{"id":"67119f8b-c2d9-4fbe-aea4-9c5f8472fa66","arxiv_id":"2504.19564","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A hybrid physics-plus-data neural model of optical fiber predicts channel power rapidly and updates fiber and amplifier parameters from measured power profiles in simulations and a field trial.","lead":"This paper builds a digital twin for optical fiber networks that combines a fast neural network model with the physics equations of light transmission, and shows it can update key hardware parameters from live measurements. If the approach holds up in independent tests, network operators could run fiber links with smaller safety margins and adapt automatically as equipment ages or is replaced.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Field-trial improvement may be in-sample: the eight OCM profile pairs used to update Λ appear to be the same data on which the 1.4 dB gain is reported, and no held-out evaluation is stated.","rationale":"The reader's conditional verdict is appropriate. I identify the same load-bearing weakness as the reader: the field-trial improvement may be computed on the same OCM profile pairs used for parameter updating, so the 1.4 dB improvement could reflect fitting rather than prediction. The full text never states that evaluation loadings were withheld from the eight updating pairs, and the inverse loss in Eqs. (6)-(8) explicitly trains on input-output boundary profiles. Because the paper reports 'improved prediction accuracy' as the key field validation of lifecycle updating, this omission is material. The simulation section does provide independent evidence that the hybrid DeepONet forward model works and that parameter refinement converges to known ground truth, which supports the methodology and justifies not rejecting the paper outright. But that simulation evidence does not resolve the field-trial evaluation question. I agree with the reader's condition: the lifecycle-management claim for real networks should be accepted only after held-out validation of the reported improvement, plus clarification of identifiability if the authors claim uniqueness of the updated Λ. Since the reader already set the verdict to CONDITIONAL for essentially this reason, I recommend no change to the verdict.","tokens_in":18023,"tokens_out":4072,"duration_ms":48276,"concrete_test":"Re-run the inverse-updating procedure on the eight field-trial first-span OCM profile pairs in a leave-one-out fashion: for each k, update Λ and the PEO parameters using the other seven pairs, freeze the updated model, and compute channel-power RMSE and per-channel GSNR error on the held-out pair. Report the mean and maximum held-out improvement relative to the old DT. If the held-out improvement is near zero or negative, the reported 0.8 dB average and 1.4 dB maximum improvement are in-sample fitting rather than predictive updating.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central lifecycle claim rests on the field-trial result in §VI, where the paper says that 'eight pairs of channel powers before and after this span measured by OCM along the regular operations are used as initial and final conditions' for updating Λ, and then reports improved RMSE and up to 1.4 dB GSNR improvement without stating that the evaluation loadings were excluded from those eight pairs. If the full-loading or partial-loading profiles used in Fig. 11 are among the updating pairs, the reported improvement is a fitting result, not a prediction: the update procedure can adjust the per-channel gain profile gn (96 parameters), per-band connector losses, and Raman strength to match the measured boundary profiles closely. The model has ample capacity to fit a handful of output profiles, so matching those profiles does not demonstrate that the twin tracks physical parameter shifts. The simulation study provides independent support for the forward operator and for convergence to known parameters in a controlled setting, but it does not establish that the field-trial accuracy gain is predictive. A leave-one-out evaluation on the eight available pairs would settle whether the improvement persists out-of-sample. The identifiability issue noted by the reader is secondary; the decisive weakness is the apparent absence of a train/test split in the field-trial evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a dynamic-updating digital twin (DT) for optical-network lifecycle management. The fiber channel is modeled by a DeepONet trained in a hybrid data-driven and physics-informed manner: the SRS-ODE residual is used as a physics regularizer, and a small set of labeled power profiles supplies the data term. An inverse updating procedure refines a set of physical parameters Λ = {r, δ_in,C(L), δ_out,C(L), g_n} from measured input/output channel-power profiles. The authors demonstrate forward prediction on the COST 239 