Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Ultrafast neuromorphic computing with nanophotonic optical parametric oscillators

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A nanophotonic optical parametric oscillator acts as an all-optical recurrent neural network that predicts chaotic time series, equalizes nonlinear channels, and classifies noisy waveforms at ~10 GHz with accuracies above 93%.

desk verdict A real 10 GHz OPO reservoir on TFLN with credible hardware results, but the chaotic-forecast benchmark lacks baselines and the all-optical framing is softer than advertised. read the letter →

arxiv 2501.16604 v1 pith:NILXE5M5 submitted 2025-01-28 physics.optics

Midya Parto , Gordon H.Y. Li , Ryoto Sekine , Robert M. Gray , Luis L. Ledezma , James Williams , Arkadev Roy , Alireza Marandi This is my paper · ORCID
classification physics.optics PACS 42.65.Yj42.79.Ta
keywords opticalparametricoscillatorthin-filmlithiumniobatephotonicneuralnetworkrecurrentchaotictimeseriespredictionchannelequalizationall-opticalcomputingultrafastphotonics
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

This paper reports a nanophotonic optical parametric oscillator, fabricated on thin-film lithium niobate, that performs machine learning directly on time-domain optical signals. Because the oscillator's cavity feeds its own output back into the parametric gain process, the device acts as a recurrent neural network whose input, memory, and nonlinear activation all remain in the optical domain. The authors demonstrate the device on three tasks: one-step-ahead forecasting of Lorenz and Mackey-Glass chaotic series, correction of nonlinear distortions in a simulated four-level communication channel, and classification of noisy sinusoid, square, and sawtooth waveforms. They report normalized mean-square errors of 0.07 and 0.06 for the two forecasting tasks and accuracies above 93% for the others, at a clock rate near 10 GHz with sub-nanosecond latency. The claim that matters is that a single integrated photonic circuit can carry out both the linear operations and the nonlinear activations of a recurrent network without converting signals to electronics and back.

What carries the argument

The central object is the degenerate optical parametric oscillator (OPO) on a thin-film lithium niobate chip: an optical cavity with a periodically poled section that gives parametric gain at half the pump frequency, here a 2090 nm signal from a 1045 nm pump. Synchronous pumping with ~2 ps pulses at a ~10 GHz repetition rate makes the cavity feedback act as an optical memory, so the signal field at each roundtrip is a nonlinear function of the current data-modulated pump and the previous signal field. This combination of memory and instantaneous parametric nonlinearity is what turns a single OPO into a recurrent neural network; a random input mask and trained output weights complete the network. The same device supplies both the linear multiply-accumulate operations and the nonlinear activation functions in the optical domain.

What would settle it

Compute the normalized mean-square error of a persistence forecast (predict the next sample equals the current one) and of a linear autoregressive model on the same Lorenz and Mackey-Glass data sampled at 10 GHz; if either matches the reported 0.07 and 0.06, the experiments do not show that the OPO's nonlinear processing is responsible for the forecasting accuracy.

Watch

Extended reading notes

Core claim

The central discovery is that a degenerate optical parametric oscillator, synchronously pumped by data-modulated optical pulses, implements a recurrent neural network in hardware. Each cavity roundtrip, the incoming pump pulse and the circulating signal pulse overlap in the periodically poled lithium niobate section; parametric gain makes the generated signal a nonlinear function of both, while the cavity preserves information from previous roundtrips. Weighted sums of the detected signal intensities form the output layer. With this single device, the authors forecast one step ahead of the Lorenz and Mackey-Glass chaotic systems, reduce the symbol error rate of a simulated nonlinear PAM4 channel from 19% to 7%, and classify noisy waveforms with 100% accuracy in their test set. The result is offered as evidence that integrated nanophotonic circuits can provide both linear and nonlinear neural-network operations in the optical domain, removing the need for optical-electrical-optical conversion.

Load-bearing premise

The load-bearing premise is that the benchmarks are nontrivial at the demonstrated clock rate; the paper evaluates one-step-ahead predictions on signals sampled at 10 GHz without comparing to a persistence or linear autoregressive baseline, so the low errors could partly come from consecutive samples being nearly identical rather than from the OPO's nonlinear computation.

