Pith. sign in

REVIEW 3 major objections 7 minor 1 cited by

Nanosecond-latency all-optical fiber sensing with in-sensor computing

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Fiber-optic measurements can be read straight from light intensity, with under 3 nanoseconds of delay and no electronic processing.

desk verdict Solid all-optical demodulation demonstration with real experimental weight; needs honest claim-trimming on resolution, generalization, and latency before publication. read the letter →

arxiv 2507.15376 v1 pith:UWB3HMJA submitted 2025-07-21 physics.optics

Yu Tao , Yangyang Wan , Ziwen Long , Wenjia Zhang , Jiangbing Du , Zuyuan He This is my paper · ORCID
classification physics.optics
keywords all-opticalfibersensingin-sensorcomputingspecklediffractiveopticalnetworkmultimodestraintorsion
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 claims that fiber-optic sensing can shed its electronic demodulation stage entirely. The authors route light from a fiber sensor through a scattering multimode fiber and then through a trained diffractive optical network, so that the physical quantity being measured—strain, torsion angle, vibration—is mapped directly onto the intensity of light at a detector. A photodetector reading alone then gives the measured value, with a reported demodulation delay below $3\,\text{ns}$ and no computer or digital signal processing in the loop. If this holds, conventional interrogators, spectrometers, and electronic processors become optional, and fiber sensing speed becomes limited mainly by detector bandwidth.

What carries the argument

The load-bearing mechanism is the pairing of a scattering medium with a trained diffractive optical network. The scattering medium (a multimode fiber) performs a high-dimensional nonlinear projection that turns tiny changes in wavelength, polarization, or mode content into large, distinguishable speckle changes; the optical diffraction network, implemented with a spatial light modulator and trained by a genetic algorithm, applies phase modulation that spatially re-routes those speckles so that total intensity in a target region equals the measurand. Two properties carry the argument: the deterministic "memory effect," in which speckle patterns decorrelate gradually and monotonically as the measurand changes, allowing interpolation from sparse training states; and linear intensity readout, so that the sensing result is available immediately at the photodetector.

What would settle it

Take the FBG strain setup, apply a dense sweep of strain values between and beyond the training states, and measure both the speckle cross-correlation and the optical-network output intensity. If the speckle correlation versus strain is non-monotonic, has plateaus, or exhibits discontinuities, or if the output intensity deviates from the calibration line by more than the reported RMSE at any untrained value, the claimed general linear mapping is falsified.

Watch

Extended reading notes

Core claim

The central discovery is an architecture, AOFS-IC, that performs sensing demodulation entirely in the optical domain. A scattering medium, here a multimode fiber, converts small optical-field changes—wavelength shift, polarization change, mode coupling—into high-dimensional speckle patterns, and a spatial-light-modulator-based diffractive optical network, trained end-to-end with a genetic algorithm, transforms those speckles so that the intensity in a designated output region is linearly proportional to the measurand. The paper demonstrates this for FBG strain sensing (RMSE $2.7554\,\mu\varepsilon$ over $150\,\mu\varepsilon$ and $0.0688\,\text{m}\varepsilon$ over $2.5\,\text{m}\varepsilon$), for torsional state classification with 100% accuracy, for simultaneous strain and torsion readout, for nanoscale strain down to $1.6160\,\text{n}\varepsilon$ RMS over $95\,\text{n}\varepsilon$, and for 3-DOF robotic-arm joint monitoring. The claimed result is that the trained optical network generalizes from only a few training states to the whole measurement range thanks to the deterministic, gradually decorrelating speckle response (the memory effect).

Load-bearing premise

The whole scheme depends on the speckle pattern evolving deterministically, monotonically, and without jumps as the measurand changes, so that a network trained on a handful of states can interpolate every value in between.

Editorial extensions

If this is right

  • A fiber sensor linked to an AOFS-IC module can report a physical quantity with a total demodulation latency below $3\,\text{ns}$, which the paper estimates as more than two orders of magnitude faster than conventional electronic demodulation.
  • Because the readout is just light intensity, the usable sensing bandwidth is set by the photodetector rather than by any computing hardware; the paper shows $10\,\text{kHz}$ and $150\,\text{kHz}$ vibrations recovered with SNR $59\,\text{dB}$ and $27\,\text{dB}$, the latter limited by the PD's $90\,\text{kHz}$ bandwidth.
  • Multiple sensors or multiple measurands can be decoded simultaneously by training the optical network to focus each sensing signal onto a separate spatial region of the detector plane, as demonstrated for simultaneous strain and torsion on one multimode fiber.
  • The same architecture works across sensor types and measurands—FBG wavelength shifts, MMF torsion and stretching, SMF polarization rotation, and robot-arm joint bending—implying a general all-optical demodulation layer rather than a single-purpose device.
  • Removing electronic demodulation eliminates the power consumption and latency of interrogators and computers from fiber sensing systems, which the paper argues makes dense, large-scale sensor arrays more practical.

