{"id":"f8e1e331-cd9c-4fa7-8905-23da166e35ca","arxiv_id":"2509.03976","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Estimating fiber parameters from a handful of modal coefficients instead of raw speckle images cuts ML training time from 9h45m to 40s and reduces training data by 5x in a six-mode fiber sensor.","lead":"This paper trains machine learning on modal coefficients extracted from speckle patterns in a few-mode fiber, instead of on raw pixel images, to estimate fiber curvature, bending, torsion, and position. The authors report a 5x reduction in training data and an 800x reduction in machine-learning training time, while also demonstrating dual-parameter and tactile sensing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central premise—that retrieved modal coefficients are the true physical mode weights—is validated only by self-consistency (0.985 speckle correlation), not by ground-truth comparison; if the SPGD-derived H_z is biased, the physics-informed advantage and the 5×/800× efficiency claims rest on an u","rationale":"The paper's central claim is that using modal coefficients instead of pixel specklegrams yields comparable/better accuracy with 5× less data and 800× less training time. For this to be a physics-informed advance—rather than just another feature extractor—the coefficients must correspond to the true fiber eigenmodes. The only experimental check in the paper is speckle reconstruction correlation (§2.2, Fig. 2), which is necessary but not sufficient: an incorrect coefficient vector can often reproduce the measured intensity because the intensity is a projection that loses phase information and the inverse problem is ill-conditioned under noise. The preparatory SPGD step that builds H_z (Eq. 2) is itself a non-convex optimization (Ref. 47) and its solution may depend on initialization; if H_z is biased, the subsequent BFGS refinement will faithfully recover coefficients in that wrong basis, and the 0.985 correlation only confirms internal consistency. This is the weakest link because every downstream result—continuous modal evolution, multi-mode importance, dual-parameter low crosstalk—inherits the error. The reader's weakest_assumption identifies exactly this. A secondary issue is that the paper uses only amplitude information despite claiming 2N−1 coefficients, and the 800× training-time claim excludes the 0.42 s per-measurement decomposition cost; these are real and should be corrected, but they do not strike at the conceptual foundation as directly as the missing ground-truth validation. The proposed test—controlled mode-selective excitation with known coefficients, plus an independent interferometric decomposition—would settle the question. If the coefficients are accurate, the physics-informed story holds and the efficiency claims stand; if not, the paper should be reframed as empirical feature learning and the generalization claims revised. Hence CONDITIONAL, agreeing with the reader.","tokens_in":9929,"tokens_out":5881,"duration_ms":57744,"concrete_test":"In the same six-mode fiber setup, use a spatial light modulator or a mode-selective photonic lantern to excite the fiber with known modal amplitudes/phases (e.g., 50 random superpositions with independently measured settings). Record the output speckle, apply the proposed anti-noise decomposition, and compute per-coefficient RMS errors (amplitude and phase) relative to the known excitation, across SNRs and mode counts. Additionally, run an independent interferometric mode decomposition (e.g., S² imaging or a reference SPGD with many restarts) on the same experimental specklegrams and compare coefficient vectors. If the per-coefficient errors are larger than the sensing resolution (0.033 m⁻¹ etc.) or disagree with the independent method by more than the noise floor, the 0.985 correlation is insufficient evidence of physical accuracy.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Equations (2)–(4) define the decomposition: an eigenmode matrix H_z is built from SPGD-retrieved coefficients in a preparatory stage, then per-speckle coefficients are obtained by inversion and BFGS refinement. The only experimental validation reported is the correlation between measured and reconstructed speckle intensities (0.985, Fig. 2a). This is a self-consistency check: it shows the extracted coefficients can reproduce the measured intensity, not that they match the true physical superposition. Since the intensity is insensitive to certain ambiguities (global phase, and combinations of mode phases/amplitudes) and the preparatory SPGD step itself can converge to local minima, H_z may be systematically biased. If so, the 'modal fields' used for training are not the true LP-mode fields, and the claimed benefits—continuity, predictability, diverse modal responses, low crosstalk—are not consequences of