{"id":"f24951b5-cb9f-4ac4-9fc5-6e04095a0e3e","arxiv_id":"2411.14410","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A neural network that takes the Wigner function as its input can control the spectrum and shape of supercontinuum light, enabling 12 fs few-cycle pulses and on-demand high-order solitons in experiments.","lead":"This paper trains a neural network that reads a two-dimensional map of a light pulse's time-frequency structure (the Wigner function) and uses it to predict and control the spectrum of supercontinuum light. The authors demonstrate on-demand shaping of laser pulses, including 12-femtosecond pulses and high-order solitons, directly from a nonlinear fiber without an external compressor.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim hinges on an unvalidated simulation-to-experiment link: the GNLSE training targets are generated without specifying the fiber parameters or comparing simulated against measured spectra for this setup.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern: the simulation-to-experiment transfer is not validated. I agree that this is the single most important issue. The entire P-CNN pipeline is trained on GNLSE-generated spectral targets; if those targets are not representative of the actual EDFA-2/HNF output, then the claimed on-demand control, the 12 fs pulse, and the high-order soliton dynamics are not causally tied to the physics-embedded network. The experimental demonstrations are genuine and include real-time feedback, which is a form of independent support, but they do not isolate the CNN's contribution from the optimization loop. Other concerns, such as the 'sub-three optical cycle' wording, the approximate form of Eq. (8), and the estimated 120 fs² FROG GDD, are secondary and would not change the overall conditional verdict. Thus the reader's CONDITIONAL recommendation remains appropriate, pending the requested simulator validation and data/code release.","tokens_in":17212,"tokens_out":8415,"duration_ms":92526,"concrete_test":"Provide the actual HNF parameters (βk, γ, R(t), noise model) and the calibrated shaper transfer function, then select 20 phase masks from the LHS training distribution and compare GNLSE-predicted SC spectra with OSA-measured spectra under identical experimental conditions, reporting the normalized Euclidean distance or correlation per mask. If the median correlation is substantially below the claimed R≈0.97 regression on GNLSE targets, the training surrogate does not match the physical system and the central claim is undermined.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core demonstrations—DW engineering, 12 fs compression, and N=3/4 soliton sequences—require that the 7000-sample GNLSE training set (Methods-2) is a faithful surrogate for the physical EDFA-2/HNF system. This is not established. Eq. (1) is written generically, but the simulation parameters for this setup are never given: the dispersion coefficients βk, the nonlinear coefficient γ, the Raman response R(t), the spontaneous-Raman noise ΓR(z,t), and the exact pulse-shaper transfer function are all absent. Methods-1 gives fiber lengths and amplifier stages but not the material/dispersion parameters that determine SC dynamics. The shaper response is justified only by a reference to previous work (Ref. 51) rather than by an in-situ calibration. Because SC generation is extremely sensitive to these parameters and to phase noise, any systematic model-experiment mismatch means the CNN's inverse solutions are not solutions of the real system. The in-loop gradient-descent optimization can partially compensate for such mismatch, but then the paper's claim that the P-CNN itself enables the observed control is not cleanly supported. No comparison between a simulated SC spectrum and a measured OSA spectrum is presented for any training configuration, leaving the surrogate model unvalidated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-embedded convolutional neural network (P-CNN) that maps a pulse-shaper spectral phase to the supercontinuum (SC) output spectrum, using the Wigner chronocyclic representation of the input field as the network input. The network is trained on 7,000 GNLSE-simulated spectra, benchmarked against a feedforward neural network under simulated phase noise, and then used experimentally to tune dispersive-wave center wavelength and bandwidth, to produce 12 fs pulses without an external compressor, and to generate spectro-temporal patterns identified as N=3 and N=4 soliton breathers. The central claim is that embedding the Wigner transform improves convergence and noise robustness and enables on-demand spectro-temporal control of SC.","tokens_in":17469,"tokens_out":7344,"duration_ms":73898,"significance":"If