{"id":"ff609eb0-118c-4df6-bda0-135a15e42ec3","arxiv_id":"2507.15376","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A fiber sensor uses a multimode fiber and a genetically-trained optical diffractive network to map physical perturbations directly to output light intensity, with no electronic signal processing.","lead":"This paper demonstrates an all-optical fiber sensing system in which a trained diffractive optical network converts strain and torsion directly into light intensity readings, removing electronic demodulation. If the latency and resolution claims hold, the architecture could enable low-power, nanosecond-response fiber sensors for robotics, structural health monitoring, and dense sensor arrays.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed linear readout from a few training states rests on an untested assumption that the trained optical output is monotonic in the measurand; if speckle decorrelation is non-monotonic, untrained states give wrong intensities.","rationale":"Agree with the reader's identification of the memory-effect/monotonicity assumption as the load-bearing concern. The paper's experimental demonstrations provide credible evidence that a specific trained ODN maps several measurands to intensity with low RMSE on a 30-state test set, so the concern is not that the concept is impossible; it is that the mechanism claimed to guarantee generalization is not established. The loss function in Eq. 2 is purely pointwise, and the genetic algorithm can fit the training states without any smoothness bias. The paper's reference to the memory effect does not by itself ensure monotonic speckle decorrelation; MMF speckle correlations can oscillate for wavelength changes. A direct fine-grained measurement of the trained output would settle whether the between-training-state behavior is actually monotonic and linear. Since the reader's verdict is already CONDITIONAL and this concern is a refinement of the stated weakest assumption, no change in verdict is recommended. The proposed test is non-destructive and uses the existing setup.","tokens_in":16917,"tokens_out":13657,"duration_ms":156119,"concrete_test":"Using the trained SLM from the FBG strain experiment, apply a continuous strain sweep across the full 150 µε calibration range in 1 µε steps and record the ODN output intensity at the designated detector region. Plot intensity versus ground-truth strain and test for strict monotonicity (all successive increments of the same sign) and for deviation from the linear fit exceeding the reported RMSE. If any non-monotonic segment or multivalued region appears, the generalization from four training states to the full range is not valid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central generalization claim (Section 2.2) is that training on only four strain states yields a linear intensity-measurand mapping over the full range, attributed to the 'memory effect' in the MMF (Fig. 2c, Supplementary Note 2). This assumes the speckle decorrelation is deterministic, monotonic, and smooth enough that the ODN output between training points is single-valued and approximately linear. However, the training loss (Eq. 2) constrains the output at only the training states; it imposes no monotonicity or smoothness condition on the response between them. A single-layer phase mask is sufficiently expressive to fit four points with a highly oscillatory mapping, and MMF spectral speckle correlations are not generally monotonic - they can exhibit side lobes due to modal phase differences. The cited memory effect (ref 37) concerns spatial shift invariance, not monotonic spectral decorrelation. The paper does not report a fine-grained measurement of the trained ODN output across the full range, so it cannot rule out local non-monotonic regions where a single intensity value corresponds to multiple measurands. If such regions exist, the claimed 'direct readout' would be ambiguous and the linear regression (Eq. 1) would fail for untrained states.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an all-optical fiber sensing architecture (AOFS-IC) in which the sensing signal from a fiber sensor passes through a multimode-fiber scattering medium and a trained single-layer phase mask (SLM), so that the physical measurand (strain, torsion, bending-related joint angle, etc.) is read directly from detected optical intensity without digital demodulation. The authors report strain regression with RMSEs of 2.7554 µε over 150 µε and 0.0688 mε over 2.5 mε, torsion classification with 100% accuracy on 9 discrete angles, dual-parameter strain/torsion demultiplexing, PD-based dynamic strain measurements up to 150 kHz, and 3-DOF robotic arm joint-angle estimation. The claimed advantages are sub-3 ns demodulation delay and elimination of electronic processing hardware.","tokens_in":1427,"tokens_out":2901,"duration_ms":117966,"significance":"If the claims hold, the architecture is a useful demonstration of in-sensor optical computing for fiber sensing: it moves demodulation into the optical domain, supports spatial multiplexing, and shows interpolation to untrained measurand values. The experimental work is substantial and mostly internally consistent, including repeated measurements, error histograms, comparisons with simulation, and multiple sensor types. The genetic-algorithm in-situ training avoids requiring a full physical model, and the PD-based high-sensitivity measurements provide a concrete path beyond camera-based demonstrations. However, the