network with up to 100x speedup over split-step methods and normalized RMSE around 1e-4 on 1,000 unseen profiles, and they show in simulation that the refinement converges to known ground-truth parameters. A field trial on a deployed C+L-band link after an EDFA replacement reports improved channel-power and QoT prediction accuracy, with up to 1.4 dB GSNR improvement.","tokens_in":18254,"tokens_out":6304,"duration_ms":57431,"significance":"If the central claim is established, the work is a valuable step toward low-margin lifecycle management: it couples a fast, physics-consistent forward operator with an inverse scheme that updates several physical parameters from monitoring data. The paper’s strengths are concrete: the forward operator is tested on 1,000 unseen input profiles with normalized RMSE around 1e-4; the simulation study recovers known ground-truth values of connector loss, Raman strength, and gain profile; and the reported 100x speedup relative to split-step is plausible for a closed-form operator. The field-trial demonstration on a real C+L link is a useful differentiator from purely simulation studies. However, the field-trial evaluation protocol as written does not establish that the reported accuracy improvements are predictive rather than in-sample, and the identifiability of Λ from boundary profiles is not examined beyond the controlled simulation.","major_comments":[{"comment":"The lifecycle-updating claim rests on the field-trial results, but the manuscript does not state that the evaluation loadings were held out from the eight OCM profile pairs used for updating. In §VI, 'Eight pairs of channel powers before and after this span measured by OCM along the regular operations are used as initial and final conditions,' and the subsequent accuracy improvements (power RMSE from 1.1 to 0.12 dB; up to 1.4 dB GSNR improvement) are then reported on full-loading and partial-loading conditions. If any of the Fig. 11 loadings are among those eight pairs, the reported gains are a fitting result: the update procedure has ample capacity (96 per-channel gain values plus per-band connector losses and Raman strength) to match a handful of boundary profiles. Since the abstract and conclusion present this as post-device-replacement prediction, the authors must either (i) explicitly state that the Fig. 11 loadings were excluded from the updating pairs, or (ii) provide a leave-one-out evaluation over the eight pairs showing that the improvement persists out-of-sample.","section":"§VI, Fig. 11"},{"comment":"The inverse updating step assumes that Λ = (r, δ_in,C(L), δ_out,C(L), g_n) is identifiable from boundary power profiles at z=0 and z=zmax. The loss in Eq. (8) is minimized over these parameters plus the network parameters θ, and the simulation in §V.C shows convergence to known ground truth in one scenario. However, no argument or experiment establishes uniqueness: different combinations of per-channel gain g_n and per-band connector losses can compensate for each other at the boundary, and the Raman strength r interacts with the power-dependent SRS term. A concrete test would be to generate synthetic boundary profiles from two different Λ combinations that both drive Eq. (8) to the same low value, and show whether the optimizer converges to the true parameters or to a different combination. Without such a test, the claim that the field-trial parameter traces in Fig. 10 represent physical parameter shifts is not fully supported.","section":"§IV.D, Eqs. (6)–(8)"}],"minor_comments":[{"comment":"The notation for the final-condition loss weight is inconsistent: Eq. (8) and the text define λd3, while the Fig. 3(b) caption uses λd2. Please align.","section":"§IV.D, Fig. 3(b)"},{"comment":"The heading 'Simulations on Large-Sacle Optical Networks' contains a typo; it should read 'Large-Scale'.","section":"§V heading"},{"comment":"In the sentence 'At start, physical parameters Λ are incorrected,' 'incorrected' should be 'incorrect' or 'not yet corrected.'","section":"§IV.D"},{"comment":"The field-trial QoT improvement is reported as a maximum over a small number of CUT channels; it would be helpful to report the per-channel distribution and the number of channels contributing to the 1.4 dB maximum.","section":"§VI, Fig. 11"},{"comment":"The sentence 'the normalized root mean-square-error (RMSE) generally falls in 1x10−4' should read 'falls below 1×10−4' or 'on the order of 1×10−4' for clarity.","section":"§V.B"},{"comment":"The caption lists two '(d)' entries ('Updating error...' and 'Convergence of connector loss'); the second should be '(e)'.","section":"Fig. 8 caption"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the potential absence of a train/test split in the field-trial evaluation; this is fixable by a leave-one-out analysis and should be the focus of the revision. The simulation study is sufficiently strong to justify not rejecting the paper. I see the paper as a good fit for the journal; no novelty concerns beyond the need to clarify the evaluation protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: the simulation content is the real contribution; the field-trial section as written cannot support the headline claim.