs. (2)-(4).
Assumed throughout; used for both training and experimental interpretation.
ad hoc to paper The measured nonlinear response (with a Gaussian beam) extrapolates to the SLM-modulated field, with higher local intensity providing a larger nonlinear phase shift.
The paper states the measured phase shift 'does not reach the full 2π range' with the Gaussian beam, but claims the SLM-modulated light 'can have a higher local intensity, resulting in a larger nonlinear phase shift' (Results, Experimental results). This unverified assumption reconciles the device with the simulated ideal activation.
domain assumption The fully connected condition (Eqs. S9-S11) ensures that each SLM pixel can contribute to each output pixel.
Used to design system geometry; the paper relies on this to justify the 3.55° incident angle and 2.86 cm distance.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Scalable multilayer diffractive neural network with all-optical nonlinear activation." pith.science (2026). https://pith.science/paper/ACPIJD4Q

@misc{pith2026250413518,
  author       = {Pith},
  title        = {Pith review of: Scalable multilayer diffractive neural network with all-optical nonlinear activation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ACPIJD4Q}},
  note         = {Machine review of arXiv:2504.13518}
}

read the original abstract

All-optical diffractive neural networks (DNNs) offer a promising alternative to electronics-based neural network processing due to their low latency, high throughput, and inherent spatial parallelism. However, the lack of reconfigurability and nonlinearity limits existing all-optical DNNs to handling only simple tasks. In this study, we present a folded optical system that enables a multilayer reconfigurable DNN using a single spatial light modulator. This platform not only enables dynamic weight reconfiguration for diverse classification challenges but crucially integrates a mirror-coated silicon substrate exhibiting instantaneous \c{hi}(3) nonlinearity. The incorporation of all-optical nonlinear activation yields substantial accuracy improvements across benchmark tasks, with performance gains becoming increasingly significant as both network depth and task complexity escalate. Our system represents a critical advancement toward realizing scalable all-optical neural networks with complex architectures, potentially achieving computational capabilities that rival their electronic counterparts while maintaining photonic advantages.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

1 extracted references

[1]

1 Erb, R. J. Introduction to backpropagation neural network computation. Pharmaceutical Research 10, 165-170 (1993). https://doi.org/10.1023/A:1018966222807 2 Chen, H. et al. Diffractive deep neural networks at visible wavelengths. Engineering 7, 1483- 1491 (2021). https://doi.org/10.1016/j.eng.2020.07.032 3 Bristow, A. D. et al. Two-photon absorption and...

doi:10.1023/a:1018966222807 1993

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

1 Erb, R. J. Introduction to backpropagation neural network computation. Pharmaceutical Research 10, 165-170 (1993). https://doi.org/10.1023/A:1018966222807 2 Chen, H. et al. Diffractive deep neural networks at visible wavelengths. Engineering 7, 1483- 1491 (2021). https://doi.org/10.1016/j.eng.2020.07.032 3 Bristow, A. D. et al. Two-photon absorption and...

doi:10.1023/a:1018966222807 1993