Editorial extensions

If this is right

  • Recurrent-neural-network inference on time-domain signals can run entirely in the optical domain at a ~10 GHz clock rate, with a latency comparable to one cycle of a state-of-the-art electronic processor.
  • Because the clock rate is set by the electronic pump source and not by the parametric process itself, the same chip structure could operate faster when driven by a higher-repetition-rate pump.
  • The channel-equalization result points toward all-optical equalizers for high-speed data-center links, where the OPO network would correct nonlinear distortion without optical-electrical-optical conversion.
  • The simple readout training, least-squares or winner-takes-all, means the hardware can be retrained for new tasks without backpropagating through the optical system.

Reading between the lines

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

  • The paper does not compare its one-step-ahead forecasts against a persistence predictor or a linear autoregressive model on the same 10 GHz-sampled data; if those baselines reach similar errors, part of the reported accuracy may reflect smoothness of consecutive samples rather than the OPO's nonlinear computation.
  • Because the device is a fixed nonlinear recurrent kernel with a trained linear readout, a fair test of the photonic nonlinearity would be to compare its forecasting error against standard nonlinear autoregressive baselines at the same sampling rate.
  • A direct extension would be multi-step-ahead forecasting: if the network has internalized the chaotic dynamics rather than local autocorrelation, its error should grow slowly with prediction horizon.
  • The present output is intensity-only at the subharmonic; reading out phase or operating in different parametric regimes could give each node more computational states per clock cycle.
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

4 major / 5 minor

Summary. The manuscript reports an integrated photonic recurrent neural network based on a degenerate optical parametric oscillator on thin-film lithium niobate. Input sequences are encoded onto ~10 GHz pump pulses via an electro-optic modulator, and the cavity feedback and parametric gain create recurrent coupling among signal pulses. The authors demonstrate three tasks: one-step-ahead prediction of Lorenz63 and Mackey-Glass chaotic time series (NMSE 0.07 and 0.06), nonlinear equalization of PAM4 signals (SER improved from 19% to 7%, compared with 11% for a linear equalizer), and classification of noisy sinusoidal, square, and sawtooth waveforms (100% accuracy). They conclude that the OPO provides ultrafast all-optical linear and nonlinear operations with sub-nanosecond latency and no OEO conversions.

Significance. If the results withstand scrutiny, the work is a valuable experimental demonstration of a nanophotonic OPO as a nonlinear recurrent reservoir at ~10 GHz, with a useful control (linear equalization) for the channel-equalization task. The device is fabricated on a TFLN platform and the experiments include error bars on the main metrics. However, the chaotic forecasting results lack baselines at the same sampling rate, the latency claim is not measured, and the 'all-optical' claim goes beyond what is demonstrated, since input encoding and output readout are electronic. These issues currently weaken the central claim of ultrafast all-optical nonlinear neuromorphic processing.

major comments (4)
  1. [Low latency time domain signal processing using OPONN; Fig. 2] The chaotic time-series prediction results (NMSE = 0.07 ± 0.017 for Lorenz63 and 0.06 ± 0.017 for Mackey-Glass) are reported without baselines. A persistence predictor x̂(n+1)=x(n) or a linear autoregressive model trained on the same data should be evaluated at the same ~10 GHz sampling rate. Because Lorenz63 and Mackey-Glass trajectories have characteristic timescales much longer than 0.1 ns, consecutive samples are strongly correlated, and a low one-step-ahead NMSE may reflect signal autocorrelation rather than the OPO's nonlinear computation. Please add these baselines and, if possible, multi-step-ahead predictions over Lyapunov times; this is necessary to support the claim that the OPO provides useful nonlinear processing in the chaotic benchmarks.
  2. [Abstract and Discussion] The claim that the OPONN is 'capable of achieving sub-nanosecond latencies' is not supported by a latency measurement or a stated definition. The reported experiments use an AWG-driven EOM for input encoding and a photodetector for output readout; end-to-end latency is not reported. Please either measure the latency with a defined input-to-output convention, or restrict the claim to the optical recurrent core and support it with the relevant roundtrip/response times.
  3. [Low latency time domain signal processing using OPONN] The sentence 'Training is performed in silico by singular value decomposition to obtain the optimal output weight matrix Wout' is ambiguous. If Wout is fitted to a simulated model of the OPO, the paper must describe how the simulation is calibrated to the device and why the experimental results then validate the model; if Wout is fitted to measured reservoir outputs, the phrase 'in silico' is misleading. Please specify the training data, the fitting procedure, and how the training/testing split avoids leakage. This clarification is needed to interpret all reported metrics.
  4. [Abstract, Fig. 1, and Discussion] The 'all-optical' framing overstates the demonstrated system. In the experiments, the input mask Win is applied electronically (the AWG prepares the masked waveform), and the output layer is formed by a fast photodetector followed by a digital weighted sum Wout. What is demonstrated is an all-optical recurrent nonlinear core, not an all-optical processor that eliminates OEO conversions. Please revise the abstract and discussion to state precisely which operations are optical and which are electronic.
minor comments (5)
  1. [Figure 1 caption] The words 'neurmorphic' and 'experiements' are typos; please correct them.
  2. [Abstract and Discussion] For the chaotic time-series tasks the paper reports NMSE, not accuracy; the summary statement 'success rates exceeding 93%' mixes metrics. Please use a consistent performance measure or explicitly define how accuracy is derived from NMSE.
  3. [Waveform classification, Fig. 4] The 100% accuracy is reported on a single test set of 300 waveforms with no confidence interval; please state the number of independent runs and the noise level used.
  4. [Low latency time domain signal processing using OPONN] The text calls the single OPO with delayed feedback a 'deep recurrent neural network'; unless there are multiple recurrent layers, this is a reservoir-computing architecture. Please align the terminology with the reservoir-computing literature (e.g., refs. 28-30) to avoid overstating the architectural depth.
  5. [Fig. 3 and accompanying text] The experimental SER values are given as mean ± standard deviation, but the number of symbols or trials is not stated; please add this information to allow the error bars to be interpreted.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the readout is trained on a training split and evaluated on a held-out split, which is standard reservoir computing rather than a hidden self-derived prediction.