the results hold, the paper offers a practical and transferable idea: replacing a raw spectral-phase vector with a physically motivated spectro-temporal representation as CNN input improves inverse design in a strongly nonlinear system. The simulation benchmark is clearly described, the network architecture is reproducible from the main text, and the experimental demonstrations are ambitious, including continuous dispersive-wave tuning, 12 fs compression, and high-order soliton patterns. The main value lies in showing that physics-embedded input representations can reduce noise sensitivity in ultrafast nonlinear control. However, the strength of the experimental conclusions depends on the fidelity of the GNLSE training surrogate, and that link is not yet established.","major_comments":[{"comment":"The experimental claims rest on a GNLSE-trained surrogate that is never validated for this setup. Eq. (1) is stated generically, and Methods-1 gives fiber lengths and amplifier stages but no values for the HNF dispersion coefficients β_k, the nonlinear coefficient γ, the Raman response R(t), or the spontaneous-Raman noise Γ_R; the pulse-shaper transfer function is justified only by reference to Ref. 51. Since the 7,000-sample training set (Methods-2) is generated from this simulation, any systematic model-experiment mismatch means the P-CNN inverse solutions are not guaranteed to be solutions of the physical system. Please provide the full parameter set and a quantitative comparison between a simulated SC spectrum and a measured OSA spectrum for representative phases, including the noise model used.","section":"Methods-1; Methods-2"},{"comment":"The FNN-versus-P-CNN benchmark is not controlled: the FNN receives a one-dimensional phase vector, while the P-CNN receives a 241×81 Wigner image and uses a convolutional architecture with additional fully connected layers. The two models therefore differ simultaneously in input representation, input dimensionality, and parameter count, so the reported threefold loss reduction and noise-robustness advantage (Fig. 2) cannot be attributed to the physics-embedded input alone. Please include matched-capacity models and/or an FNN fed the same Wigner features, and report parameter counts for all models.","section":"Benchmarking P-CNN performance; Methods-2"},{"comment":"The central experimental demonstrations (Figs. 3(c,d), 4(a), and 5(a)) are single-shot sequences with no error bars, no repeated trials, and no quoted measurement uncertainty. Claims of agreement with targets, such as the 3.7% center-wavelength modulation and the 10–70 nm bandwidth tracking, therefore do not yet establish reproducibility. Please add repeated measurements for at least a subset of points and quantify residual errors, for example as RMS deviation from the target curve.","section":"Spectral Domain Engineering; Temporal Domain Engineering; Spectro-Temporal Engineering"},{"comment":"The effective-z interpretation is not justified. The claim that varying the input spectral phase is equivalent to scanning the physical propagation distance z is not generally true for the GNLSE, and Eq. (8) omits the fiber parameters γ and β_2 from the expression for N², so the constant-N evidence in Fig. 5(d) depends on unstated constants. Please either derive the equivalence explicitly or reframe the demonstration as target-spectrum matching; if solitonic dynamics is claimed, validate it with an independent measure, such as full-field evolution or phase-sensitive characterization.","section":"Spectro-Temporal Engineering: High-Order Solitons; Eq. (8)"}],"minor_comments":[{"comment":"The claim of a 'sub-three optical cycle' 12 fs pulse should be reconciled with the carrier wavelength. At 1070 nm, three optical cycles correspond to about 10.7 fs, so a 12 fs pulse is not sub-three-cycle at that wavelength; please specify the carrier wavelength or revise the statement.","section":"Abstract; Discussion"},{"comment":"'Elicude distance' should be 'Euclidean distance', and the sentence 'in does serve as an ideal benchmarking tool' contains a grammatical error.","section":"Fig. 7(d) caption; main text"},{"comment":"The table in the bottom-left panel is difficult to parse because the columns for GNLSE, FNN, P-FNN, and P-CNN are not clearly aligned with the metric rows; please reformat it.","section":"Fig. 2(c) table"},{"comment":"The notation I_tarG is confusing; please define the reference spectrum as I_tar and the noisy predictions as I_N, or use consistent subscripts throughout.","section":"Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of the journal and the idea is appealing. The deciding issue is the unvalidated simulation-to-experiment