headline claims of sub-nano strain resolution and nanosecond latency are overstated relative to the reported RMSE and the 90 kHz PD bandwidth, and the interpolation argument from four training states needs additional characterization.","major_comments":[{"comment":"The abstract's claim of 'sub-nano strain resolution' is not supported by the reported RMSE of 1.6160 nε over a 95 nε range, since 1.6160 nε is larger than 1 nε. Please replace 'sub-nano' with 'nanostrain-level' or provide a measurement with RMSE below 1 nε; as written, the headline overstates the demonstrated sensitivity.","section":"Abstract; §2.5, Fig. 5(c)"},{"comment":"The '<3 ns demodulation delay' is an optical-propagation estimate for the 0.5 m encoding MMF and the free-space module (τ_encode = 2.44 ns, τ_compu = 0.08 ns), but it excludes the photodetector. The experimental system uses an InGaAs PD with 90 kHz bandwidth, which limits the recovered 150 kHz signal to a 27 dB SNR and visibly distorts the waveform; a 90 kHz bandwidth corresponds to a response time of order 1.8 μs. Thus the demonstrated end-to-end sensing latency is microseconds, not nanoseconds. Please either measure and report the true step response of the PD-based system, or explicitly restrict the '<3 ns' claim to the optical computing module and revise the abstract's 'nanosecond-latency' wording.","section":"§4.4; §2.5"},{"comment":"The generalization argument that training on four strain states yields a linear mapping over the full range is not fully established. Equation (2) constrains the ODN output only at the training states; it imposes no monotonicity or smoothness constraint on the response between them, and the cited memory effect (ref. 37) concerns spatial shift invariance rather than monotonic spectral decorrelation. Figure 2(c) shows the autocorrelation of speckle patterns decreasing with strain, but monotonic decorrelation of the speckle does not by itself guarantee that the trained output intensity is a single-valued monotonic function of strain. Since Eq. (1) assumes a single-valued linear readout, please report a fine-grained measured output-intensity-versus-strain curve across the full range, not only the 30 sampled states, and quantify monotonicity/no ambiguity, or add a constraint and discussion that rules out non-monotonic regions.","section":"§2.2, Eq. (2), Fig. 2(c)"}],"minor_comments":[{"comment":"The main text states the high-accuracy strain resolution as 2.7754 µε while the Fig. 2(e) caption reports RMSE = 2.7554 µε; please reconcile these numbers.","section":"§2.2"},{"comment":"'optical nerual network' should be 'optical neural network'.","section":"Introduction"},{"comment":"'ig. 3(c)' should be 'Fig. 3(c)'.","section":"Fig. 3 caption"},{"comment":"'espcially' should be 'especially' in the noise analysis paragraph.","section":"§4.5"},{"comment":"'Detecor 1 Detector 2' contains a typo and should read 'Detector 1 Detector 2'.","section":"Fig. 4 caption"},{"comment":"Please define I_obj and describe how the normalized intensity Inorm is computed (which spatial region and normalization procedure) before Eq. (1); currently the reader must infer this from the figures.","section":"§4.2"},{"comment":"For reproducibility, please include the GA population size, number of generations, and number of training states used in each experiment, or at least state explicitly that these details are given in Supplementary Note 6.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The experimental core is credible and the genetic in-situ training is a strength; the main problem is overstatement in the abstract (sub-nano resolution, nanosecond latency) and in the interpolation justification. The requested response-curve characterization is a reasonable and bounded addition. I would support acceptance after the claims are calibrated and the response-curve data are added."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this is a real experimental demonstration, not a simulation. The authors combine a multimode-fiber scattering medium with an SLM-based diffractive optical network trained by a genetic algorithm, and show direct intensity readout for FBG strain, torsion classification, dual-parameter strain/torsion, and a three-joint robotic arm. The strain regression over 30 unseen states after training on four is the strongest evidence: interpolation actually works, and the speckle decorrelation curve in Fig. 2c supports the monotonicity assumption. That is a genuine step beyond the same group's earlier speckle-spectrometer work, and the cited references for those building blocks are appropriate.\n\nThe main soft spots are overstatement, not core physics. The abstract claims 'sub-nano strain resolution,' but the reported RMSE is 1.6160 nε; that is nano, not sub-nano. The '100% torsional angle classification' is on the nine trained angles, so it is a classification demo, not a generalization result. The '<3 ns demodulation delay' is estimated from fiber length and free-space path, not measured end-to-end, and the SLM is electronically driven, so 'no electronic hardware' should be read as 'no digital signal processing.' Data and code are not public, and the supplementary notes with training details were not part of the review package.\n\nOn the stress-test note: the worry that four training points could fit a non-monotonic mapping is legitimate in principle, but the paper's own 30-state test set and the monotonic decorrelation curve in Fig. 2c largely defuse it. A rigorous proof would require a fine-grained sweep, but the evidence is consistent with a smooth, single-valued mapping.