physics but of an arbitrary, possibly overfitted feature compression. The reported R² values (0.9987 curvature, etc.) then do not demonstrate a physics-informed advantage; they only show that a LightGBM can fit a particular feature map. The paper's SI simulations (Section 2) apparently use known coefficients, but no experimental ground-truth validation (e.g., via mode-selective excitation or an independent interferometric technique) is presented. Without that, the central claim that 'harnessing modal fields' causes the improvements is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a machine-learning pipeline for speckle-based fiber metrology in which pixel-level specklegrams are replaced by 2N−1 modal coefficients (amplitudes and relative phases) recovered by a two-stage anti-noise mode-decomposition method. An SPGD-based preparatory stage builds an eigenmode matrix Hz, and each specklegram is then decomposed by matrix inversion followed by BFGS refinement. The resulting coefficients are used to train a LightGBM regressor for estimating fiber curvature, bending position, bending angle, and torsion angle, and for 2D tactile-pattern reconstruction. The authors report high R² values (0.9987 for curvature, up to 0.9999 for position), a 5× reduction in training data, an 800× reduction in training time, and a 100× reduction in inference time relative to a CNN trained directly on specklegrams.","tokens_in":10317,"tokens_out":3629,"duration_ms":36950,"significance":"If the claims held as stated, the work would offer a practical and low-cost route to multi-dimensional fiber sensing by compressing high-dimensional speckle data into a small set of physically motivated features. The demonstration of tactile pattern reconstruction and simultaneous dual-parameter estimation is interesting, and the authors provide a reasonably detailed experimental description. However, the central premise—that the retrieved coefficients are true physical LP-mode weights—is not validated against any ground truth. The performance comparisons with the CNN baseline are incomplete and partly misleading (the reported inference and training times exclude the mode-decomposition cost). The paper does not ship code, data, or machine-checked proofs, so reproducibility rests on the experimental description alone. The practical idea is plausible, but the current evidence is insufficient to support the strong physics-informed claims.","major_comments":[{"comment":"The decomposition accuracy is validated only by the correlation between measured and reconstructed speckle intensities (0.985). This is a self-consistency check: it shows the coefficients reproduce the measured intensity, not that they equal the true physical mode weights. Since Hz is calibrated from SPGD in a noisy environment, any SPGD bias is inherited by the fast decomposition. The paper needs an independent validation of the modal coefficients, e.g., using mode-selective excitation, an interferometric measurement, or synthetic specklegrams with known coefficients under matched noise statistics. Without this, the claim that the features are physical modal fields—and the associated advantages of continuity, predictability, and low crosstalk—is not established.","section":"§2.2, Eqs. (2)–(4) and Fig. 2"},{"comment":"The inference and training-time comparisons are misleading. The 0.015 ms inference time for the modal-weight scheme excludes the 0.42 s per-speckle mode decomposition; including that cost, end-to-end inference is about 0.42 s, which is far slower than the 1.4 ms CNN inference. Likewise, the '40 s training time' excludes the mode-decomposition preprocessing; the text later says the total time is 'around 20 minutes', but 4,200 samples at 0.42 s/sample is 29.4 min, not 20 min. The 800×/100× speed-up claims need to be recomputed consistently, and the decomposition time should be included or explicitly separated in the comparisons.","section":"Table 2 and §2.3"},{"comment":"The manuscript states 'we utilize only the amplitude information, which is the modal weight ρ2, for sensing.' This conflicts with the surrounding text and figures, which use multiple modal weights (Fig. 4a, Fig. 7, Fig. 8a) and emphasize diverse responses of different modes. If only ρ2 is used, the multi-mode diversity argument collapses; if all modal weights are used, the sentence is an error. This must be corrected and the exact feature set used in each experiment must be stated unambiguously.","section":"§2.3, text near Fig. 4"},{"comment":"The CNN baseline's accuracy is never reported. The claim that 'our approach based on modal fields achieves comparable estimation accuracy with only 4,200 samples' requires the same metrics (R², RMSE, error distributions) for the CNN trained on the same dataset, same splits, and same platform. Comparing against literature values (Table 