full rationale

The paper's derivation chain is a standard reservoir-computing experiment: a fixed random input mask Win modulates pump pulses, the OPO provides recurrent nonlinear delayed dynamics, and only the output weights Wout are trained (by SVD) on a training portion of each benchmark and then evaluated on a separate testing portion. This is supervised fitting of a readout, not circularity, because the test data are not used in the fit and the metrics (NMSE, SER, classification accuracy) are computed on held-out data. The nonlinear role of the OPO is corroborated by an explicit control in the channel-equalization task, where a purely linear equalization gives SER = 11% whereas the OPONN gives SER = 7%, supporting the claim that the OPO contributes nonlinear processing. The chaotic time-series benchmarks lack a persistence or linear autoregressive baseline, and this is a legitimate benchmark-rigor concern about whether the tasks are nontrivial at the 10 GHz sampling rate; however, absence of a baseline is not circularity, because the predicted signal is not defined in terms of the fitted weights or the measured output by construction. No equation in the paper equates a predicted quantity with an input quantity, and no load-bearing premise is justified solely by a self-citation chain: citations to prior OPO device work [24,25] supply independent experimental characterization of the degenerate OPO and TFLN chip, and the machine-learning benchmarks are standard external tasks. Therefore the central derivation is self-contained with respect to circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claims rest on standard reservoir computing assumptions: the OPO cavity provides fading memory and the parametric nonlinearity creates a rich feature space. These are not directly characterized in the paper. The readout weights and input mask are fitted or randomly chosen, and the reported performance depends on them.

free parameters (3)
  • Readout weights Wout = not listed (optimized via SVD)
    The final accuracies are obtained after fitting Wout to training data. This is a standard trained parameter, and the reported performance depends on it.
  • Input mask Win = random values, seed not given
    Input weights are chosen randomly; the specific realization affects the reservoir states.
  • Task-specific hyperparameters (Nin, Nout, sampling rate) = Nin=3/5/5, Nout=10/21/3, 10 GHz
    Chosen by hand for each task; not optimized but influence performance.
assumptions (3)
  • domain assumption The OPO cavity provides fading memory (echo state property) sufficient for the tasks.
    The recurrent dynamics are assumed to act as a reservoir; no memory capacity measurement is presented.
  • domain assumption The parametric nonlinearity of the PPLN section acts as a nonlinear activation that enriches the reservoir states.
    The paper does not characterize the exact input-output nonlinear function of the OPO; it relies on its existence for the ML performance.
  • domain assumption A linear readout of the detected intensities suffices for the three tasks.
    The output layer is a linear weighted sum; the results validate this for the specific tasks, but it is an assumption about the reservoir's linear separability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ultrafast neuromorphic computing with nanophotonic optical parametric oscillators." pith.science (2026). https://pith.science/paper/NILXE5M5

@misc{pith2026250116604,
  author       = {Pith},
  title        = {Pith review of: Ultrafast neuromorphic computing with nanophotonic optical parametric oscillators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NILXE5M5}},
  note         = {Machine review of arXiv:2501.16604}
}
read the original abstract

