{"id":"666897dc-8686-41f2-917e-951b0b6bfe98","arxiv_id":"2501.02974","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Surrogate models of the time-domain microlensing amplification factor for point-mass and singular isothermal sphere lenses match numerical waveforms with mismatches below about 5e-4 and evaluate in about 100 ms.","lead":"The authors build fast computer models that imitate how gravitational waves are distorted by point-mass or galaxy-like lenses, returning the same waveforms in milliseconds instead of seconds. The work is a proof-of-principle that a standard GW surrogate technique works for the lensing amplification factor, clearing a path toward practical parameter estimation of microlensed events.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Surrogate accuracy claims rest on a reconstructed reference truth whose 5% peak-matching window is empirically tuned; if that window is not robust across parameter space, the reported mismatches understate the error against the true lensed waveform.","rationale":"The paper makes a credible proof-of-principle that surrogate modeling can accelerate evaluation of the time-domain microlensing amplification factor for PML and SIS. The validation uses training and test parameter sets that are disjoint, which is good practice, and the reported mismatches are small. However, the central accuracy claim depends on the fidelity of the reference amplification factor, which is not a direct evaluation of the Kirchhoff integral but a constructed quantity: the area method of Eq. (7) plus the peak reconstruction of Section II B that splices in the analytic logarithmic approximation using an empirically chosen 5% matching window. The authors state that the numerical reference is calibrated to the analytic PML solution and the SIS summation form to better than 1e-4, but no calibration error map, no dependence of the best window choice on y or ML, and no convergence test for the truncated SIS series are provided. If that calibration fails in some region of the tested parameter space, the surrogate mismatches would be measured against a biased truth, and the headline accuracy would not hold against the actual lensed waveform. This is the same weakest assumption identified by the reader, and it justifies the conditional verdict. The concern is not that the surrogate method is wrong, but that the demonstrated accuracy and the claim of reproducing the original lens models require the reconstructed reference to be independently validated and made available. Releasing code and data, and including the proposed window-sensitivity and calibration-error tests, would resolve this issue and support acceptance. Therefore the verdict remains CONDITIONAL, unchanged from the reader's assessment.","tokens_in":22018,"tokens_out":6031,"duration_ms":57198,"concrete_test":"Compute the analytic amplification factor for PML (Eq. A6) and the SIS series (Eq. A13) at a dense grid of 100 y-values spanning the full ranges (PML: 0.1-2, SIS: 0.01-0.95) for ML = 100, 500, 1000 M_sun and zL = 0.05. Then apply the exact peak-reconstruction pipeline to the area-method eF(t) and FFT to obtain F_ref(f), and compare against the analytic F(f) to produce a calibration-error map. Separately, rebuild the surrogate and recompute mismatches against the analytic waveforms while varying the reconstruction window width from 2% to 10% on each side of tpeak. If the calibration error exceeds 1e-4 at any grid point, or if the surrogate-to-analytic mismatch varies by more than a factor of a few from the reported <=5e-4 across that window range, the reference-bias concern lands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the reconstructed time-domain amplification factor eF(t) used as training and validation truth is itself accurate. It is built from the area-between-time-delay-curves expression (Eq. 7) with the peak spliced in from the analytic logarithmic approximation (Eq. 12) inside an empirically chosen 5% matching window (Section II B). The paper states the numerical reference is calibrated to the analytic PML solution (Eq. A6) and SIS series (Eq. A13) to mismatches < 1e-4, but no calibration error map or convergence study is shown, and the SIS series is truncated at n=500 without a convergence check. If the optimal matching-window width varies with impact parameter y or lens mass ML, or if the logarithmic approximation is inaccurate at the high dimensionless frequencies sampled in the ML up to 1000 M_sun cases, the reference could carry a systematic error larger than the claimed 1e-4 in some parts of the grid. Because the surrogate is trained and validated against this same reconstructed reference, the reported mismatches up to 5e-4 would then be relative to a biased truth, and the headline claim that the surrogates reproduce the original lens models would overstate true accuracy. This is the single most load-bearing concern because it directly affects the accuracy claim, not the speed claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes constructing surrogate models of the time-domain gravitational-wave microlensing amplification factor \\tilde F(t) for two spherically symmetric lens models, the point-mass lens (PML) and the singular isothermal sphere (SIS), and then obtaining fast microlensed waveforms by Fourier transforming the surrogate amplification factor and multiplying by an unlensed frequency-domain waveform. The training data are generated from the area-between-time-delay-curves expression of Eq. (7), with the logarithmic peak reconstructed from the analytic near-peak formula Eq. (12) inside an empirically chosen 5% matching window, followed by the regularization of Eq. (13). The surrogate itself uses reduced-basis greedy selection, empirical interpolation, and Gaussian process regression. The authors report mismatches \\lesssim 5 \\times 10^{-4} against the numerical microlensed waveforms across the tested ML-y grids, evaluation times of about 100 ms, and speedups between 5 and 10^3 relative to the underlying lensing models.","tokens_in":22210,"tokens_out":5590,"duration_ms":130970,"significance":"If the accuracy claim holds, the paper supplies a practically useful ingredient for parameter estimation of microlensed gravitational waves, and the methodology is in principle extendable to more complex, asymmetric lens models. The internal validation is careful in several respects: training and test lens parameters are disjoint, mismatches are computed with a