1) is not a controlled comparison because datasets, fiber types, and tasks differ. Without this baseline, the 5× data-reduction claim is unsubstantiated.","section":"§2.3, Table 1 and CNN comparison"}],"minor_comments":[{"comment":"The equations are poorly rendered and hard to parse. The definitions of X, the eigenmode matrix H, and the sign convention in Eq. (4) should be rewritten clearly; in particular, the expression for X is garbled and the role of Hz versus H is not explicitly defined.","section":"§2.2, Eqs. (2)–(4)"},{"comment":"The abbreviations MEA, REC, DIS, COR are used in the caption but not defined in the main text. Please define them or remove the unexplained acronyms.","section":"Fig. 2 caption"},{"comment":"The abstract states 'training time ... reduced by 800 times, from 9 h 45 min to 40 s', while §2.3 says the total time (including mode decomposition) is 'around 20 minutes, reduced by 30 times'. These two statements should be reconciled, and the definition of 'training time' should be spelled out consistently.","section":"Abstract and §2.3"},{"comment":"The column headers mix 'Resolution' with standard deviations of estimation errors. For a fair comparison, specify how resolution is defined in each cited work and whether the reported values are standard deviations, mean absolute errors, or quantization steps.","section":"Table 1"},{"comment":"The symbol 'R²' is sometimes written as 'R2' or 'R^2'. Also, no error bars or repeated-measurements statistics are given for the R² and RMSE values; please report the number of independent trials and variability.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is legitimate: the decomposition is validated only by self-consistency, so the physics-informed advantage is not proven. The timing comparisons are also unbalanced. I recommend revision rather than acceptance, with emphasis on (i) independent validation of mode coefficients, (ii) fair end-to-end timing, and (iii) a proper CNN baseline. The missing baseline accuracy is especially important because the paper's central practical claim is a controlled comparison against that baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid applied fiber-sensing paper with a real if incremental idea — use mode decomposition to reduce speckle data to 2N-1 coefficients and train a LightGBM on those instead of raw specklegrams. If the comparisons hold up, that cuts training data ~5x and training time ~800x, which matters for rapid deployment. I think it deserves a serious referee, but the authors need to fix three things before the strongest claims are fair.\n\nWhat is actually new: the specific pipeline is not present in the cited speckle-sensor papers I know — they train directly on pixels. The experiments cover four parameters plus tactile reconstruction, and the reported R² values are strong. The anti-noise mode decomposition is a reasonable engineering contribution; the 0.42 s per speckle is slower than IMS but apparently much more stable under noise. The authors deserve credit for demonstrating a wide set of sensing tasks with one compact feature set.\n\nSoft spots, in order of severity. First, the inference-time accounting is misleading. The 0.015 ms is just the LightGBM forward pass; the per-measurement mode decomposition takes 0.42 s, so real per-measurement latency is orders of magnitude worse than the CNN baseline. They do acknowledge total training+decomposition time (20 min), but Table 2 and the '100x inference' claim in the abstract are cherry-picked. Second, the 5x training-data claim lacks the CNN accuracy number; 'comparable accuracy' is asserted but not shown, and the reader is left to trust the comparison. Third, the stress-test concern about validation is real: the decomposition is validated only by speckle reconstruction (0.985 correlation), not by ground truth. The SI simulations use known coefficients, so the algorithm is tested in silico, but in the experiments the 'modal features' could be a smooth compressed representation of the speckle rather than the true LP modes. That doesn't necessarily sink the practical result, but it undercuts the strong physics-informed story. Minor: the text says 'we utilize only the amplitude information, which is the modal weight ρ2,' yet all the figures use many modal weights; I assume a typo, but it should be cleaned up.