{"id":"3461da4e-839e-4dc6-87bc-2b0e2e52c410","arxiv_id":"2501.03550","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"An intelligent controller using simulated-spectrum-trained CNN-Transformer classification and an evolutionary algorithm can find and restore single-cavity dual-comb mode locking in under two seconds.","lead":"This paper reports an automated control system for a single-cavity dual-comb laser that uses a neural network trained on simulated spectra to recognize mode-locked states, and a search algorithm to adjust polarization controllers, restoring a desired state in under two seconds. A generalist might read it because stable dual-comb sources matter for precision spectroscopy and ranging, and this work is a step toward hands-free operation of such lasers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Experimental transfer of the simulated-data-trained CNN-Transformer is unquantified; the central intelligent search collapses if the near-100% accuracy on real DFT spectra is not reproduced.","rationale":"The reader's weakest_assumption matches mine exactly: the load-bearing premise is that the CNN-Transformer trained only on simulated data generalizes to real time-stretch DFT spectra with near-perfect accuracy. The paper's strongest claim depends on that premise, yet the only evidence is a single sentence with no supporting numbers, experimental confusion matrix, or protocol. This is not an internal inconsistency but a missing empirical validation of a critical component. A secondary concern is that the random-collision stability algorithm is never explicitly shown to trigger: Fig. 4(b) records standard deviations for 6 h, but the text does not state that the threshold of 0.08 was ever exceeded or that a collision successfully restored the state. That is also worth checking, but it is secondary because even the stability algorithm relies on the same weak-soliton evaluation and would not rescue a failed classifier. These gaps are addressable with additional data and analysis rather than being fundamental flaws, so the reader's CONDITIONAL verdict remains appropriate; I recommend no change.","tokens_in":5112,"tokens_out":3988,"duration_ms":36584,"concrete_test":"Run the trained CNN-Transformer on no fewer than 100 independently labeled experimental time-stretch DFT traces, covering CW, single-wavelength, harmonic, and dual-comb states, with labels assigned by optical spectrum (e.g., two peaks at ~1536 and ~1561 nm), RF beat notes (~12.0 MHz with ~800 Hz separation), and autocorrelation traces. Report the full confusion matrix and per-class accuracy. If the experimental accuracy on the harmonic/dual-comb pair is not close to the claimed ~100%, the simulation-to-experiment transfer is unvalidated and the central search claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The mode-locking search relies entirely on the CNN-Transformer to distinguish dual-comb from second-order harmonic states, because Eq. (1)'s fitness function explicitly cannot separate them: 'the pulse trains of the second-order harmonic and dual-comb mode locking are very similar, this makes it difficult to efficiently distinguish with the fitness function.' The network is trained 'exclusively' on simulated Ginzburg-Landau time-stretch DFT spectra, and the only statement about experimental performance is 'the accuracy is close to 100%' (discussion of Fig. 2(d)). No experimental confusion matrix, number of test traces, labeling protocol, or simulation parameters are given. If the classifier confuses harmonic and dual-comb states on real laser output, the EA will converge to the wrong operating point, and the claimed '<2 s' library recall and 6-h stability are not guaranteed by the presented evidence. This is the most load-bearing assumption because every downstream claim depends on the classifier's correctness on experimental spectra.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":5343,"tokens_out":4838,"duration_ms":42406,"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":[{"comment":"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":"Section 2, Fig. 2(d)"},{"comment":"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":"Section 2, CNN-Transformer training dataset"},{"comment":"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":"Section 3, Library recall"},{"comment":"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.","section":"Section 3, Fig. 4"}],"minor_comments":[{"comment":"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.","section":"Eq. (1)"},{"comment":"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":"Abstract and Introduction"},{"comment":"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.","section":"Section 2, after Eq. (1)"},{"comment":"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":"Fig. 3(b)"},{"comment":"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.","section":"Section 2, last paragraph"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope, but the missing experimental confusion matrix and simulation parameters are essential before publication. I would not recommend acceptance without those additions. The authors should also temper the 'fastest record' claim unless more statistics are provided, and should ensure the comparison with Ref. [10] is fair and clearly stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nThe 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.\n\nThere 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.","headline":"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.","tokens_in":5866,"tokens_out":2492,"would_cite":false,"duration_ms":23264,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["42.55.Wd","42.60.By","42.65.Re"],"model":"deepseek-v4-flash","headline":"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…","keywords":["dual-comb laser","single-cavity dual comb","mode locking","time-stretch dispersive Fourier transform","CNN-Transformer","evolutionary algorithm","motorized polarization controller","soliton stability"],"falsifier":"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.","tokens_in":4925,"feed_emoji":"⏱️","tokens_out":7574,"duration_ms":64888,"temperature":0.7,"pith_summary":"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.","feed_headline":"Self-tuning laser finds dual-comb mode locking in 2 seconds","feed_subtitle":"A simulated-data-trained classifier restores the state in under 2 seconds and holds it stable for 6 hours.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the baseline of a prior intelligent single-cavity dual-comb search with 2.48-second mean recall time, which the paper's sub-2-second claim is positioned against.","marker":"[10]"},{"why":"Reports automatic dual-wavelength mode locking via a two-stage genetic algorithm, the alternative approach this work extends and compares against.","marker":"[11]"},{"why":"Provides the dual-region count scheme that the paper adopts as the preliminary pulse-train fitness criterion.","marker":"[12]"},{"why":"Establishes that a Lyot filter permits single-soliton mode locking at high pump power, supporting the cavity design that avoids multi-soliton artifacts.","marker":"[13]"}],"fun_headline_variants":["Self-tuning dual-comb laser locks in 2 seconds","Simulation-trained AI steers laser to dual-comb lock","Two-second dual-comb lock via evolutionary search","Dual-comb laser self-restores, stable for 6 hours","Single-cavity laser: 2s to lock dual-comb mode"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Self-tuning dual-comb laser locks in 2 seconds","Simulation-trained AI steers laser to dual-comb lock","Two-second dual-comb lock via evolutionary search","Dual-comb laser self-restores, stable for 6 hours","Single-cavity laser: 2s to lock dual-comb mode"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000638,"raw_usage":{"total_tokens":2914,"prompt_tokens":894,"completion_tokens":2020,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":1933}},"tokens_in":510,"tokens_out":2020,"duration_ms":15563,"temperature":1.0,"reasoning_tokens":1933,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:53:41.941799+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the baseline of a prior intelligent single-cavity dual-comb search with 2.48-second mean recall time, which the paper's sub-2-second claim is positioned against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reports automatic dual-wavelength mode locking via a two-stage genetic algorithm, the alternative approach this work extends and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the dual-region count scheme that the paper adopts as the preliminary pulse-train fitness criterion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that a Lyot filter permits single-soliton mode locking at high pump power, supporting the cavity design that avoids multi-soliton artifacts."}],"review_version":1}