\n\nWhat is genuinely new: the authors integrate forward prediction and inverse multi-parameter updating in one hybrid DeepONet, and in simulation they update Raman gain strength, per-channel EDFA gain profile, and connector losses simultaneously from boundary power profiles. The forward model is tested on 1,000 unseen loading profiles with normalized RMSE around 1e-4, and the inverse scheme converges to known ground-truth parameters. The 100x speedup over split-step is also credible because the closed-form operator removes the iterative loop.\n\nThe soft spot is the field-trial evaluation. Section VI says eight pairs of OCM channel powers “are used as initial and final conditions” for updating, then reports improved RMSE and up to 1.4 dB GSNR improvement without stating that the evaluation profiles in Fig. 11 were excluded from those eight pairs. With 96 per-channel gain parameters plus connector losses and Raman strength, the model has ample capacity to memorize eight output profiles. So the reported improvement may be fitting error, not prediction error. This is not a minor quibble: the lifecycle-updating claim depends on generalization to new operating conditions. The authors need a leave-one-out evaluation on those eight pairs, or a clear statement that the full/partial-loading profiles used for display were different from the updating pairs.\n\nThe identifiability of Lambda is a real concern but secondary: the simulation converges to known values, which is evidence that the inverse problem is well-posed in at least some regime. More important is the missing train/test separation in the field data. Also minor: no error bars, no code/data release, and the paper doesn't discuss measurement jitter quantitatively despite mentioning it.\n\nIn sum: this is a plausible engineering contribution from a group that has done serious prior work in this area. The simulation results are reproducible in principle and the method is interesting. The field-trial validation needs to be re-done or re-reported before the central claim can be accepted. A good referee would not desk-reject this; they would ask for major revision and specifically for an out-of-sample test. I'd bring it to reading group to discuss train/test discipline in digital-twin papers.","headline":"Solid simulation study of a physics-informed DeepONet for fast fiber-channel modeling, but the field-trial updating result is likely in-sample and does not yet establish predictive lifecycle tracking.","tokens_in":18853,"tokens_out":1891,"would_cite":true,"duration_ms":19153,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["42.79.Sz","42.65.Dr"],"model":"deepseek-v4-flash","headline":"An optical-network digital twin that re-fits its physical parameters from routine power measurements can track lifecycle changes and improve post-replacement performance estimation by up to 1.4 dB.","keywords":["digital twin","lifecycle management","optical networks","physics-informed neural network","DeepONet","stimulated Raman scattering","power evolution operator","quality of transmission"],"falsifier":"Take the same field link after a device replacement, split the OCM power-profile pairs into an update set and a held-out evaluation set, run the inverse updating on the update set only, and recompute the reported average and maximum power/GSNR improvements on the held-out set; if the improvement vanishes, the result is in-sample fitting. Alternatively, search for two different parameter combinations that produce the same boundary power profiles under the SRS-ODE and show that updating converges to the wrong one.","tokens_in":17745,"feed_emoji":"📡","tokens_out":5755,"duration_ms":56030,"temperature":0.7,"pith_summary":"The paper tries to show that a digital twin of an optical fiber network can be kept accurate over the network's whole lifetime by re-estimating a small set of physical parameters from channel-power measurements already available from optical channel monitors. The twin is built on a physics-informed neural operator that learns power evolution along the fiber, with the stimulated-Raman-scattering equations folded into the training loss, so it predicts channel power and quality of transmission quickly while staying physically consistent. On top of that forward model, the paper adds an inverse updating step that refits connector losses, fiber Raman gain strength, and amplifier gain profiles from input and output power profiles at the two ends of a span. The claimed payoff is lifecycle management without manual parameter lookup: accurate prediction at