Pith. sign in

REVIEW 3 major objections 4 minor 31 references

Multi-dimensional optical neural network

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper proposes that mode-division multiplexing can be layered on wavelength-division multiplexing to increase the input dimension of micro-ring optical neural networks, demonstrated in a foundry-built 2x2 matrix multiplier.

desk verdict A real but modest experimental step: the 2x2 MDM matrix multiplier works, but the WDM-compatibility claim is not actually demonstrated as a 2xN channel grid. read the letter →

arxiv 2411.16140 v1 pith:6OCLF4CF submitted 2024-11-25 physics.optics

Zhetao Jia , Hector Rubio , Lilian Neim , Jagang Park , Stefan Preble , Boubacar Kanté This is my paper · ORCID
classification physics.optics
keywords mode-divisionmultiplexingwavelength-divisionopticalneuralnetworkmicro-ringresonatormatrixmultiplicationphotoniccomputingthermo-opticaltunermultimodewaveguide
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 proposes that mode-division multiplexing (MDM) — carrying separate data streams in different spatial patterns of light within the same waveguide — can be layered on top of wavelength-division multiplexing (WDM) to enlarge the input dimension of micro-ring resonator optical neural networks. It identifies the three components needed for this: a multimode beam splitter, a thermo-optical tuner that couples a specific high-order mode (TE01) from a multimode bus into a micro-ring, and a low-loss multimode waveguide bend. The central demonstration is a 2-by-2 matrix multiplication system fabricated in a foundry that works with two spatial modes alone, and also with two spatial modes combined with two wavelengths, performing element-wise multiplication and row addition. A sympathetic reader would care because WDM-based optical neural networks are hitting a practical limit on channel count set by wavelength spacing and crosstalk, and spatial modes provide an independent axis that can multiply the matrix size without adding more laser sources.

What carries the argument

The central mechanism is the mode-selective thermo-optical micro-ring tuner: a micro-ring resonator with an asymmetric coupling region that phase-matches the TE01 mode of the multimode bus waveguide to the fundamental mode of the ring, and a symmetric coupling region that returns the ring's fundamental mode to a single-mode output. An integrated photoconductive sensor on the ring lets its resonance be aligned by electrical current measurement rather than by probing the optical signal. The supporting components are a multimode beam splitter that distributes each mode's intensity equally across the row of weight rings, and a multimode waveguide bend that guides the high-order mode with low scattering loss.

What would settle it

A concrete test would be to fabricate a larger MDM matrix multiplier (say 4x4) with two modes using the same foundry process and measure the crosstalk: excite only the TE01 input and read the current in the TE00 row's photodetector. If the ratio of desired signal to leaked current is no better than the average signal-to-noise ratio of 5 already observed in the 2x2 device, the claim that MDM increases the usable input dimension fails to scale.

Watch

Extended reading notes

Core claim

The paper claims that a micro-ring resonator can be made mode-selective: an asymmetric coupling region matches the effective index of the TE01 mode of a 1.1-µm multimode waveguide to the fundamental mode of the ring, so only the high-order mode is dropped into the ring, while the fundamental TE00 mode passes with negligible coupling. By pairing this with a symmetric coupler on the output side and a heater for thermal tuning, a single device acts as a weight for one mode of the input vector. The fabricated system uses six such rings (two input tuners and four weight rings) to implement a 2x2 matrix-vector multiplier, and the measured photodetector currents for input vectors (1,0), (0,1), and (1,1) show the expected element-wise multiplication and summation. The same chip is then used with two lasers at 1472.6 nm and 1550.7 nm, demonstrating that the MDM computation is compatible with WDM: the output current adds the contributions from both wavelength-mode combinations. The paper also reports that the photoconductive response of the ring can be used to calibrate each tuner electrically, without reading the optical output, which is necessary for scaling to larger weight banks.

Load-bearing premise

The load-bearing premise is that the TE00 and TE01 spatial modes remain sufficiently orthogonal as they propagate through the multimode waveguide, beam splitter, bends, and tuners; if intermodal crosstalk grows with matrix size, the added channels will not be usable for computation.