Reading between the lines

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

  • The training-data sparsity suggests a testable scaling law: the required number of training states should grow with the width of the measurement range relative to the decorrelation length of the speckle pattern, so measuring that length directly could predict where the linear mapping breaks down.
  • Replacing the spatial light modulator with passive etched phase plates—which the paper names as a possibility—would remove the only actively powered optical component, making the sensing head fully passive and potentially deployable in hard-to-reach locations.
  • The same speckle-to-intensity mapping could be repurposed as an all-optical spectrometer, polarization analyzer, or temperature sensor whenever the measurand leaves a deterministic fingerprint in the speckle pattern, a direction the paper gestures at with its spectrometer and polarization-analyzer extensions.
  • The accuracy-versus-range trade-off the paper reports looks like a consequence of the memory effect, not merely an engineering fix: extending the dynamic range compresses the intensity response per unit measurand, so system design must choose a range matched to the decorrelation curve.
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 / 7 minor

Summary. The manuscript proposes an all-optical fiber sensing architecture (AOFS-IC) in which the sensing signal from a fiber sensor passes through a multimode-fiber scattering medium and a trained single-layer phase mask (SLM), so that the physical measurand (strain, torsion, bending-related joint angle, etc.) is read directly from detected optical intensity without digital demodulation. The authors report strain regression with RMSEs of 2.7554 µε over 150 µε and 0.0688 mε over 2.5 mε, torsion classification with 100% accuracy on 9 discrete angles, dual-parameter strain/torsion demultiplexing, PD-based dynamic strain measurements up to 150 kHz, and 3-DOF robotic arm joint-angle estimation. The claimed advantages are sub-3 ns demodulation delay and elimination of electronic processing hardware.

Significance. If the claims hold, the architecture is a useful demonstration of in-sensor optical computing for fiber sensing: it moves demodulation into the optical domain, supports spatial multiplexing, and shows interpolation to untrained measurand values. The experimental work is substantial and mostly internally consistent, including repeated measurements, error histograms, comparisons with simulation, and multiple sensor types. The genetic-algorithm in-situ training avoids requiring a full physical model, and the PD-based high-sensitivity measurements provide a concrete path beyond camera-based demonstrations. However, the headline claims of sub-nano strain resolution and nanosecond latency are overstated relative to the reported RMSE and the 90 kHz PD bandwidth, and the interpolation argument from four training states needs additional characterization.

