Intelligent Mode-Locked Single-Cavity Dual-Comb Laser Utilizing Time-Stretch Dispersive Fourier Transform Spectroscopy

Pan Guo; Sha Wang; Yongjie Pu; Yuan Gao; Zhenhua Cong; Zhigang Zhao

REVIEW 4 major objections 5 minor 13 references

Intelligent Mode-Locked Single-Cavity Dual-Comb Laser Utilizing Time-Stretch Dispersive Fourier Transform Spectroscopy

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 An evolutionary algorithm guided by a CNN-Transformer trained entirely on simulated time-stretch dispersive Fourier transform spectra can find and restore single-cavity dual-comb mode locking in under two seconds, with stable operation…

desk verdict The system is genuinely new as an end-to-end combination, but the simulated-to-experimental classifier transfer is the load-bearing claim and it isn't quantified. read the letter →

arxiv 2501.03550 v3 pith:BICGMYWE submitted 2025-01-07 physics.optics

Pan Guo , Yuan Gao , Yongjie Pu , Zhigang Zhao , Zhenhua Cong , Sha Wang This is my paper · ORCID

classification physics.optics PACS 42.55.Wd42.60.By42.65.Re

keywords dual-comblasersingle-cavitydualcombmodelockingtime-stretchdispersiveFouriertransformCNN-Transformerevolutionaryalgorithmmotorizedpolarizationcontrollersolitonstability

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 tries to establish that stable single-cavity dual-comb mode locking can be found and kept automatically, without active phase locking or a pre-collected experimental dataset. It proposes a two-part evaluation criterion: an intensity-count fitness function for pulse-train quality and a CNN-Transformer classifier, trained solely on simulated time-stretch dispersive Fourier transform spectra, that tells dual-comb states apart from second-order harmonic and other states. An evolutionary algorithm controlling the six paddles of two motorized polarization controllers uses that criterion to search the cavity, and a real-time library of saved paddle angles restores known states in under two seconds. The paper also claims a random-collision maintenance algorithm holds the dual-comb state for over six hours by monitoring the standard deviation of weak soliton peaks.

What carries the argument

The carrying mechanism is the pairing of a two-part evaluation criterion with an evolutionary search loop. The fitness function, based on a dual-region intensity count with thresholds $T_1$ and $T_2$, scores how close a time-stretch spectrum is to the ideal pulse count; the CNN-Transformer, whose convolutional layers catch local spikes and whose attention layers catch global structure, classifies the spectrum as continuous wave, single-wavelength, harmonic, or dual-comb. The training data are generated entirely by numerical simulation of the Ginzburg-Landau equation, and the search loop is driven by an evolutionary algorithm whose six genes are the voltages of the six MPC paddles. The stability loop is driven by random collision: small Gaussian-distributed rotations of the paddles, accepted or reversed according to the standard deviation of weak-soliton peak heights, and a state library records the fitness and paddle angles of each successful state, with state filtering deleting entries that no longer reproduce.

What would settle it

Take a large set of experimental time-stretch DFT waveforms whose true states are identified by an independent method, run the same CNN-Transformer on them, and compare; if real-spectrum accuracy is not close to 100 percent, the simulation-only training assumption is falsified and the claimed guarantee on classification collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that single-cavity dual-comb mode locking can be treated as a searchable, restorable operating point rather than a fragile accident. The authors show that a two-part evaluation criterion—an intensity-count fitness function plus a CNN-Transformer classifier trained entirely on numerically simulated Ginzburg-Landau time-stretch spectra—can guide an evolutionary algorithm controlling the six paddles of two motorized polarization controllers to find dual-comb states and to tell them apart from second-order harmonic states. They further report that a real-time library of saved paddle angles restores any recorded dual-comb state in less than two seconds, and that a random-collision algorithm using the standard deviation of weak soliton peaks maintains the state for over six hours.

Load-bearing premise

The load-bearing premise is that spectra simulated from the Ginzburg-Landau equation mimic the real laser output closely enough that the CNN-Transformer, trained only on those simulations, classifies genuine experimental spectra with near-perfect accuracy; if that transfer fails, the evolutionary search cannot reliably distinguish dual-comb from harmonic mode locking.