deployment and automatic re-alignment after ageing, repair, or device replacement. In a field-trial C+L-band link, updating the parameters after a device replacement improved per-channel performance estimation by up to 1.4 dB for the channels under test.","feed_headline":"Self-updating twin cuts optical prediction error by 1.4 dB","feed_subtitle":"Hybrid physics-plus-data twin refits Raman gain, connector loss, and amplifier profiles from routine power measurements.","key_machinery":"The central object is the power evolution operator (PEO), a DeepONet composed of a branch net that encodes the launch-power profile and a trunk net that encodes distance, merged as a vector product to output channel power at any distance z. Physics enters through the stimulated-Raman-scattering ODE as a residual loss computed by automatic differentiation, and hybrid training runs in three steps: pure physics guidance, data-driven speedup, then combined refinement. For inverse updating, the parameter set Lambda = (r, delta_in,C(L), delta_out,C(L), g_n) is attached to the model as a physical-parameter layer and fitted by alternating initial-condition loss, final-condition loss, and SRS-ODE loss, using input and output power profiles measured by optical channel monitors at the two ends of a span.","core_discovery":"The central discovery is that one hybrid neural operator can do both jobs a living digital twin needs: solve the forward problem fast and solve the inverse problem cheaply. Trained as a DeepONet with branch and trunk networks, the operator maps a launch-power profile to the full distance-resolved power profile of all C+L-band channels; the SRS-ODE residual is part of the loss, so the outputs obey the coupled Raman equations rather than merely fitting labels. The same architecture, with a physical-parameter layer, can then be updated using only a few measured power profiles at the span input and output: the connector losses for C and L bands, the Raman gain strength, and the amplifier gain profile are recovered by alternating between matching the boundary data and satisfying the SRS-ODE. The paper reports normalized RMSE around $10^{-4}$ for forward predictions on unseen loadings in the COST 239 simulation, up to 100 times speedup over split-step Fourier methods, and, on a field link after replacement, average channel-power RMSE dropping from 1.1 to 0.12 dB with a maximum GSNR improvement of 1.4 dB.","pith_inferences":["A test the paper leaves implicit is blind evaluation: update the twin on one set of OCM pairs after a replacement and score the improvement on a second, later set of pairs, since the reported 1.4 dB gain may otherwise reflect in-sample fitting.","Before deploying the full six-parameter update, one could probe identifiability by using only connector-loss updates with more diverse launch-power profiles, which would separate practical convergence from fundamental parameter ambiguity.","The inverse-updating recipe should transfer to other distributed-constant systems governed by coupled ODEs, such as Raman amplifiers or sensing fibers, wherever boundary power spectra are measurable.","A fully autonomous loop would add a prediction-error trigger, and the simulation's 0.5 dB threshold suggests such closed-loop control is feasible, though the field-trial section does not demonstrate the trigger in operation."],"forward_implications":["A trained twin can predict channel power and GSNR about 100 times faster than split-step numerical integration, with accuracy that degrades only mildly with distance and channel count.","Operators can re-sync the twin to a changed span using routine OCM measurements, without manual re-characterization of amplifiers or connectors.","The same updating procedure simultaneously recovers several parameter types, fiber Raman gain, both end connector losses on C and L bands, and the EDFA gain profile, rather than tuning one parameter at a time.","If the field-trial improvement holds, lifecycle management can shift from conservative static margins to a low-margin mode in which performance predictions track device ageing and repair.","The approach also works under partial loading, with ASE fill channels removed, so updating does not require full-band traffic."],"supporting_citations":[{"why":"Supplies the physics-informed machine learning / PINN methodology whose loss-embedding idea the hybrid training adopts.","marker":"[25]"},{"why":"Supplies the DeepONet architecture and universal operator approximation rationale used as the power evolution operator.","marker":"[33]"},{"why":"Establishes physics-informed DeepONets, the basis for combining operator learning with SRS-ODE regularization.","marker":"[34]"},{"why":"The prior conference paper whose physics-informed parameter refinement on a field-trial C+L-band link this work extends.","marker":"[17]"},{"why":"Provides the closest prior