major comments (3)
  1. [Abstract; §2.5, Fig. 5(c)] The abstract's claim of 'sub-nano strain resolution' is not supported by the reported RMSE of 1.6160 nε over a 95 nε range, since 1.6160 nε is larger than 1 nε. Please replace 'sub-nano' with 'nanostrain-level' or provide a measurement with RMSE below 1 nε; as written, the headline overstates the demonstrated sensitivity.
  2. [§4.4; §2.5] The '<3 ns demodulation delay' is an optical-propagation estimate for the 0.5 m encoding MMF and the free-space module (τ_encode = 2.44 ns, τ_compu = 0.08 ns), but it excludes the photodetector. The experimental system uses an InGaAs PD with 90 kHz bandwidth, which limits the recovered 150 kHz signal to a 27 dB SNR and visibly distorts the waveform; a 90 kHz bandwidth corresponds to a response time of order 1.8 μs. Thus the demonstrated end-to-end sensing latency is microseconds, not nanoseconds. Please either measure and report the true step response of the PD-based system, or explicitly restrict the '<3 ns' claim to the optical computing module and revise the abstract's 'nanosecond-latency' wording.
  3. [§2.2, Eq. (2), Fig. 2(c)] The generalization argument that training on four strain states yields a linear mapping over the full range is not fully established. Equation (2) constrains the ODN output only at the training states; it imposes no monotonicity or smoothness constraint on the response between them, and the cited memory effect (ref. 37) concerns spatial shift invariance rather than monotonic spectral decorrelation. Figure 2(c) shows the autocorrelation of speckle patterns decreasing with strain, but monotonic decorrelation of the speckle does not by itself guarantee that the trained output intensity is a single-valued monotonic function of strain. Since Eq. (1) assumes a single-valued linear readout, please report a fine-grained measured output-intensity-versus-strain curve across the full range, not only the 30 sampled states, and quantify monotonicity/no ambiguity, or add a constraint and discussion that rules out non-monotonic regions.
minor comments (7)
  1. [§2.2] The main text states the high-accuracy strain resolution as 2.7754 µε while the Fig. 2(e) caption reports RMSE = 2.7554 µε; please reconcile these numbers.
  2. [Introduction] 'optical nerual network' should be 'optical neural network'.
  3. [Fig. 3 caption] 'ig. 3(c)' should be 'Fig. 3(c)'.
  4. [§4.5] 'espcially' should be 'especially' in the noise analysis paragraph.
  5. [Fig. 4 caption] 'Detecor 1 Detector 2' contains a typo and should read 'Detector 1 Detector 2'.
  6. [§4.2] Please define I_obj and describe how the normalized intensity Inorm is computed (which spatial region and normalization procedure) before Eq. (1); currently the reader must infer this from the figures.
  7. [§4.2] For reproducibility, please include the GA population size, number of generations, and number of training states used in each experiment, or at least state explicitly that these details are given in Supplementary Note 6.

Circularity Check

1 steps flagged · score 4.0 of 10

Torsion-classification accuracy is evaluated on the same nine states used to train the ODN; the main strain regression is tested on untrained states and is not circular.

  1. fitted input called prediction [Section 2.3, Fig. 3(e–f), Methods Eq. (4)]
    "By appropriately adjusting the objective function during ODN training, the speckles corresponding to different torsion angles can be focused onto distinct spatial positions. ... Both the experimental and simulated confusion matrices demonstrate ideal classification accuracy, as shown in Fig. 3(e–f). All predicted torsion angles perfectly match GT, confirming the capability to reliably distinguish discrete torsion states with 45◦ resolution steps."

    The nine discrete torsion angles are the classes encoded as one-hot targets in Eq. (4), and the training loss is minimized by matching those targets. The paper then reports confusion matrices and 100% accuracy for the same nine angles (0°, 45°, ..., 360°) without describing any held-out or untrained torsion angle. Thus the reported classification performance measures how well the SLM phase pattern fits the training classes, not how well the system predicts unseen torsion states. The detector-region mapping is the optimization objective itself, so the accuracy is partly forced by construction. This is a secondary result; the central strain-regression claim was trained on only four states and evaluated on 30 continuous states, so it is not similarly circular.

full rationale

The main strain-regression chain is self-contained and not circular. The ODN is trained on a small set of strain states via Eq. (2), and the evaluation then uses 30 continuous strain states, most of which were not used in training; the reported RMSE therefore measures interpolation to unseen measurand values, not merely the training fit. Eq. (1) is a calibration rescaling, but it is not the source of the claimed linearity: the empirical claim is that the trained optical output remains linear between sparse training points, which the experiment directly tests. The memory-effect argument (external ref. 37) and the speckle-correlation curve in Fig. 2(c) provide a physical rationale for interpolation, not a circular input. Self-citations (refs. 39–41) support only a peripheral statement in the Discussion about MMF length versus spectral resolution and are not load-bearing for the central architecture. The one partial circularity is the torsion classification in Sec. 2.3: the nine angles used to define the one-hot classification targets in Eq. (4) are the same nine angles evaluated in the confusion matrices, and no held-out discrete angle is reported. The claimed 100% classification accuracy is therefore a training-set result rather than evidence of generalization to untrained discrete states. Concerns about possible non-monotonic speckle decorrelation between training strain states are correctness and generalization risks, not circularity, because the strain experiment does test untrained states.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The work rests on several domain assumptions about speckle stability and deterministic decorrelation, and on the fitness of a genetically trained phase mask. The central mapping is fitted to training data, and the experimental results verify interpolation. No new physical entities are introduced.