\n\nBottom line: the paper is worth engaging with. The core demonstration is plausible and the experimental set is fairly rich. It needs a careful referee to push on the inference accounting, the CNN comparison, and the ground-truth validation. I would send it to peer review with a request for meaningful revision, not desk reject.","headline":"The core idea — train on modal coefficients instead of raw speckle — is plausible and the 5x/800x training claims are striking, but the paper overstates the physics grounding and underreports the mode decomposition cost in inference.","tokens_in":10751,"tokens_out":3173,"would_cite":true,"duration_ms":30633,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that compressing a speckle image into 2N-1 fiber-mode coefficients makes speckle-based metrology train 800 times faster without sacrificing accuracy.","keywords":["speckle metrology","few-mode fiber","mode decomposition","modal fields","machine learning","fiber-optic sensing","tactile sensing","multi-dimensional metrology"],"falsifier":"Measure the same fiber states with a phase-sensitive reference technique that reports the true modal amplitudes and phases, and compare those against the SPGD-BFGS outputs. If the two disagree while the retrieved coefficients still reconstruct the recorded speckle, the claimed accuracy rests on a self-consistent but physically wrong representation, and a model trained on it should fail on unseen optical states.","tokens_in":9870,"feed_emoji":"📏","tokens_out":14197,"duration_ms":116877,"temperature":0.7,"pith_summary":"The paper's central claim is that the main bottleneck in speckle-based optical metrology — large training sets and long training times — comes from treating the speckle pattern as raw pixels, and that compressing each specklegram into the modal field coefficients of a few-mode fiber removes that bottleneck. It proposes a two-stage anti-noise fast mode decomposition: an SPGD calibration stage builds a mode-to-speckle matrix for the noisy environment, followed by fast matrix inversion and BFGS refinement that recovers the amplitudes and relative phases of the fiber's LP modes. Feeding these modal coefficients to a LightGBM regressor, rather than 256x256 pixel specklegrams to a CNN, yields high-accuracy estimation of fiber curvature, bending position, bending angle, and torsion, and enables 2D tactile pattern reconstruction. On the reported experiments the approach needs about five times fewer training samples and 800 times less training time (9 h 45 min to 40 s). If the decomposition is physically faithful, this makes speckle sensing substantially cheaper to calibrate and deploy, and opens the same sensor to simultaneous multi-parameter measurement.","feed_headline":"Fiber modes cut speckle-sensor training from 9h45m to 40s","feed_subtitle":"Replacing pixel specklegrams with 2N-1 modal coefficients cuts training data 5x and sharpens resolution.","key_machinery":"The central object is the modal-coefficient vector of length 2N-1: the normalized amplitudes and relative phases of the N linearly polarized (LP) modes making up the field in the few-mode fiber. The key mechanism is the two-stage anti-noise fast mode decomposition. In a preparatory stage, many specklegrams are decomposed with stochastic parallel gradient descent (SPGD) to construct an eigenmode matrix H_z that accounts for the actual noisy environment. In the running stage, each new specklegram is converted to coefficients by the fast inverse matrix solution X=(H_z)^-1 I, and the initial values are refined with the BFGS optimizer. This reduces a 256x256 pixel intensity pattern to at most 2N-","core_discovery":"On its own terms, the discovery is that a few-mode fiber's speckle pattern is better represented for machine learning by the underlying 2N-1 modal coefficients (amplitudes and relative phases of its LP eigenmodes) than by raw pixels. These coefficients are retrieved reliably under noise by a two-stage decomposition: an SPGD calibration builds an eigenmode matrix for the noisy environment, then fast matrix inversion plus BFGS refinement solves each specklegram. Trained on these coefficients, a LightGBM model estimates curvature, bending position, bending angle, and torsion with R² values from 0.9933 to 0.9999, reconstructs 2D tactile patterns, and cuts training data about 5x and training time","pith_inferences":["A direct cross-validation against phase-resolved reference mode decomposition would separate the gain from true physical information from the gain from a convenient learned coordinate system; the paper's 0.985 correlation is with the measured speckle, not with ground-truth mode weights.","The current sensor uses only modal amplitudes for regression, so the N-1 relative phases are an untapped channel that could improve accuracy or add sensing axes.","The same decomposition-plus-regression recipe should carry over to other fiber types and to scattering media that admit a low-dimensional modal basis, with decomposition noise setting the practical ceiling.","Because the dual-parameter experiment already shows crosstalk, scaling to three or more simultaneous parameters will require either more modes or a model that explicitly separates modal responses."],"forward_implications":["A speckle-based fiber sensor can be recalibrated for a new environment in about 40 seconds on