link: the central experimental demonstrations depend on a GNLSE surrogate whose parameters and validation are not provided. I do not see grounds for rejection, but the authors should be asked to supply those parameters and a simulated-versus-measured comparison before the claims can be accepted. The uncontrolled FNN/CNN comparison is a secondary but important issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nYou should know two things about this one. The experiments are real and the best part of the paper: 12 fs pulses out of a fiber without a compressor, continuous control of dispersive-wave wavelength/bandwidth, and high-order soliton (N=3,4) breather sequences reconstructed by FROG. That is a solid demonstration of what a shaper plus an optimized neural inverse model can do in a real nonlinear system. The other thing: the 'physics-embedded' novelty is thinner than the title suggests. Embedding the Wigner function as the input to a CNN is a sensible feature choice, not a new learning paradigm. The paper does not oversell too hard, but the reader should not expect physics-informed training.\n\nThe benchmarking of P-CNN vs FNN is suggestive but not controlled. The FNN gets a 1D vector of phase samples while the CNN gets a 241×81 image, so the comparison conflates input representation with model capacity. That said, the P-FNN baseline (Wigner input with a plain FNN) partly addresses this, and the noise-robustness story is plausible.\n\nThe soft spot that matters is the simulation-to-experiment link. The GNLSE training set is the backbone of the inverse model, yet the paper never gives the HNF dispersion coefficients, gamma, Raman response, or noise level, and it never shows a measured SC spectrum against a GNLSE prediction for the same input. The shaper calibration is delegated to Ref. 51, which is fine as prior work, but the fiber model is unvalidated in this setup. The real-time gradient-descent loop can absorb some model mismatch, but it weakens the claim that the P-CNN itself enables the control. This needs to be fixed before I trust the quantitative claims.\n\nThere are also smaller internal inconsistencies: the 12 fs pulse is called 'sub-three optical cycle,' but at the dispersive-wave wavelength near 1070 nm, 12 fs is ~3.4 cycles (unless they mean another wavelength); and P-CNN's Euclidean distance is said to be a 2.5-fold increase over GNLSE, but the numbers give 0.46/0.13 ≈ 3.5. Minor, but sloppy. Experimental data lack error bars and repeated trials.\n\nThe effective propagation-length scan via shaper phase is actually plausible—NLS scaling makes it work—but it deserves an explicit derivation in the main text rather than a hand-wave.\n\nRecommendation: send it to peer review. It is a useful paper with genuine experimental results. Require code/data release, full GNLSE parameters, a simulation-vs-measurement validation plot, and error bars. Then it can be a good contribution. I would bring it to reading group in the meantime.","headline":"Real experiments and a plausible but oversold 'physics-embedded' label; the missing simulator-to-experiment validation is the main thing standing between this and a strong paper.","tokens_in":17984,"tokens_out":4124,"would_cite":false,"duration_ms":36729,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["42.65.Wi","42.65.Re","07.05.Mh"],"model":"deepseek-v4-flash","headline":"Feeding a neural network the Wigner function of the input pulse makes supercontinuum shaping precise enough to produce 12 fs pulses without an external compressor.","keywords":["supercontinuum generation","physics-embedded neural networks","Wigner function","few-cycle pulse shaping","high-order solitons","dispersive wave engineering","ultrafast photonics","spectro-temporal control"],"falsifier":"Take a P-CNN trained on the nominal GNLSE model and test it on a fiber with a measurably shifted zero-dispersion wavelength; if its optimized phase patterns no longer produce the targeted dispersive-wave spectra and 12 fs compression, the claim that the physics-embedded network transfers to real-world supercontinuum control is falsified.","tokens_in":17018,"feed_emoji":"⚡","tokens_out":16309,"duration_ms":129056,"temperature":0.7,"pith_summary":"This paper claims that a convolutional neural network whose input is the Wigner function—the joint time-frequency representation of the pulse—can learn the nonlinear mapping from input spectral phase to supercontinuum output spectrum well enough to control that output on demand. The authors argue that embedding this spectro-temporal correlation into the network input, rather than feeding raw phase or intensity data, cuts prediction loss by about a factor of three and reduces noise-induced