free parameters (3)
  • SLM phase pattern (trained) = position-dependent phase profile, not specified
    Each sensing task requires training the spatial light modulator phase pattern via genetic algorithm to map speckle intensity to the desired output; this is a high-dimensional set of fitted parameters.
  • Calibration range [M_min, M_max] = e.g., 0-150 µε, 0-2.5 mε, 0-95 nε, 0-30 degrees / 0-60 degrees
    The range over which the linear regression is applied is chosen by the experimenter per task, affecting dynamic range and resolution.
  • MMF length for scattering medium = 0.5 m (also 5 m in one setup)
    The length of the multimode fiber used as scattering medium is selected to balance resolution and stability; shorter lengths reduce encoding delay.
assumptions (6)
  • domain assumption Memory effect in the scattering medium ensures monotonic, deterministic decorrelation of speckle patterns with changes in the measurand.
    The linear regression from training states to untested states relies on gradual decorrelation (Fig. 2c); if decorrelation is non-monotonic or discontinuous, unseen strain values could be misestimated. Mentioned in Section 2.2 and Supplementary Note 2.
  • domain assumption Speckle patterns remain stable over the measurement period (correlation 0.98 over 24 hours under temperature control).
    Section 4.5 states the speckle correlation degraded to 0.98 over 24 h; this stability is assumed to hold during experiments, otherwise mapping would drift.
  • domain assumption The SLM provides pure phase modulation with a linear 2π response at 1550 nm.
    Methods Section 4.1 states the SLM is a reflective phase modulator with linear 2π phase response; any nonlinearity or phase flicker would perturb the mapping.
  • ad hoc to paper The genetic algorithm converges to a phase pattern that approximates the desired linear mapping.
    No proof of convergence or optimality is given; the algorithm is described in Supplementary Note 6 which is not provided in this review.
  • domain assumption Photodetector intensity is linearly related to the measurand within the calibrated range.
    Equation (1) assumes a linear regression; if the photodetector saturates or the mapping is nonlinear outside training points, accuracy degrades.
  • domain assumption The calibration setup (translation stage, rotation mount, PZT) provides accurate ground truth.
    The calibration setup has strain accuracy 2 µε and angular accuracy 0.5 degrees, cited in Section 2.4; the reported RMSEs approach these baselines.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nanosecond-latency all-optical fiber sensing with in-sensor computing." pith.science (2026). https://pith.science/paper/UWB3HMJA

@misc{pith2026250715376,
  author       = {Pith},
  title        = {Pith review of: Nanosecond-latency all-optical fiber sensing with in-sensor computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UWB3HMJA}},
  note         = {Machine review of arXiv:2507.15376}
}
read the original abstract

Optical fiber sensing plays a crucial role in modern measurement systems and holds significant promise for a wide range of applications. This potential, though, has been fundamentally constrained by the intrinsic latency and power limitations associated with electronic signal processing. Here, we propose an all-optical fiber sensing architecture with in-sensor computing (AOFS-IC) that achieves fully optical-domain sensing signal demodulation at the speed of light. By integrating a scattering medium with an optimized diffractive optical network, AOFS-IC enables linear mapping of physical perturbations to detected intensity, and sensing results can be directly read out without electronic processing. The proposed system maintains high accuracy across various sensing tasks, providing sub-nano strain resolution and 100% torsional angle classification accuracy, as well as multiplexed sensing of multiple physical quantities, and performing multi-degree-of-freedom robot arm monitoring. AOFS-IC eliminates computing hardware requirements while providing <3 ns demodulation delay, which is more than 2 orders of magnitude faster than conventional fiber optic sensing systems. This work demonstrates the potential of next-generation optical sensing systems empowered by all-optical computing, and paves the way for expanded applications of fiber sensing through the integration of fully optical components, ultrafast measurement speed, and low power consumption.

Figures

Figures reproduced from arXiv: 2507.15376 by the authors.

Figure 1
Figure 1. All-optical fiber sensing architecture with in-sensor computing (AOFS-IC). (a) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. All-optical estimation of strain over varying dynamic ranges. (a) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. All-optical classification of discrete torsion states. (a) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Simultaneous all-optical sensing of strain and torsion. (a) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: High-speed and high-accuracy all-optical strain sensing. (a) [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Multi-degree-of-freedom (DOF) joint angle estimation on a robotic arm using AOFS-IC. (a) [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Self-Pulsing Microring Resonator Networks for Bandwidth-Efficient Event Detection in an Optical Fiber Sensor

    physics.optics 2026-06 unverdicted novelty 6.0 of 10

    Experimental demonstration that self-pulsing microring resonator networks retain fiber sensor perturbation information, reducing required sampling rate by at least one order of magnitude.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · cited by 1 Pith paper