\n\nThis is a solid experimental paper that deserves a serious referee. The concept is novel enough for the fiber-sensing community and the experiments are well-executed. I would send it to review, expecting major revisions to trim the abstract and discussion back to what is actually measured.","headline":"Solid all-optical demodulation demonstration with real experimental weight; needs honest claim-trimming on resolution, generalization, and latency before publication.","tokens_in":17689,"tokens_out":3402,"would_cite":true,"duration_ms":36956,"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":"Fiber-optic measurements can be read straight from light intensity, with under 3 nanoseconds of delay and no electronic processing.","keywords":["all-optical fiber sensing","in-sensor computing","speckle","diffractive optical network","multimode fiber","strain sensing","torsion sensing","optical computing"],"falsifier":"Take the FBG strain setup, apply a dense sweep of strain values between and beyond the training states, and measure both the speckle cross-correlation and the optical-network output intensity. If the speckle correlation versus strain is non-monotonic, has plateaus, or exhibits discontinuities, or if the output intensity deviates from the calibration line by more than the reported RMSE at any untrained value, the claimed general linear mapping is falsified.","tokens_in":16679,"feed_emoji":"⚡","tokens_out":8045,"duration_ms":74345,"temperature":0.7,"pith_summary":"This paper claims that fiber-optic sensing can shed its electronic demodulation stage entirely. The authors route light from a fiber sensor through a scattering multimode fiber and then through a trained diffractive optical network, so that the physical quantity being measured—strain, torsion angle, vibration—is mapped directly onto the intensity of light at a detector. A photodetector reading alone then gives the measured value, with a reported demodulation delay below $3\\,\\text{ns}$ and no computer or digital signal processing in the loop. If this holds, conventional interrogators, spectrometers, and electronic processors become optional, and fiber sensing speed becomes limited mainly by detector bandwidth.","feed_headline":"Fiber sensors read out by light alone in under 3 ns","feed_subtitle":"A speckle-generating fiber plus a trained diffractive network turns strain and torsion directly into detector intensity.","key_machinery":"The load-bearing mechanism is the pairing of a scattering medium with a trained diffractive optical network. The scattering medium (a multimode fiber) performs a high-dimensional nonlinear projection that turns tiny changes in wavelength, polarization, or mode content into large, distinguishable speckle changes; the optical diffraction network, implemented with a spatial light modulator and trained by a genetic algorithm, applies phase modulation that spatially re-routes those speckles so that total intensity in a target region equals the measurand. Two properties carry the argument: the deterministic \"memory effect,\" in which speckle patterns decorrelate gradually and monotonically as the measurand changes, allowing interpolation from sparse training states; and linear intensity readout, so that the sensing result is available immediately at the photodetector.","core_discovery":"The central discovery is an architecture, AOFS-IC, that performs sensing demodulation entirely in the optical domain. A scattering medium, here a multimode fiber, converts small optical-field changes—wavelength shift, polarization change, mode coupling—into high-dimensional speckle patterns, and a spatial-light-modulator-based diffractive optical network, trained end-to-end with a genetic algorithm, transforms those speckles so that the intensity in a designated output region is linearly proportional to the measurand. The paper demonstrates this for FBG strain sensing (RMSE $2.7554\\,\\mu\\varepsilon$ over $150\\,\\mu\\varepsilon$ and $0.0688\\,\\text{m}\\varepsilon$ over $2.5\\,\\text{m}\\varepsilon$), for torsional state classification with 100% accuracy, for simultaneous strain and torsion readout, for nanoscale strain down to $1.6160\\,\\text{n}\\varepsilon$ RMS over $95\\,\\text{n}\\varepsilon$, and for 3-DOF robotic-arm joint monitoring. The claimed result is that the trained optical network generalizes from only a few training states to the whole measurement range thanks to the deterministic, gradually decorrelating speckle response (the memory effect).","pith_inferences":["The training-data sparsity suggests a testable scaling law: the required number of training states should grow with the width of the measurement range relative to the decorrelation length of the speckle pattern, so measuring that length directly could predict where the linear mapping breaks down.","Replacing the spatial light modulator with passive etched phase plates—which the paper names as a possibility—would remove the only actively powered optical component, making the sensing head fully passive and potentially deployable in hard-to-reach locations.","The same speckle-to-intensity mapping could be repurposed as an all-optical spectrometer, polarization analyzer, or temperature sensor whenever the measurand leaves a deterministic fingerprint in the speckle pattern, a direction the paper gestures at with its spectrometer and polarization-analyzer extensions.","The accuracy-versus-range trade-off the