parameter-refinement approach for DT performance prediction, the baseline this work contrasts with.","marker":"[20]"},{"why":"Documents system drift and data-driven parameter optimization in a live production network, motivating dynamic updating.","marker":"[22]"},{"why":"Direct precursor using PINN for power evolution prediction and Raman gain spectrum identification.","marker":"[30]"},{"why":"Supports connector-loss estimation from stimulated-Raman-scattering strength in C+L-band systems.","marker":"[21]"},{"why":"Supplies the GSNR decomposition and transceiver-noise model used for QoT estimation.","marker":"[35]"}],"fun_headline_variants":["Self-updating optical twin: 100x speedup, 1.4 dB gain","Hybrid twin recalibrates itself from few power probes","Physics-informed twin auto-refits via measured power profiles","Network twin cuts prediction error by 1.4 dB, self-updating","100x faster predictions from self-learning optical twin"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the connector losses, Raman gain strength, and amplifier gain profile can be uniquely recovered from just the input and output power profiles of a span, and that the eight field-trial profile pairs used for updating are not also the ones on which the reported accuracy improvement is measured.","fun_headline_variants_meta":{"raw":{"variants":["Self-updating optical twin: 100x speedup, 1.4 dB gain","Hybrid twin recalibrates itself from few power probes","Physics-informed twin auto-refits via measured power profiles","Network twin cuts prediction error by 1.4 dB, self-updating","100x faster predictions from self-learning optical twin"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00135,"raw_usage":{"total_tokens":5515,"prompt_tokens":1013,"completion_tokens":4502,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":4413}},"tokens_in":629,"tokens_out":4502,"duration_ms":33569,"temperature":1.0,"reasoning_tokens":4413,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:49:07.463342+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same field link after a device replacement, split the OCM power-profile pairs into an update set and a held-out evaluation set, run the inverse updating on the update set only, and recompute the reported average and maximum power/GSNR improvements on the held-out set; if the improvement vanishes, the result is in-sample fitting. Alternatively, search for two different parameter combinations that produce the same boundary power profiles under the SRS-ODE and show that updating converges to the wrong one.","supporting_citations":[{"cited_title":"Physics-informed machine learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the physics-informed machine learning / PINN methodology whose loss-embedding idea the hybrid training adopts."},{"cited_title":"Learning the solution operator of parametric partial differential equations with physics-informed deep- onets,","cited_arxiv_id":null,"evidence_quote":"Establishes physics-informed DeepONets, the basis for combining operator learning with SRS-ODE regularization."},{"cited_title":"Physics-informed digital twin with parameter refinement for a field-trial c+ l-band transmission link,","cited_arxiv_id":null,"evidence_quote":"The prior conference paper whose physics-informed parameter refinement on a field-trial C+L-band link this work extends."},{"cited_title":"Machine learning enhancement of a digital twin for wavelength division mul- tiplexing network performance prediction leveraging quality of trans- mission parameter refinement,","cited_arxiv_id":null,"evidence_quote":"Provides the closest prior parameter-refinement approach for DT performance prediction, the baseline this work contrasts with."},{"cited_title":"Improved qot estimations through refined signal power measurements and data-driven parameter optimizations in a disaggregated and partially loaded live production network,","cited_arxiv_id":null,"evidence_quote":"Documents system drift and data-driven parameter optimization in a live production network, motivating dynamic updating."},{"cited_title":"Pinn for power evolution prediction and raman gain spectrum identification in c+ l-band transmission system,","cited_arxiv_id":null,"evidence_quote":"Direct precursor using PINN for power evolution prediction and Raman gain spectrum identification."},{"cited_title":"Improving the accuracy of qot estimation with insertion loss distribution evaluation for c + l band transmission systems,","cited_arxiv_id":null,"evidence_quote":"Supports connector-loss estimation from stimulated-Raman-scattering strength in C+L-band systems."},{"cited_title":"On the impact of launch power optimization and transceiver noise on the performance of ultra-wideband transmission systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the GSNR decomposition and transceiver-noise model used for QoT estimation."}],"review_version":1}