  1. [1]

    Elsherif, A

    M. Elsherif, A. E. Salih, M. G. Mu \ n oz, F. Alam, B. AlQattan, D. S. Antonysamy, M. F. Zaki, A. K. Yetisen, S. Park, T. D. Wilkinson, et al. , ``Optical fiber sensors: Working principle, applications, and limitations,'' Advanced Photonics Research , vol. 3, no. 11, p. 2100371, 2022

    work page 2022

  2. [2]

    Pend \ a o and I

    C. Pend \ a o and I. Silva, ``Optical fiber sensors and sensing networks: overview of the main principles and applications,'' Sensors , vol. 22, no. 19, p. 7554, 2022

    work page 2022

  3. [3]

    N. J. Lindsey and E. R. Martin, ``Fiber-optic seismology,'' Annual Review of Earth and Planetary Sciences , vol. 49, no. 1, pp. 309--336, 2021

    work page 2021

  4. [4]

    C. Du, S. Dutta, P. Kurup, T. Yu, and X. Wang, ``A review of railway infrastructure monitoring using fiber optic sensors,'' Sensors and Actuators A: Physical , vol. 303, p. 111728, 2020

    work page 2020

  5. [5]

    T. Wu, G. Liu, S. Fu, and F. Xing, ``Recent progress of fiber-optic sensors for the structural health monitoring of civil infrastructure,'' Sensors , vol. 20, no. 16, p. 4517, 2020

    work page 2020

  6. [6]

    Ashry, Y

    I. Ashry, Y. Mao, B. Wang, F. Hveding, A. Y. Bukhamsin, T. K. Ng, and B. S. Ooi, ``A review of distributed fiber--optic sensing in the oil and gas industry,'' Journal of Lightwave Technology , vol. 40, no. 5, pp. 1407--1431, 2022

    work page 2022

  7. [7]

    Z. Yi, Z. Chen, K. Yin, L. Wang, and K. Wang, ``Sensing as the key to the safety and sustainability of new energy storage devices,'' Protection and Control of Modern Power Systems , vol. 8, no. 2, pp. 1--22, 2023

    work page 2023

  8. [8]

    Shimizu, L.-C

    Y. Shimizu, L.-C. Chen, D. W. Kim, X. Chen, X. Li, and H. Matsukuma, ``An insight into optical metrology in manufacturing,'' Measurement Science and Technology , vol. 32, no. 4, p. 042003, 2021

    work page 2021

Show all 44 references
  1. [9]

    F. Zhou, X. Fu, S. Chen, E. Kim, and M. B. Jun, ``Fiber optic sensor for smart manufacturing,'' International Journal of Precision Engineering and Manufacturing-Smart Technology , vol. 1, no. 2, 2023

    2023

  2. [10]

    Venketeswaran, N

    A. Venketeswaran, N. Lalam, J. Wuenschell, P. R. Ohodnicki Jr, M. Badar, K. P. Chen, P. Lu, Y. Duan, B. Chorpening, and M. Buric, ``Recent advances in machine learning for fiber optic sensor applications,'' Advanced Intelligent Systems , vol. 4, no. 1, p. 2100067, 2022

    2022

  3. [11]

    Gherlone, P

    M. Gherlone, P. Cerracchio, and M. Mattone, ``Shape sensing methods: Review and experimental comparison on a wing-shaped plate,'' Progress in Aerospace Sciences , vol. 99, pp. 14--26, 2018

    2018

  4. [12]

    L. Zhu, G. Sun, W. Bao, Z. You, F. Meng, and M. Dong, ``Structural deformation monitoring of flight vehicles based on optical fiber sensing technology: A review and future perspectives,'' Engineering , vol. 16, pp. 39--55, 2022

    2022

  5. [13]

    B. Wang, L. Wang, N. Guo, Z. Zhao, C. Yu, and C. Lu, ``Deep neural networks assisted botda for simultaneous temperature and strain measurement with enhanced accuracy,'' Optics express , vol. 27, no. 3, pp. 2530--2543, 2019

    2019

  6. [14]

    Li, Z.-W

    B. Li, Z.-W. Tan, P. P. Shum, C. Wang, Y. Zheng, and L. J. Wong, ``Dilated convolutional neural networks for fiber bragg grating signal demodulation,'' Optics Express , vol. 29, no. 5, pp. 7110--7123, 2021