Editorial extensions

If this is right

  • With M supported modes and N wavelengths, a single bus waveguide can carry an input vector of dimension M times N, multiplying the matrix size available to a micro-ring weight bank without adding laser lines.
  • The approach is compatible with existing WDM micro-ring neural networks, because the mode channels are added on the same multimode bus and the weight rings convert each mode back to the fundamental mode for detection.
  • The photoconductive calibration mechanism allows each mode-selective tuner to be set electrically, removing the need to measure the optical output during alignment as the number of weights grows.
  • The demonstrated element-wise multiplication and row addition for both MDM-only and MDM-WDM inputs confirms that the added mode dimension is usable for matrix-vector multiplication on a foundry platform.
  • The paper's own assessment is that intermodal crosstalk can be reduced by optimizing the multimode waveguide width and the coupling length of the multimode directional coupler, which would improve the measured average signal-to-noise ratio of 5.

Reading between the lines

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

  • If mode-selective rings can be extended to higher-order modes (TE02 and beyond) with acceptable crosstalk, the same bus waveguide could carry three or more spatial channels, multiplying the effective channel count without denser wavelength packing.
  • The MDM-WDM combination may be a natural fit for convolutional neural network preprocessing, where the mode count M can match the kernel channels and the wavelength dimension handles positions, potentially reducing the number of laser sources required.
  • Because the measured average signal-to-noise ratio is about 5 (roughly 7 dB), a useful benchmark would be to track how the crosstalk accumulates as the matrix size grows; if it scales superlinearly, the usable mode count will be limited even after optimization.
  • The photoconductive response of the tuners could plausibly be used as an in-situ training signal for the weight rings, allowing the optical matrix to be adapted without external photodetectors, though the paper does not demonstrate this.
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 / 4 minor

Summary. The paper proposes using mode-division multiplexing (MDM) in silicon micro-ring resonator weight banks as a complement to wavelength-division multiplexing (WDM). It reports a foundry-fabricated thermo-optic tuner that couples the TE01 mode to the TE00 mode, with photoconductive monitoring, and a 2-by-2 matrix-vector multiplication unit. Measured photocurrents for input combinations (1,0), (0,1), and (1,1) show element-wise multiplication and row addition. A two-laser experiment demonstrates current summation when laser 1 (coupled to TE00) and laser 2 (coupled to TE01) are both on. The authors claim this verifies MDM-WDM computing and increases the input vector size for matrix multiplication.

Significance. If fully substantiated, the work would be a useful advance: it adds a new degree of freedom to the established micro-ring weight-bank architecture, uses foundry-compatible components, and includes a photoconductive calibration scheme that avoids optical feedback. The experimental data are direct observations with no fitted model, which is a strength. The main limitation is that the WDM experiment does not yet demonstrate the claimed dimensionality increase: each wavelength is tied to a distinct mode, so the test is equivalent to a two-input MDM system with two independent laser sources. The crosstalk SNR of 5 also needs to be connected to computational accuracy. The paper is therefore a promising proof-of-concept, but its central quantitative claims currently outrun the evidence.

major comments (3)
  1. [System-level (Fig. 4)] The MDM-WDM demonstration uses lambda1 on the TE00 mode and lambda2 on the TE01 mode, which is a diagonal pairing in the wavelength-mode grid. No measurement shows two wavelengths on the same mode or two modes at the same wavelength being weighted independently. Consequently, the data do not support the abstract's statement that the system works for both MDM and MDM-WDM computing in the sense of WDM multiplying the MDM dimension; they support a two-input MDM system whose two inputs are generated by separate lasers. Please either measure at least one off-diagonal channel (for example, lambda2 on TE00 with its own weight) or revise the claim to state that MDM and WDM sources were operated together sequentially rather than that WDM increases the input vector dimension.
  2. [System-level, Fig. 3(c)] The only quantitative quality metric is an average signal-to-noise ratio of 5, but no definition of SNR, no expected-versus-measured comparison for the matrix-vector product, and no error bars are given. The text also states that intermodal coupling can be reduced by optimizing the width of the multimode waveguide in future iterations. To support the claim of a system that successfully increases the input vector size, the paper should report the measured photocurrent values against the ideal values for all four matrix elements and quantify the effect of crosstalk on the computed sums.
  3. [Component-level] The paper claims to experimentally demonstrate key components, but the multimode beam splitter and the multimode waveguide bend results are only presented in the supplementary material, with no summary in the main text. Because the system-level SNR of 5 and the crosstalk behavior depend directly on the splitting ratio and bending loss of these components, the main text should provide at least their measured insertion loss and uniformity; otherwise the reader cannot assess the system-level claims from the main text alone.
minor comments (4)
  1. [Abstract and Conclusion] The term multi-dimensional in the title and conclusion overstates the demonstrated system, which is two-dimensional in the mode dimension and only diagonal in wavelength-mode space; consider a more qualified phrasing such as mode- and wavelength-multiplexed.
  2. [Component-level, Fig. 2(e)] The negative photoconductive response observed in Fig. 2(e) is mentioned but not explained; a brief mechanistic description would help the reader interpret the calibration scheme.
  3. [References] Reference [26] is incomplete; it lists only the author and arXiv number with no title and no full author list. Please correct it.
  4. [System-level] The definition of signal-to-noise ratio in Fig. 3(c) should be stated explicitly, for example as the ratio of on-state current to off-state crosstalk current, so that the reader can interpret the reported value of 5.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are experimental demonstrations verified by direct photocurrent measurements, not derivations from fitted inputs or self-citation chains.