Editorial extensions

If this is right

Dual-comb operation can be restored by recalling saved polarization angles, so a library of previously found states makes re-entry to dual-comb mode locking a sub-two-second routine rather than a fresh search.
Because the classifier is trained on simulated spectra, the same intelligent controller can in principle be deployed on a different fiber laser without collecting a new experimental data set for that specific cavity.
Changing the ideal pulse count in the fitness function points the same search machinery at other target states, so higher-order harmonic mode locking can also be sought and classified from the spectral information.
The random-collision maintenance loop, which reverses ineffective paddle rotations and retries, provides an automatic mechanism to keep a drift-prone dual-comb state alive for hours rather than requiring operator intervention.

Reading between the lines

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

The reported sub-2-second restoration is a library recall, so the practical speed advantage is bounded by how often the library degrades; the paper's stated 15-minute library build time is the relevant cost when a cavity loses its saved states.
The claimed 'close to 100%' accuracy on experimental spectra is asserted without a published experimental confusion matrix; the simulation-to-real transfer is the single most testable point, and a mismatch would make harmonic states masquerade as dual-comb states during search.
The standard-deviation stability metric and its 0.08 threshold are calibrated to this cavity; applying the maintenance algorithm to another dual-comb laser would likely require re-measuring that threshold, but the reversal-and-retry logic should transfer.

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.

Desk Editor's Note

read the letter

First, what is actually new. The paper puts together three things I haven't seen in one system: a CNN-Transformer trained exclusively on simulated Ginzburg-Landau time-stretch DFT spectra, an evolutionary algorithm controlling two motorized polarization controllers, and a real-time library that stores MPC angles and lets you recall a mode-locked state in under 2 seconds. On top of that there is a random-collision routine intended to keep the dual-comb state stable for hours. The experimental evidence for a true dual-comb state is solid: two repetition rates around 12 MHz with an 800 Hz difference, SNR above 50 dB, autocorrelation traces, and spectral peaks at 1535.8 and 1561.3 nm. That part I buy.

The soft spot is the classifier transfer. The text admits the fitness function cannot distinguish second-order harmonic from dual-comb, so the CNN-Transformer is doing essential work. Yet the only statement about experimental performance is "the accuracy is close to 100%" with no confusion matrix on real traces, no number of test traces, no labeling protocol, no simulation parameters. If the classifier confuses harmonic and dual-comb states on the real laser, the evolutionary search collapses to the wrong operating point. This is a real omission, but it is fixable: they just need to hand-label a few hundred experimental DFT traces and report the confusion matrix.

The headline times have a similar but smaller problem. The 2-second recall and 15-minute library build are stated without statistics or run counts. I'd like to see how many repeats, and what the spread is. The 6-hour stability plot is nice but doesn't show the random-collision controller actually triggering and recovering from a perturbation; if it never kicks in, the plot only shows that the laser was stable, not that the algorithm works. That's a minor gap because the algorithm is described clearly.

There is no data or code release, which makes it harder to verify or build on. That's a practical limitation, not a flaw in the physics. In all, the central claim is plausible, the combination is new, and the missing pieces are standard supplementary material. I would not cite this yet, but it deserves serious review.

Referee Report

4 major / 5 minor

Summary. This manuscript reports an intelligent single-cavity dual-comb fiber laser system in which an evolutionary algorithm controls two motorized polarization controllers based on a two-part evaluation criterion: a fitness function (Eq. (1)) and a CNN-Transformer classifier. The classifier is trained exclusively on time-stretch dispersive Fourier transform spectra generated by numerical simulation of the Ginzburg-Landau equation. The authors report that their system builds a library of mode-locked states within 15 minutes, recovers a stored dual-comb state in under 2 seconds on average, and maintains dual-comb operation for over 6 hours using a random-collision stability-maintenance algorithm based on the standard deviation of weak soliton peaks. The paper's central claims are reliable intelligent search for dual-comb states without experimental training data and record-fast library recall.