paper reports looks like a consequence of the memory effect, not merely an engineering fix: extending the dynamic range compresses the intensity response per unit measurand, so system design must choose a range matched to the decorrelation curve."],"forward_implications":["A fiber sensor linked to an AOFS-IC module can report a physical quantity with a total demodulation latency below $3\\,\\text{ns}$, which the paper estimates as more than two orders of magnitude faster than conventional electronic demodulation.","Because the readout is just light intensity, the usable sensing bandwidth is set by the photodetector rather than by any computing hardware; the paper shows $10\\,\\text{kHz}$ and $150\\,\\text{kHz}$ vibrations recovered with SNR $59\\,\\text{dB}$ and $27\\,\\text{dB}$, the latter limited by the PD's $90\\,\\text{kHz}$ bandwidth.","Multiple sensors or multiple measurands can be decoded simultaneously by training the optical network to focus each sensing signal onto a separate spatial region of the detector plane, as demonstrated for simultaneous strain and torsion on one multimode fiber.","The same architecture works across sensor types and measurands—FBG wavelength shifts, MMF torsion and stretching, SMF polarization rotation, and robot-arm joint bending—implying a general all-optical demodulation layer rather than a single-purpose device.","Removing electronic demodulation eliminates the power consumption and latency of interrogators and computers from fiber sensing systems, which the paper argues makes dense, large-scale sensor arrays more practical."],"supporting_citations":[{"why":"Supplies the multimode-fiber speckle spectrometer concept that AOFS-IC uses as its scattering medium and spectral-resolution dependence.","marker":"[35]"},{"why":"Establishes that random scattering media can encode spectral and polarimetric information into speckle, the basis for scattering-medium encoding.","marker":"[33]"},{"why":"Provides the memory effect in multimode fibers that the paper relies on for generalization from sparse training states.","marker":"[37]"},{"why":"Supplies the genetic-algorithm optimization used to train the optical diffractive network without a physical model.","marker":"[36]"},{"why":"Supports the use of high-dimensional nonlinear projection to amplify subtle optical-field changes into measurable speckle variations.","marker":"[34]"},{"why":"Defines the wavelength-accuracy baseline for commercial FBG interrogation that the optical approach is compared against.","marker":"[32]"}],"fun_headline_variants":["All-optical fiber sensing with 3-ns in-sensor computing","Fiber sensing demodulated by light alone in under 3 ns","No electronics: fiber sensor reads out in 3 ns optically","Ultrafast all-optical sensing: speckle + diffractive network","In-sensor optical computing cuts fiber sensing latency to 3 ns"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme depends on the speckle pattern evolving deterministically, monotonically, and without jumps as the measurand changes, so that a network trained on a handful of states can interpolate every value in between.","fun_headline_variants_meta":{"raw":{"variants":["All-optical fiber sensing with 3-ns in-sensor computing","Fiber sensing demodulated by light alone in under 3 ns","No electronics: fiber sensor reads out in 3 ns optically","Ultrafast all-optical sensing: speckle + diffractive network","In-sensor optical computing cuts fiber sensing latency to 3 ns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000618,"raw_usage":{"total_tokens":2901,"prompt_tokens":1011,"completion_tokens":1890,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":1798}},"tokens_in":627,"tokens_out":1890,"duration_ms":12308,"temperature":1.0,"reasoning_tokens":1798,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:33:13.488515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the FBG strain setup, apply a dense sweep of strain values between and beyond the training states, and measure both the speckle cross-correlation and the optical-network output intensity. If the speckle correlation versus strain is non-monotonic, has plateaus, or exhibits discontinuities, or if the output intensity deviates from the calibration line by more than the reported RMSE at any untrained value, the claimed general linear mapping is falsified.","supporting_citations":[{"cited_title":"Redding, M","cited_arxiv_id":null,"evidence_quote":"Supplies the multimode-fiber speckle spectrometer concept that AOFS-IC uses as its scattering medium and spectral-resolution dependence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that random scattering media can encode spectral and polarimetric information into speckle, the basis for scattering-medium encoding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the memory effect in multimode fibers that the paper relies on for generalization from sparse training states."},{"cited_title":"Michalewicz, Genetic algorithms+ data structures= evolution programs","cited_arxiv_id":null,"evidence_quote":"Supplies the genetic-algorithm optimization used to train the optical diffractive network without a physical model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the use of high-dimensional nonlinear projection to amplify subtle optical-field changes into measurable speckle variations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the wavelength-accuracy baseline for commercial FBG interrogation that the optical approach is compared against."}],"review_version":1}