full rationale

The paper's central claims are experimental: it fabricates components, measures photodetector currents, and demonstrates element-wise multiplication and addition for MDM and MDM-WDM inputs. There is no fitted model whose parameters are later renamed as predictions, and no quantity is defined in terms of the result it is supposed to establish. The MDM tuner design uses simulated coupling efficiencies, but the experimental extinction ratio and resonance shift are measured directly and compared with simulation rather than derived from the intended conclusion. The photoconductive calibration mechanism cites prior work [28-31] for a known technique, but that citation is not load-bearing for the claimed increase in input vector size; it supports an implementation detail. The WDM-compatibility evidence gap noted in the skeptic's reading, namely that Fig. 4 uses one wavelength per mode rather than a full 2x2 wavelength-mode grid, is a correctness or completeness concern about whether the demonstration matches the abstract's claim, not a circularity. Likewise, the acknowledgment that intermodal crosstalk can be reduced in future iterations is a limitation statement, not a circular step. No self-definitional reduction, fitted-input prediction, or author-imported uniqueness theorem appears. The paper's results stand or fall on measurement quality and experimental design, which are outside the circularity definition.

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

The paper is an experimental demonstration, so no free parameters are fitted to data. The central claim rests on standard photonic domain assumptions about mode orthogonality, the photoconductive calibration mechanism, and foundry fidelity. No new physical entities are postulated.

assumptions (3)
  • domain assumption TE00 and TE01 spatial modes in the multimode waveguide remain sufficiently orthogonal through the splitter, bends, and ring couplers so that each mode can be independently modulated.
    The entire MDM scheme relies on mode orthogonality for independent channels. The paper measures an average crosstalk SNR of ~5 (Fig. 3c), so orthogonality is only approximate. Invoked in the Design Architecture and System-level sections.
  • domain assumption The photoconductive effect in the doped micro-ring yields a detectable conductance change proportional to optical coupling, enabling calibration without measuring optical output.
    Used to claim built-in calibration. The paper reports a negative conductance change, opposite to prior work, attributed to thermal effects from higher doping without direct verification. Section on photoconductive effect.
  • domain assumption Foundry fabrication on AIM Photonics MPW reproduces the simulated coupling coefficients closely enough for the demonstrated function.
    The system-level discussion notes mismatch between simulation and fabricated device causing unwanted intermodal coupling; the central demonstration assumes the device functions as designed. System-level section.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-dimensional optical neural network." pith.science (2026). https://pith.science/paper/6OCLF4CF

@misc{pith2026241116140,
  author       = {Pith},
  title        = {Pith review of: Multi-dimensional optical neural network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6OCLF4CF}},
  note         = {Machine review of arXiv:2411.16140}
}
read the original abstract

The development of deep neural networks is witnessing fast growth in network size, which requires novel hardware computing platforms with large bandwidth and low energy consumption. Optical computing has been a potential candidate for next-generation computing systems. Specifically, wavelength-division multiplexing (WDM) has been widely adopted in optical neural network architecture to increase the computation bandwidth. Although existing WDM neural networks architectures have shown promise, they face challenges in the integration of light sources and further increase of the computing bandwidth. To overcome these issues, we introduce a mode-division multiplexing (MDM) strategy, offering a new degree of freedom in optical computing within the micro-ring resonator platform. We propose a MDM approach for small-scale networks and a multi-dimensional architecture for large-scale applications, supplementing WDM with MDM to enhance channel capacity for computations. In this work, we design and experimentally demonstrate key components for the MDM computing system, i.e., a multimode beam splitter, a thermo-optical tuner for the high-order mode, and a multimode waveguide bend. We further show a 2-by-2 matrix multiplexing system fabricated in a foundry that works for both MDM and MDM-WDM computing, which confirms that our approach successfully increases the input vector size for computing and ensures compatibility with existing WDM networks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 27 canonical work pages