output error by more than half compared with a standard feed-forward network. They support this with simulations benchmarked against the generalized nonlinear Schrödinger equation and with experiments in an erbium-fiber system: one P-CNN tunes the center wavelength and bandwidth of a dispersive wave, compresses pulses to 12 fs without an external compressor, and another recreates the breathing spectra of third- and fourth-order solitons. If correct, the result would make broadband spectro-temporal pulse shaping a routine, machine-learnable step in ultrafast laser systems.","feed_headline":"Neural network sculpts supercontinuum into 12-fs pulses","feed_subtitle":"A Wigner-function input lets one neural network tune spectrum and duration of supercontinuum light on demand.","key_machinery":"The central object is the Wigner function $W(t,\\omega)=\\int A(t+t'/2)A^*(t-t'/2)e^{i\\omega t'}dt'$, a joint time-frequency (chronocyclic) representation of the pulse field. The paper places this transform at the front of the network: instead of learning from raw spectral phase or from the output spectrum alone, the CNN sees a two-dimensional image of the pulse's spectro-temporal correlations. That representation suppresses phase variations such as the carrier-envelope offset phase and gives the convolution layers local structure to exploit, which the authors identify as the reason for faster convergence and noise robustness. The supporting machinery is a dataset of 7,000 phase profiles sampled by Latin hypercube sampling, each propagated through the generalized nonlinear Schrödinger equation to produce a target SC spectrum; the Wigner image is cropped and downsampled to $241\\times81$ pixels as CNN input.","core_discovery":"The central claim is that the Wigner distribution, computed from the shaped input pulse, is the right input representation for neural-network control of supercontinuum generation. The paper shows that feeding this spectro-temporal image into a convolutional network (P-CNN) yields a regression factor $R\\approx 0.97$ against GNLSE targets, versus $R\\approx 0.91$ for a three-layer feed-forward network, and that the network stays stable when Gaussian phase noise is added, with typical output error $d_E\\approx 0.46$ instead of $1.17$. Experimentally, the same idea is implemented as two networks—Net-D for dispersive waves and Net-S for solitons—operating on an erbium-fiber laser, 4f pulse shaper, amplifier, and highly nonlinear fiber. With Net-D the network continuously tunes the dispersive-wave center wavelength (3.7% peak-to-peak modulation around 1070 nm) and bandwidth (60–100 nm), and produces 12 fs near-transform-limited infrared pulses—a 70-fold compression of the 850 fs soliton—without any pulse compressor. With Net-S it generates high-order solitons with $N=3$ and $4$, reproducing the spectro-temporal breather dynamics as if scanning fiber length, with measured compression factors around 33–36. The paper attributes this improvement not to the CNN itself but to the physics-embedded input: a plain feed-forward network given Wigner inputs (P-FNN) also outperforms the data-driven FNN.","pith_inferences":["Extension: the paper's training pipeline should carry over to gas-filled hollow-core fibers or chip-scale waveguides, which would test whether the approach scales to µJ–mJ pulses; the authors note this direction but do not demonstrate it.","Extension: the same Wigner-embedding could apply to spatio-temporal correlation functions to design 'space-time' wavepackets with arbitrary transverse profiles, generalizing the $(\\omega,t)$ control shown here to $(k,x)$ degrees of freedom.","Extension: a cleaner test of the mechanism would compare Wigner images against other two-dimensional representations (e.g., spectrograms from FROG) as network inputs; if the advantage persists for any joint time-frequency image, the essential ingredient is the correlation structure, not the Wigner function specifically.","Extension: the noise-robustness result suggests the Wigner input could benefit other sensitive inverse problems in nonlinear optics, such as soliton-molecule control or extreme-event suppression; this is an inference, not a claim of the paper."],"forward_implications":["The same P-CNN architecture can target spectral features (center wavelength, bandwidth) or temporal features (pulse duration) of a supercontinuum, and can switch between dispersive-wave and soliton regimes by retraining on the appropriate spectral band.","Because the trained network returns an output spectrum from a phase pattern almost instantly, real-time closed-loop shaping of few-cycle pulses becomes feasible in this class of fiber systems, replacing minute-scale genetic-algorithm