Over the past decade, artificial intelligence (AI) has led to disruptive advancements in fundamental sciences and everyday technologies. Among various machine learning algorithms, deep neural networks have become instrumental in revealing complex patterns in large datasets with key applications in computer vision, natural language processing, and predictive analytics. On-chip photonic neural networks offer a promising platform that leverage high bandwidths and low propagation losses associated with optical signals to perform analog computations for deep learning. However, nanophotonic circuits are yet to achieve the required linear and nonlinear operations simultaneously in an all-optical and ultrafast fashion. Here, we report an ultrafast nanophotonic neuromorphic processor using an optical parametric oscillator (OPO) fabricated on thin-film lithium niobate (TFLN). The input data is used to modulate the optical pulses synchronously pumping the OPO. The consequent signal pulses generated by the OPO are coupled to one another via the nonlinear delayed dynamics of the OPO, thus forming the internal nodes of a deep recurrent neural network. We use such a nonlinearly coupled OPO network for chaotic time series prediction, nonlinear error correction in a noisy communication channel, as well as noisy waveform classification and achieve accuracies exceeding 93% at an operating clock rate of ~ 10 GHz. Our OPO network is capable of achieving sub-nanosecond latencies, a timescale comparable to a single clock cycle in state-of-the-art digital electronic processors. By circumventing the need for optical-electronic-optical (OEO) conversions, our ultrafast nanophotonic neural network paves the way for the next generation of compact all-optical neuromorphic processors with ultralow latencies and high energy efficiencies.

Figures

Figures reproduced from arXiv: 2501.16604 by the authors.

Figure 1
Figure 1. Nanophotonic OPO-based neuromorphic processor. Schematic view of the nanophotonic OPO utilized in our experiements for neuromorphic computing. The top panel depicts the microscope image of the device (more information about the chip design can be found in [25]). The central part of the racetrack resonator is removed for clarity. The structure is designed in such a way that only the degenerate signal/idler resonate i… view at source ↗
Figure 2
Figure 2. Chaotic time series prediction using OPONN. a, Numerically calculated values of the x(t) signal associated with the Lorenz63 model (target) and the experimentally measured signal at the output of our OPONN (predicted). As shown in the figure, the entire span of the signal with a ∼ 500 ns duration is split into two equal sections for training (the shaded area) and testing. As evident, the predictions of the OPONN clo… view at source ↗
Figure 3
Figure 3. Nonlinear channel equalization of PAM4 signals using ultrafast nonlinear re￾sponse of the OPO network. a, Simulated SERs obtained using direct detection versus those achieved after OPONN equalization for various SNR levels. b, Experimentally obtained eye diagrams without (left) and with (right) OPONN equalization. we also considered a purely linear equalization applied directly to the input pump, for which a SER = 1… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Noisy waveform classification task. a, We use the AWG to encode samples rep￾resenting two cycles of periodic waveforms randomly chosen from three different classes of sawtooth, sinusoid and square signals with additive noise on the input pump (left). The mea￾sured sign…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 25 canonical work pages

  1. [1]

    & Hinton, G

    LeCun, Y ., Bengio, Y . & Hinton, G. Deep learning. Nature 521, 436–444 (2015)

    work page 2015

  2. [2]

    Ashtiani, F., Geers, A. J. & Aflatouni, F. An on-chip photonic deep neural network for image classification. Nature 606, 501–506 (2022)

    2022

  3. [3]

    Chen, Y . et al. All-analog photoelectronic chip for high-speed vision tasks. Nature 623, 48–57 (2023)

    work page 2023

  4. [4]

    Pai, S. et al. Experimentally realized in situ backpropagation for deep learning in photonic neural networks. Science 380, 398–404 (2023)

    2023

  5. [5]

    Wetzstein, G. et al. Inference in artificial intelligence with deep optics and photonics. en. Nature 588, 39–47 (2020)

    work page 2020

  6. [6]

    Shastri, B. J. et al. Photonics for artificial intelligence and neuromorphic computing. Na- ture Photonics 15, 102–114 (2021)

    work page 2021

  7. [7]

    Vaswani, A. et al. Attention is All you Needin Advances in Neural Information Processing Systems 30 (Curran Associates, Inc., 2017)

    work page 2017

  8. [8]

    Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021)

    2021

Show all 33 references
  1. [9]

    H., Wu, Y ., Le, Q

    Trinh, T. H., Wu, Y ., Le, Q. V ., He, H. & Luong, T. Solving olympiad geometry without human demonstrations. Nature 625, 476–482 (2024)

    2024

  2. [10]