Significance. If the transfer of the simulated-trained classifier to experimental spectra is quantitatively established, the work would be a useful step toward practical, self-starting single-cavity dual-comb lasers. The idea of training exclusively on synthetic data, if shown to work, would reduce the burden of data collection and improve generalization. The paper also clearly demonstrates the library-based recall concept and a stability-maintenance heuristic. However, the current manuscript does not yet provide the experimental confusion matrix needed to verify the central classifier assumption, and the simulation parameters are absent, so the significance of the result is conditional on that evidence.

major comments (4)

[Section 2, Fig. 2(d)] The experimental classification accuracy of the CNN-Transformer is stated only as 'close to 100%' with no supporting confusion matrix, no number of experimental test traces, and no description of how ground-truth labels were assigned for real laser states. Because the text explicitly notes in Section 2 that the fitness function alone cannot distinguish second-order harmonic from dual-comb mode locking, the CNN-Transformer is the only component separating these two states in the evolutionary search loop. Without a quantitative demonstration of simulation-to-real transfer, the central claim that the EA reliably selects dual-comb states is not supported. The authors should provide an experimental confusion matrix for the four classes, including the harmonic versus dual-comb row, with trace counts and labeling criteria.
[Section 2, CNN-Transformer training dataset] The training data are described as generated by numerical simulations of the Ginzburg-Landau equation, but no simulation parameters (gain, dispersion, nonlinear coefficient, filter profile, noise model, etc.) are reported. The transferability of the near-100% simulated validation accuracy to real spectra depends on the simulation reproducing the experimental waveform statistics. Without these parameters, the work is not reproducible and the claim that the model 'does not rely on specific experimental data' cannot be assessed. Provide the simulation details and, at minimum, overlay representative simulated and experimental time-stretch spectra to demonstrate similarity.
[Section 3, Library recall] The claim of a 'fastest record' of less than 2 seconds for restoring a dual-comb state from the library is not quantified beyond an average. The authors should report the number of recall trials, the distribution of recall times, and the exact procedure used to measure the interval (e.g., whether it includes the time for the oscilloscope to confirm successful mode locking). This is important because the comparison in the text is against the 2.48 s mean time reported in Ref. [10].
[Section 3, Fig. 4] The 6-hour stability demonstration is presented as a single recorded trajectory, and the random-collision maintenance algorithm is characterized only by a threshold of 0.08 on the standard deviation of weak soliton peaks. The manuscript does not report how many times the algorithm was triggered during the 6 hours, the success rate of the collision-restoration steps, or the distribution of restoration durations. Without these statistics, the 'robust restoration' claim is not quantitatively supported.

minor comments (5)

[Eq. (1)] Eq. (1) appears in the submitted text in a garbled form; please ensure the equation is typeset correctly and that the symbols I_pulse, C_real, C_ideal, α, β, and I_noise are clearly defined.
[Abstract and Introduction] The sentence 'As the generation of single-cavity dual combs fundamentally rely on...' contains grammatical errors; the manuscript should be proofread by a native English speaker.
[Section 2, after Eq. (1)] The value Cideal = 10 for the dual-comb state is not explained; clarify how this count relates to the time-stretch DFT period and the presence of two solitons.
[Fig. 3(b)] The inset captions in Fig. 3(b) should state the scale and units of the autocorrelation traces; the text refers to pulse durations of 885 fs and 783 fs but does not discuss the deconvolution factor used.
[Section 2, last paragraph] The statement that 'higher-order harmonic mode locking above the third order is not observed due to the limitation of pump power' is an important limitation and should be restated in the conclusion for completeness.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CNN is trained on external simulations, the fitness function is a control objective, and library recall is stored-setting replay rather than prediction.