ordinary computing hardware rather than roughly ten hours.","The same sensor can estimate multiple physical parameters simultaneously — curvature, bending position, bending angle, and torsion — with a mild accuracy trade-off as the parameter count increases.","Because modal fields are continuous, the model can interpolate between discrete calibration points, yielding 3x to 7x better resolutions than classification-style speckle sensors in the reported comparisons.","The expensive SPGD calibration is paid once; afterward, mode decomposition takes about 0.42 s per sample, so the front end can be reused across sensing tasks without retraining.","Two-dimensional tactile imaging works without prior knowledge of the written pattern by classifying the 10x10 pressed positions from modal weights."],"supporting_citations":[{"why":"Supplies the fast inverse-matrix mode decomposition that the new method starts from and makes more noise-robust.","marker":"[51]"},{"why":"Supplies the SPGD algorithm used in the preparatory stage to build the eigenmode matrix in the noisy environment.","marker":"[47]"},{"why":"Supplies the BFGS optimization used to refine each specklegram's modal coefficients after the fast initial solution.","marker":"[58]"},{"why":"Supplies the LightGBM algorithm used to regress physical parameters from the modal coefficients.","marker":"[55]"},{"why":"Provides a speckle-based bending/position sensor baseline and the fiber-bending experimental method whose resolution is improved 7x.","marker":"[27]"},{"why":"Serves as the MMF-CNN curvature classification baseline whose resolution is compared against the curvature sensing result.","marker":"[24]"},{"why":"Serves as the MMF-CNN curvature regression baseline whose resolution and range are compared with the proposed approach.","marker":"[25]"}],"fun_headline_variants":["Speckle metrology: physics-informed ML trains 800x faster","Physics-informed speckle metrology: 5x less data, 800x faster","Modal fields from fiber speckle: 800x faster machine learning","Multi-dimensional metrology from fiber speckle, 800x faster training"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the routine that pulls the mode amplitudes and phases out of a speckle image is recovering the true physical state of the light, not just a convenient set of numbers that can re-create the measured speckle; if the calibration pattern library is skewed by noise, every downstream estimate inherits that skew, and matching the measured speckle does not prove the mode weights are right.","fun_headline_variants_meta":{"raw":{"variants":["Speckle metrology: physics-informed ML trains 800x faster","Physics-informed speckle metrology: 5x less data, 800x faster","Modal fields from fiber speckle: 800x faster machine learning","Multi-dimensional metrology from fiber speckle, 800x faster training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000808,"raw_usage":{"total_tokens":3394,"prompt_tokens":767,"completion_tokens":2627,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":2546}},"tokens_in":511,"tokens_out":2627,"duration_ms":17520,"temperature":1.0,"reasoning_tokens":2546,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:29:06.853831+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the same fiber states with a phase-sensitive reference technique that reports the true modal amplitudes and phases, and compare those against the SPGD-BFGS outputs. If the two disagree while the retrieved coefficients still reconstruct the recorded speckle, the claimed accuracy rests on a self-consistent but physically wrong representation, and a model trained on it should fail on unseen optical states.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the fast inverse-matrix mode decomposition that the new method starts from and makes more noise-robust."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SPGD algorithm used in the preparatory stage to build the eigenmode matrix in the noisy environment."},{"cited_title":"Nocedal, S","cited_arxiv_id":null,"evidence_quote":"Supplies the BFGS optimization used to refine each specklegram's modal coefficients after the fast initial solution."},{"cited_title":"Cheng, X","cited_arxiv_id":null,"evidence_quote":"Supplies the LightGBM algorithm used to regress physical parameters from the modal coefficients."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides a speckle-based bending/position sensor baseline and the fiber-bending experimental method whose resolution is improved 7x."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the MMF-CNN curvature classification baseline whose resolution is compared against the curvature sensing result."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the MMF-CNN curvature regression baseline whose resolution and range are compared with the proposed approach."}],"review_version":1}