  1. [1]

    Huang, S

    C. Huang, S. Fujisawa, T. F. de Lima, A. N. Tait, E. C. Blow, Y. Tian, S. Bilodeau, A. Jha, F. Yaman, H. T. Peng, H. G. Batshon, B. J. Shastri, Y. Inada, T. Wang, and P. R. Prucnal, Nature Electronics4, 837 (2021)

    work page 2021

  2. [2]

    X. Xu, M. Tan, B. Corcoran, J. Wu, A. Boes, T. G. Nguyen, S. T. Chu, B. E. Little, D. G. Hicks, R. Moran- dotti, A.Mitchell, andD.J.Moss,Nature 589,44(2021)

    work page 2021

  3. [3]

    Feldmann, N

    J. Feldmann, N. Youngblood, M. Karpov, H. Gehring, X. Li, M. Stappers, M. Le Gallo, X. Fu, A. Lukashchuk, A. S. Raja, J. Liu, C. D. Wright, A. Sebastian, T. J. Kippenberg, W. H. Pernice, and H. Bhaskaran, Nature 589, 52 (2021)

    work page 2021

  4. [4]

    van Niekerk, A

    M. van Niekerk, A. Rizzo, H. Rubio, G. Leake, D. Cole- man, C. Tison, M. Fanto, K. Bergman, and S. Pre- ble, Neuromorphic Computing and Engineering2, 034012 (2022)

    work page 2022

  5. [5]

    L. G. Wright, T. Onodera, M. M. Stein, T. Wang, D. T. Schachter, Z. Hu, and P. L. McMahon, Nature601, 549 (2022)

    2022

  6. [6]

    Zhang, M

    H.Zhou, J.Dong, J.Cheng, W.Dong, C.Huang, Y.Shen, Q. Zhang, M. Gu, C. Qian, H. Chen, Z. Ruan, and X. Zhang, Light: Science and Applications 11 (2022), 10.1038/s41377-022-00717-8

    doi:10.1038/s41377-022-00717-8 2022

  7. [7]

    S. Pai, Z. Sun, T. W. Hughes, T. Park, B. Bartlett, I. A. Williamson, M. Minkov, M. Milanizadeh, N. Abebe, F. Morichetti, A. Melloni, S. Fan, O. Solgaard, and D. A. Miller, Science (New York, N.Y.)380, 398 (2023)

    work page 2023

  8. [8]

    Rizzo, A

    A. Rizzo, A. Novick, V. Gopal, B. Y. Kim, X. Ji, S. Daudlin, Y. Okawachi, Q. Cheng, M. Lipson, A. L. 6 Gaeta, and K. Bergman, Nature Photonics (2023), 10.1038/s41566-023-01244-7

    doi:10.1038/s41566-023-01244-7 2023

Show all 31 references
  1. [9]

    S. Xu, J. Wang, S. Yi, and W. Zou, Nature Communica- tions 13 (2022), 10.1038/s41467-022-35723-2

    2022 doi

  2. [10]

    M. J. Filipovich, Z. Guo, M. Al-Qadasi, B. A. Marquez, H. D. Morison, V. J. Sorger, P. R. Prucnal, S. Shekhar, and B. J. Shastri, Optica9, 1323 (2022)

    2022

  3. [11]

    T. Wang, S. Y. Ma, L. G. Wright, T. Onodera, B. C. Richard, and P. L. McMahon, Nature Communications 13 (2022), 10.1038/s41467-021-27774-8

    2022 doi

  4. [12]

    E. Luan, S. Yu, M. Salmani, M. S. Nezami, B. J. Shastri, L. Chrostowski, and A. Eshaghi, Scientific Reports13, 1 (2023)

    2023

  5. [13]

    T. Wang, M. M. Sohoni, L. G. Wright, M. M. Stein, S. Y. Ma, T. Onodera, M. G. Anderson, and P. L. McMahon, Nature Photonics 17, 408 (2023)