full rationale

The paper's derivation chain is not circular. The CNN-Transformer is trained exclusively on simulated Ginzburg-Landau time-stretch DFT spectra, which are external to the experimental optimization loop, and the network is then applied to experimental spectra; this is a transfer-learning claim rather than a self-referential one. The fitness function in Eq. (1) is an explicit control objective, not a fitted prediction, and the paper itself acknowledges that it cannot distinguish second-order harmonic from dual-comb states, so the CNN is added as an independent classifier. The mode-locked library stores measured MPC angles and fitness values for previously found states; recalling saved angles in under 2 seconds is a replay of stored experimental settings, not a prediction derived from those settings. The stability-maintenance algorithm uses a real-time standard-deviation criterion on weak soliton peaks to trigger reversible random collisions, which is feedback control rather than an equation containing its own conclusion. The self-citations (Refs. 6 and 13) support prior intelligent-control techniques and a Lyot-filter design choice, but they are not load-bearing for the central demonstration. The unquantified near-100 percent accuracy of the simulated-trained classifier on experimental spectra is a reproducibility and validity concern, not a circularity. Overall, the central claims are grounded in external simulation data and independent experimental characterization. Therefore, the score is 0.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several user-chosen control parameters and on the unstated representativeness of the Ginzburg-Landau simulations and CNN-Transformer transfer. No new physical entities are introduced; the library and random collision are algorithmic constructs, not invented physical entities.

free parameters (8)

Fitness function weights alpha and beta = alpha = 0.2, beta = 0.01
Chosen by hand to balance pulse intensity, pulse-count deviation, and noise terms in Eq. (1); affects which states the EA selects.
Intensity thresholds T1 and T2 for pulse and noise counting = not stated (T1 > T2)
Used in the dual-region count scheme to count pulses and noise; values are not reported, so the pulse-count term is incompletely specified.
Ideal pulse count Cideal = 5 for single-wavelength, 10 for dual-comb
Target pulse count defining the desired mode-locked state in Eq. (1); user-defined objective for the search.
Random collision step size range = Gaussian, -3 to +3 degrees
Chosen to balance search speed and accuracy in the stability maintenance loop.
Random collision iteration threshold = 20
Number of random collision attempts before the EA returns to single-wavelength search.
Weak-soliton standard deviation threshold = 0.08
Threshold for triggering the random collision stability algorithm; chosen without stated statistical justification.
CNN-Transformer hyperparameters = not specified
Network architecture details such as layers, heads, and hidden dimensions are not given, though validation accuracy is reported.
Ginzburg-Landau simulation parameters = not specified
The synthetic training data depend on simulation parameters that are not disclosed, so the representativeness of the data cannot be assessed.

assumptions (5)

domain assumption The Ginzburg-Landau equation captures the relevant dynamics of the single-cavity dual-comb fiber laser, including mode-locked states and their time-stretch DFT spectra.
The training datasets for the CNN-Transformer are completely provided by numerical simulations of the Ginzburg-Landau equation; if the simulations omit experimental effects, the classifier may not transfer.
domain assumption A classifier trained only on simulated spectra will classify experimental spectra with near-perfect accuracy.
The paper states experimental accuracy is close to 100 percent but provides no experimental confusion matrix or transfer-validation protocol.
domain assumption The standard deviation of weak soliton peak intensities is a reliable real-time indicator of dual-comb stability.
The random collision maintenance algorithm is triggered solely by this standard deviation exceeding 0.08; its sensitivity and false-alarm rate are not characterized.
domain assumption Motorized polarization controller angles map reproducibly to mode-locked states over time.
Library restoration in under 2 seconds assumes that replaying saved paddle angles re-creates the same dual-comb state despite environmental drift.
standard math Time-stretch dispersive Fourier transform maps the optical spectrum to a temporal waveform through the dispersion compensation module.
The DFT readout requires that the DCM dispersion of -654 ps/nm stretches the spectrum linearly onto the oscilloscope time trace.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Intelligent Mode-Locked Single-Cavity Dual-Comb Laser Utilizing Time-Stretch Dispersive Fourier Transform Spectroscopy." pith.science (2026). https://pith.science/paper/BICGMYWE