    2021

  7. [15]

    Wang, Z.-Q

    Q. Wang, Z.-Q. Li, Y.-J. Fan, and B. Zhou, ``High precision and speed demodulation algorithm for optical fiber fabry-perot temperature sensor,'' IEEE Transactions on Instrumentation and Measurement , 2025

    2025

  8. [16]

    Hu and W

    C. Hu and W. Bai, ``High-speed interrogation for large-scale fiber bragg grating sensing,'' sensors , vol. 18, no. 2, p. 665, 2018

    2018

  9. [17]

    Cheung, Z

    Y. Cheung, Z. Jing, Y. Guo, Y. Liu, Q. Liu, A. Li, D.-P. Zhou, and W. Peng, ``Parallel cross-correlation algorithm for the fast interrogation of high-finesse fiber-optic fabry-perot sensors,'' Journal of Lightwave Technology , vol. 41, no. 13, pp. 4522--4530, 2023

    2023

  10. [18]

    H. Wang, T. Zhai, Y. Wang, Y. Liu, R. Zhou, X. Peng, and Z. Zhang, ``Real-time sensing approach for optical frequency domain reflectometry using an fpga-based high-speed demodulation algorithm,'' Optics Express , vol. 32, no. 19, pp. 33247--33261, 2024

    2024

  11. [19]

    P. L. McMahon, ``The physics of optical computing,'' Nature Reviews Physics , vol. 5, no. 12, pp. 717--734, 2023

    2023

  12. [20]

    Zhang, Q

    Y. Zhang, Q. Zhang, H. Yu, Y. Zhang, H. Luan, and M. Gu, ``Memory-less scattering imaging with ultrafast convolutional optical neural networks,'' Science Advances , vol. 10, no. 24, p. eadn2205, 2024

    2024

  13. [21]

    F. Xia, K. Kim, Y. Eliezer, S. Han, L. Shaughnessy, S. Gigan, and H. Cao, ``Nonlinear optical encoding enabled by recurrent linear scattering,'' Nature Photonics , vol. 18, no. 10, pp. 1067--1075, 2024

    2024

  14. [22]

    Oguz, J.-L

    I. Oguz, J.-L. Hsieh, N. U. Dinc, U. Te g in, M. Yildirim, C. Gigli, C. Moser, and D. Psaltis, ``Programming nonlinear propagation for efficient optical learning machines,'' Advanced Photonics , vol. 6, no. 1, pp. 016002--016002, 2024

    2024

  15. [23]

    Z. Xue, T. Zhou, Z. Xu, S. Yu, Q. Dai, and L. Fang, ``Fully forward mode training for optical neural networks,'' Nature , vol. 632, no. 8024, pp. 280--286, 2024

    2024

  16. [24]

    Z. Wang, H. Chen, J. Li, T. Xu, Z. Zhao, Z. Duan, S. Gao, and X. Lin, ``Opto-intelligence spectrometer using diffractive neural networks,'' Nanophotonics , vol. 13, no. 20, pp. 3883--3893, 2024

    2024

  17. [25]

    Y. Guan, H. Li, Y. Zhang, Y. Qiu, L. Zhang, X. Ji, H. Wang, Q. Chen, L. Ma, X. Wang, et al. , ``Photon-efficient camera with in-sensor computing,'' Nature Communications , vol. 16, no. 1, p. 3201, 2025

    2025

  18. [26]

    Y. Luo, Y. Zhao, J. Li, E. C etinta s , Y. Rivenson, M. Jarrahi, and A. Ozcan, ``Computational imaging without a computer: seeing through random diffusers at the speed of light,'' elight , vol. 2, no. 1, p. 4, 2022

    2022

  19. [27]

    Mengu and A

    D. Mengu and A. Ozcan, ``All-optical phase recovery: diffractive computing for quantitative phase imaging,'' Advanced Optical Materials , vol. 10, no. 15, p. 2200281, 2022

    2022

  20. [28]

    T. Wang, M. M. Sohoni, L. G. Wright, M. M. Stein, S.-Y. Ma, T. Onodera, M. G. Anderson, and P. L. McMahon, ``Image sensing with multilayer nonlinear optical neural networks,'' Nature Photonics , vol. 17, no. 5, pp. 408--415, 2023

    2023

  21. [29]