    2023

  6. [14]

    Y. Shen, N. C. Harris, S. Skirlo, M. Prabhu, T. Baehr- Jones, M. Hochberg, X. Sun, S. Zhao, H. Larochelle, D. Englund, and M. Soljacic, Nature Photonics11, 441 (2017)

    2017

  7. [15]

    A. N. Tait, T. F. De Lima, E. Zhou, A. X. Wu, M. A. Nahmias, B. J. Shastri, and P. R. Prucnal, Scientific Reports 7, 1 (2017)

    2017

  8. [16]

    A. N. Tait, Physical Review Applied17, 1 (2022)

    2022

  9. [17]

    A. L. Gaeta, M. Lipson, and T. J. Kippenberg, Nature Photonics 13, 158 (2019)

    2019

  10. [18]

    Chang, S

    L. Chang, S. Liu, and J. E. Bowers, Nature Photonics 16, 95 (2022)

    2022

  11. [19]

    H. Shu, L. Chang, Y. Tao, B. Shen, W. Xie, M. Jin, A. Netherton, Z. Tao, X. Zhang, R. Chen, B. Bai, J. Qin, S. Yu, X. Wang, and J. E. Bowers, Nature 605, 457 (2022)

    2022

  12. [20]

    B. Bai, Q. Yang, H. Shu, L. Chang, F. Yang, B. Shen, Z. Tao, J. Wang, S. Xu, W. Xie, W. Zou, W. Hu, J. E. Bowers, and X. Wang, Nature Communications14 (2023), 10.1038/s41467-022-35506-9

    2023 doi

  13. [21]

    Padmaraju, X

    K. Padmaraju, X. Zhu, L. Chen, M. Lipson, and K. Bergman, IEEE Photonics Technology Letters 26, 1478 (2014)

    2014

  14. [22]

    Deng, IEEE Signal Processing Magazine 29, 141 (2012)

    L. Deng, IEEE Signal Processing Magazine 29, 141 (2012)

    2012

  15. [23]

    L. W. Luo, N. Ophir, C. P. Chen, L. H. Gabrielli, C. B. Poitras, K. Bergmen, and M. Lipson, Nature Communi- cations 5, 1 (2014)

    2014

  16. [24]

    X. Wu, C. Huang, K. Xu, C. Shu, and H. K. Tsang, Journal of Lightwave Technology35, 3223 (2017)

    2017

  17. [25]

    K. Y. Yang, C. Shirpurkar, A. D. White, J. Zang, L. Chang, F. Ashtiani, M. A. Guidry, D. M. Lukin, S. V. Pericherla, J. Yang, H. Kwon, J. Lu, G. H. Ahn, K. Van Gasse, Y. Jin, S. P. Yu, T. C. Briles, J. R. Stone, D. R. Carlson, H. Song, K. Zou, H. Zhou, K. Pang, H. Hao, L. Tras...

    2022

  18. [26]

    E. K. Gordon, . , 1 (2018), arXiv:1810.06652

    2018 arXiv

  19. [27]

    N. M. Fahrenkopf, C. McDonough, G. L. Leake, Z. Su, E. Timurdogan, and D. D. Coolbaugh, IEEE Journal of Selected Topics in Quantum Electronics25, 1 (2019)

    2019

  20. [28]

    Jayatilleka, K

    H. Jayatilleka, K. Murray, M. Á. Guillén-Torres, M. Caverley, R. Hu, N. A. F. Jaeger, L. Chrostowski, and S. Shekhar, Optics Express23, 25084 (2015)

    2015

  21. [29]

    A. N. Tait, Diss. Princeton University (2018)

    2018

  22. [30]

    Huang, S

    C. Huang, S. Bilodeau, T. Ferreira De Lima, A. N. Tait, P. Y. Ma, E. C. Blow, A. Jha, H. T. Peng, B. J. Shastri, and P. R. Prucnal, APL Photonics 5 (2020), 10.1063/1.5144121

    2020 doi

  23. [31]

    Huang, A

    C. Huang, A. Jha, T. F. De Lima, A. N. Tait, A. N. Tait, B. J. Shastri, B. J. Shastri, and P. R. Prucnal, IEEE Journal of Selected Topics in Quantum Electronics 27 (2021), 10.1109/JSTQE.2020.2998073

    2021

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.