searches.","The 12 fs result implies that octave-spanning spectra need not be paired with external prism or chirped-mirror compressors; the fiber itself can deliver near-transform-limited few-cycle pulses when the input phase is chosen correctly.","High-order soliton breather dynamics, normally observed by physically altering fiber length or energy, can instead be scanned synthetically by reshaping the input pulse, because the P-CNN maintains a nearly constant soliton number while varying dispersion and nonlinear lengths.","Because a plain feed-forward network (P-FNN) also gains from Wigner inputs, the benefit lies in the representation rather than in the convolutional architecture, so other network designs should inherit the same robustness."],"supporting_citations":[{"why":"Supplies the generalized nonlinear Schrödinger equation used to generate the training targets and to benchmark predictions.","marker":"[4]"},{"why":"The feed-forward neural network architecture and training configuration used as the baseline learner.","marker":"[41]"},{"why":"The feed-forward network as nonlinear dynamics integrator for supercontinuum, the comparison ML model for prediction.","marker":"[42]"},{"why":"The chronocyclic time-frequency representation of ultrashort pulses that motivates the Wigner-function input.","marker":"[47]"},{"why":"Latin hypercube sampling, the method used to generate the diverse random phase profiles in the training set.","marker":"[49]"},{"why":"The experimentally validated pulse-shaper transfer function that lets the authors compute Wigner inputs without characterizing every instance.","marker":"[51]"},{"why":"The experimental observation of pulse narrowing and solitons in fibers backing the high-order soliton compression interpretation.","marker":"[55]"}],"fun_headline_variants":["Wigner function teaches neural net to tailor supercontinuum","Neural net uses Wigner image to control supercontinuum shape","Physics-embedded CNN achieves 70-fold compression without compressors","Wigner-based CNN achieves 12-fs pulses despite noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme rests on the assumption that the computer model of the fiber—its dispersion, nonlinearity, and the pulse shaper's response—matches the real experiment closely enough that solutions found in simulation still work when applied to the actual laser.","fun_headline_variants_meta":{"raw":{"variants":["Wigner function teaches neural net to tailor supercontinuum","Neural net uses Wigner image to control supercontinuum shape","Physics-embedded CNN achieves 70-fold compression without compressors","Wigner-based CNN achieves 12-fs pulses despite noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00069,"raw_usage":{"total_tokens":3173,"prompt_tokens":1041,"completion_tokens":2132,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":2061}},"tokens_in":657,"tokens_out":2132,"duration_ms":15598,"temperature":1.0,"reasoning_tokens":2061,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:12:45.471992+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a P-CNN trained on the nominal GNLSE model and test it on a fiber with a measurably shifted zero-dispersion wavelength; if its optimized phase patterns no longer produce the targeted dispersive-wave spectra and 12 fs compression, the claim that the physics-embedded network transfers to real-world supercontinuum control is falsified.","supporting_citations":[{"cited_title":"M., Genty, G","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized nonlinear Schrödinger equation used to generate the training targets and to benchmark predictions."},{"cited_title":"& Finot, C","cited_arxiv_id":null,"evidence_quote":"The feed-forward neural network architecture and training configuration used as the baseline learner."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The feed-forward network as nonlinear dynamics integrator for supercontinuum, the comparison ML model for prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The chronocyclic time-frequency representation of ultrashort pulses that motivates the Wigner-function input."},{"cited_title":"D., Beckman, R","cited_arxiv_id":null,"evidence_quote":"Latin hypercube sampling, the method used to generate the diverse random phase profiles in the training set."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The experimentally validated pulse-shaper transfer function that lets the authors compute Wigner inputs without characterizing every instance."},{"cited_title":"F., Stolen, R","cited_arxiv_id":null,"evidence_quote":"The experimental observation of pulse narrowing and solitons in fibers backing the high-order soliton compression interpretation."}],"review_version":1}