    T. Yan, T. Zhou, Y. Guo, Y. Zhao, G. Shao, J. Wu, R. Huang, Q. Dai, and L. Fang, ``Nanowatt all-optical 3d perception for mobile robotics,'' Science Advances , vol. 10, no. 27, p. eadn2031, 2024

    2024

  22. [30]

    H. Yu, Z. Huang, S. Lamon, B. Wang, H. Ding, J. Lin, Q. Wang, H. Luan, M. Gu, and Q. Zhang, ``All-optical image transportation through a multimode fibre using a miniaturized diffractive neural network on the distal facet,'' Nature Photonics , pp. 1--8, 2025

    2025

  23. [31]

    J. Hu, D. Mengu, D. C. Tzarouchis, B. Edwards, N. Engheta, and A. Ozcan, ``Diffractive optical computing in free space,'' Nature Communications , vol. 15, no. 1, p. 1525, 2024

    2024

  24. [32]

    C. Li, J. Tang, C. Cheng, L. Cai, and M. Yang, ``Fbg arrays for quasi-distributed sensing: A review,'' Photonic Sensors , vol. 11, pp. 91--108, 2021

    2021

  25. [33]

    T. W. Kohlgraf-Owens and A. Dogariu, ``Transmission matrices of random media: means for spectral polarimetric measurements,'' Optics letters , vol. 35, no. 13, pp. 2236--2238, 2010

    2010

  26. [34]

    S. Yuan, C. Ma, E. Fetaya, T. Mueller, D. Naveh, F. Zhang, and F. Xia, ``Geometric deep optical sensing,'' Science , vol. 379, no. 6637, p. eade1220, 2023

    2023

  27. [35]

    Redding, M

    B. Redding, M. Alam, M. Seifert, and H. Cao, ``High-resolution and broadband all-fiber spectrometers,'' Optica , vol. 1, no. 3, pp. 175--180, 2014

    2014

  28. [36]

    Michalewicz, Genetic algorithms+ data structures= evolution programs

    Z. Michalewicz, Genetic algorithms+ data structures= evolution programs . springer science & business media, 1996

    1996

  29. [37]

    S. Li, S. A. Horsley, T. Tyc, T. C i z m \'a r, and D. B. Phillips, ``Memory effect assisted imaging through multimode optical fibres,'' Nature communications , vol. 12, no. 1, p. 3751, 2021

    2021

  30. [38]

    Redding, S

    B. Redding, S. M. Popoff, and H. Cao, ``All-fiber spectrometer based on speckle pattern reconstruction,'' Optics express , vol. 21, no. 5, pp. 6584--6600, 2013

    2013

  31. [39]

    Y. Wan, X. Fan, S. Wang, Z. Zhang, S. Zhao, and Z. He, ``Wavemeter capable of simultaneously achieving ultra-high resolution and broad bandwidth by using rayleigh speckle from single mode fiber,'' Journal of Lightwave Technology , vol. 39, no. 7, pp. 2223--2229, 2021

    2021

  32. [40]

    Y. Wan, X. Fan, and Z. He, ``Review on speckle-based spectrum analyzer,'' Photonic Sensors , vol. 11, no. 2, pp. 187--202, 2021

    2021

  33. [41]

    Y. Wan, X. Fan, B. Xu, and Z. He, ``Reconstructive spectrum analyzer with high-resolution and large-bandwidth using physical-model and data-driven model combined neural network,'' Laser & Photonics Reviews , vol. 17, no. 7, p. 2201018, 2023

    2023

  34. [42]

    T. Zhou, X. Lin, J. Wu, Y. Chen, H. Xie, Y. Li, J. Fan, H. Wu, L. Fang, and Q. Dai, ``Large-scale neuromorphic optoelectronic computing with a reconfigurable diffractive processing unit,'' Nature Photonics , vol. 15, no. 5, pp. 367--373, 2021

    2021

  35. [43]

    Spall, X

    J. Spall, X. Guo, and A. I. Lvovsky, ``Hybrid training of optical neural networks,'' Optica , vol. 9, no. 7, pp. 803--811, 2022

    2022

  36. [44]

    Zheng, Z

    Z. Zheng, Z. Duan, H. Chen, R. Yang, S. Gao, H. Zhang, H. Xiong, and X. Lin, ``Dual adaptive training of photonic neural networks,'' Nature Machine Intelligence , vol. 5, no. 10, pp. 1119--1129, 2023

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.