@misc{pith2026250103550,
  author       = {Pith},
  title        = {Pith review of: Intelligent Mode-Locked Single-Cavity Dual-Comb Laser Utilizing Time-Stretch Dispersive Fourier Transform Spectroscopy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BICGMYWE}},
  note         = {Machine review of arXiv:2501.03550}
}

read the original abstract

As dual combs play a significant role in numerous high-precision measurements, their efficient generation has been widely researched. Although the single-cavity dual-comb generation can avoid the complex active stabilization methods, achieving and maintaining stable dual-comb mode locking within a single cavity remains a critical challenge. To break through this constraint, a two-part evaluation criterion containing a fitness function and a CNN-Transformer network is employed to achieve mode locking and classify the dual-comb mode-locked state. Simulated time-stretch dispersive Fourier transform (DFT) spectra are used as datasets, which simplifies the optimization process and does not rely on specific experimental data. A developed evolutionary algorithm (EA) for paddle-based motorized polarization controllers (MPCs) is proposed, enabling the intelligent attainment of dual-comb mode-locked states. A real-time library stores fitness and MPC angles, facilitating mode-locked state achievement within 2 seconds. Finally, long term running of dual-comb mode locking is ensured by a random collision algorithm utilizing an evaluation criterion of weak soliton peaks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

[10]

G. Pu, R. Liu, C. Luo, et al., J. Lightwave Technol. 41(2), 593-598 (2022)

work page 2022
[1]

Caldwell, L.C

E.D. Caldwell, L.C. Sinclair, N.R. Newbury, et al., Nature 610, 667-673 (2022)

work page 2022
[2]

Fellinger, G

J. Fellinger, G. Winkler, P.E.C. Aldia, et al., Opt. Lett. 46(15), 3677-3680 (2021)

work page 2021
[3]

Cho, T.H

B. Cho, T.H. Yoon, and M. Cho, Phys. Rev. A 97(3), 033831 (2018)

work page 2018
[4]

Millot, S

G. Millot, S. Pitois, M. Yan, et al., Nat. Photonics, 10(1), 27-30 (2016)

work page 2016
[5]

Z. Wang, Q. Nie, H. Sun, et al., Light Sci. Appl. 13(1), 11 (2024)

work page 2024
[6]

P. Guo, M. Fan, H. Li, et al., Laser Photonics Rev. 18(8), 2301209 (2024)

work page 2024
[7]

Girardot, F

J. Girardot, F. Billard, A. Coillet, et al., IEEE J. Sel. Top. Quant. 26(5), 1-8 (2020)

work page 2020

Show all 13 references

[8]

Andral, R.S

U. Andral, R.S. Fodil, F. Amrani, et al., Optica 2(4), 275-278 (2015)

2015
[9]

Kokhanovskiy, E

A. Kokhanovskiy, E. Kuprikov, K. Serebrennikov, et al., Nanophotonics 13(16), 2891-2901 (2024)

2024
[11]

Q. Yan, Y. Tian, T. Zhang, et al., Photonics 11(1), 47 (2024)

2024
[12]

G. Pu, L. Yi, L. Zhang, et al., Optica 6(3), 362-369 (2019)

2019
[13]

Y. Pu, P. Guo, Y. Gao, et al., Opt. Lett. 49(8), 2149-2152 (2024)

2024

Pith tools

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

[1] [10]

G. Pu, R. Liu, C. Luo, et al., J. Lightwave Technol. 41(2), 593-598 (2022)

work page 2022

[2] [1]

Caldwell, L.C

E.D. Caldwell, L.C. Sinclair, N.R. Newbury, et al., Nature 610, 667-673 (2022)

work page 2022

[3] [2]

Fellinger, G

J. Fellinger, G. Winkler, P.E.C. Aldia, et al., Opt. Lett. 46(15), 3677-3680 (2021)

work page 2021

[4] [3]

Cho, T.H

B. Cho, T.H. Yoon, and M. Cho, Phys. Rev. A 97(3), 033831 (2018)

work page 2018

[5] [4]

Millot, S

G. Millot, S. Pitois, M. Yan, et al., Nat. Photonics, 10(1), 27-30 (2016)

work page 2016

[6] [5]

Z. Wang, Q. Nie, H. Sun, et al., Light Sci. Appl. 13(1), 11 (2024)

work page 2024

[7] [6]

P. Guo, M. Fan, H. Li, et al., Laser Photonics Rev. 18(8), 2301209 (2024)

work page 2024

[8] [7]

Girardot, F

J. Girardot, F. Billard, A. Coillet, et al., IEEE J. Sel. Top. Quant. 26(5), 1-8 (2020)

work page 2020