    Sevilla, J. et al. Compute Trends Across Three Eras of Machine Learning in 2022 Inter- national Joint Conference on Neural Networks (IJCNN) (2022), 1–8

    2022

  3. [11]

    McMahon, P. L. The physics of optical computing. Nature Reviews Physics 5, 717–734 (2023)

    2023

  4. [12]

    Feng, H. et al. Integrated lithium niobate microwave photonic processing engine. Nature 627, 80–87 (2024)

    2024

  5. [13]

    Shen, Y . et al. Deep learning with coherent nanophotonic circuits. Nature Photonics 11. Publisher: Nature Publishing Group, 441–446 (2017)

    2017

  6. [14]

    SeyedinNavadeh, S. et al. Determining the optimal communication channels of arbitrary optical systems using integrated photonic processors.Nature Photonics18, 149–155 (2024). 13

    2024

  7. [15]

    Chen, Z. et al. Deep learning with coherent VCSEL neural networks. Nature Photonics 17, 723–730 (2023)

    2023

  8. [16]

    D., Bhaskaran, H

    Feldmann, J., Youngblood, N., Wright, C. D., Bhaskaran, H. & Pernice, W. H. P. All- optical spiking neurosynaptic networks with self-learning capabilities. Nature 569, 208– 214 (2019)

    2019

  9. [17]

    Feldmann, J. et al. Parallel convolutional processing using an integrated photonic tensor core. Nature 589, 52–58 (2021)

    2021

  10. [18]

    Br ¨uckerhoff-Pl¨uckelmann, F. et al. Event-driven adaptive optical neural network. Science Advances 9, eadi9127 (2023)

    2023

  11. [19]

    Li, G. H. Y . et al. All-optical ultrafast ReLU function for energy-efficient nanophotonic deep learning. Nanophotonics 12, 847–855 (2023)

    2023

  12. [20]

    Wright, L. G. et al. Deep physical neural networks trained with backpropagation. Nature 601, 549–555 (2022)

    2022

  13. [21]

    & Lvovsky, A

    Spall, J., Guo, X. & Lvovsky, A. I. Hybrid training of optical neural networks. Optica 9, 803–811 (2022)

    2022

  14. [22]

    Roques-Carmes, C. et al. Biasing the quantum vacuum to control macroscopic probability distributions. Science 381, 205–209 (2023)

    2023

  15. [23]

    Choi, S. et al. Photonic probabilistic machine learning using quantum vacuum noise. Na- ture Communications 15, 7760 (2024)

    2024

  16. [24]

    Roy, A. et al. Visible-to-mid-IR tunable frequency comb in nanophotonics. Nature Com- munications 14, 6549 (2023)

    2023

  17. [25]

    Ledezma, L. et al. Octave-spanning tunable infrared parametric oscillators in nanopho- tonics. Science Advances 9. Publisher: American Association for the Advancement of Science, eadf9711. (2024) (July 2023)

    2024

  18. [26]

    Lorenz, E. N. Deterministic Nonperiodic Flow. Journal of the Atmospheric Sciences 20, 130–141 (1963)

    1963

  19. [27]

    Mackey, M. C. & Glass, L. Oscillation and Chaos in Physiological Control Systems. Sci- ence 197. Publisher: American Association for the Advancement of Science, 287–289 (1977)

    1977

  20. [28]

    & Haas, H

    Jaeger, H. & Haas, H. Harnessing Nonlinearity: Predicting Chaotic Systems and Saving Energy in Wireless Communication. Science 304, 78–80 (Apr. 2004)

    2004

  21. [29]

    Vinckier, Q. et al. High-performance photonic reservoir computer based on a coherently driven passive cavity. Optica 2, 438–446 (2015)

    2015

  22. [30]

    & Massar, S

    Lupo, A., Picco, E., Zajnulina, M. & Massar, S. Deep photonic reservoir computer based on frequency multiplexing with fully analog connection between layers. Optica 10, 1478–1485 (2023). 14

    2023

  23. [31]

    & Yoon, S

    Min, S., Lee, B. & Yoon, S. Deep learning in bioinformatics. Briefings in Bioinformatics 18, 851–869 (2017)

    2017

  24. [32]

    Dong, B. et al. Higher-dimensional processing using a photonic tensor core with continuous- time data. Nature Photonics 17, 1080–1088 (2023)

    2023

  25. [33]

    Mousavi, S. M. & Beroza, G. C. Deep-learning seismology. Science 377, eabm4470 (2022). 15

    2022

Pith tools

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