standard detector-weighted match using the Einstein Telescope PSD, the greedy convergence curves in Fig. 5 plateau, and speed is measured against a lookup table for PML and against direct numerics for SIS. The main unresolved risk is that the validation truth is itself a reconstructed numerical reference; the manuscript does not supply a calibration error map, an SIS series-convergence check, or a sensitivity study of the empirical 5% reconstruction window. These omissions directly affect the central accuracy claim rather than the speed claim.","major_comments":[{"comment":"The central accuracy claim is validated only against a training/validation truth that is itself reconstructed: \\tilde F(t) from Eq. (7) is not a direct Kirchhoff-integral solution, and the logarithmic peak is replaced by the analytic approximation Eq. (12) inside an empirically chosen 5% matching window. Section III states that the numerical amplification factors are calibrated to the analytic PML form (Eq. A6) and the SIS series (Eq. A13) so that the maximum mismatch is < 10^{-4}, but no calibration error map, no convergence study for the n=500 SIS series truncation, and no sensitivity of the 5% window width are shown. If the reconstruction bias is larger than 10^{-4} at the high dimensionless frequencies sampled at the largest lens masses, the reported surrogate mismatches, which are also computed against this reconstructed reference, would understate the error relative to the true lensed waveform. Please add a calibration error map for both lens models, a convergence test for the SIS series, and a variation of the matching window width (for example 2% and 10%) with the resulting mismatch values.","section":"Section II B and Section III"},{"comment":"Because the PML has the closed-form F(w) in Eq. (A6), the final surrogate should be compared directly against this analytic solution, not only against the calibrated numerical pipeline. Such a comparison would remove the reconstructed-reference issue for the PML entirely and would directly support the headline claim that the surrogate reproduces the original lens model. As it stands, Figs. 7 and 8 only compare against the numerical method that was used to construct the surrogate, so the validation does not independently confirm the accuracy claim for the one model where an exact reference is available.","section":"Section III, Fig. 7"},{"comment":"The greedy search is stopped when the normalized L2 error in the regularized time domain is below 10^{-2}, while the basis tolerance is 5 \\times 10^{-4}, but the relation between these thresholds and the advertised frequency-domain mismatches below 5 \\times 10^{-4} is not established. The final mismatches are computed separately, so this is not an internal contradiction, but the manuscript should clarify whether the reported mismatches are sensitive to the L2 stopping criterion and the basis tolerance; a small variation of these two parameters with the resulting mismatch statistics would make the empirical choices in the greedy loop less opaque.","section":"Section II D 3 and Fig. 5"}],"minor_comments":[{"comment":"The symbol \\mu_- is signed negative for the saddle image in Eq. (A8), but Eq. (13) takes its square root; the formula should use |\\mu_-|, or \\mu_- should be redefined as the absolute magnification.","section":"Eq. (13)"},{"comment":"The summation indices in Eqs. (15) and (20) appear as 'i=i' and 'j=i'; these should be 'i=1' and 'j=1'.","section":"Eqs. (15) and (20)"},{"comment":"The set C in min_{c_i \\in C} is not defined; presumably it is \\mathbb{R}^m, and this should be stated for clarity.","section":"Eq. (23)"},{"comment":"The relationship between the number of basis vectors (27 for PML, 15 for SIS) and the number of greedy training parameters (35 for PML, 100 for SIS) should be clarified; the text is easy to misread as saying the basis is built from all 35 or 100 greedy points.","section":"Section III"},{"comment":"The abstract states mismatches \\lesssim 5 \\times 10^{-4}, the introduction mentions O(10^{-7}-10^{-3}), and Section IV reports 10^{-8}-10^{-4}; these ranges should be made consistent with the summary statistics behind Fig. 8.","section":"Abstract and Section IV"},{"comment":"The histogram of evaluation times is described as comparing the surrogate with the lookup table, but the caption does not state whether the surrogate time includes the FFT that the text identifies as the rate-limiting step; please specify.","section":"Fig. 8, top-right panel"},{"comment":"A data and code availability statement would improve reproducibility, since the numerical method and the peak-reconstruction procedure involve several implementation choices that are only described at a high level.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"I agree with the conditional assessment: the method is promising and the paper is within the journal's scope, but the missing calibration error map and convergence tests are load-bearing for the accuracy claim. The PML direct-validation point is the easiest and most convincing fix; if the authors add it, along with an SIS convergence check and a window-sensitivity test, the paper would in my view be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is a clean, incremental methods paper. It builds reduced-basis/GPR surrogates for the time-domain lensing amplification factor Ftilde(t) for a point mass and a singular isothermal sphere, reports mismatches below 5e-4 against the numerical reference, and is 5–1000x faster. The speed claim holds up. The novelty is applying surrogate modeling to the amplification factor itself, in the time domain, rather than to the final waveform; that is a real and useful step, not a conceptual breakthrough.\n\nWhat the paper does well: the validation is honest. Training and test parameters are disjoint, the match is a standard detector-weighted overlap, and the greedy convergence curves plateau. They benchmark the PML surrogate against the LVK lookup table and show it is faster. They are also upfront about the caustic/multiple-image limitation, which sets the right scope for a proof-of-principle. The numerical reference is calibrated to analytic results (closed-form PML, SIS series truncated at n=500) at a claimed mismatch below 1e-4, so the surrogate is not purely self-referential.\n\nThe soft spots are real but not fatal. The main one is the reference truth. The numerical Ftilde(t) is built from the area-between-time-delay-curves expression plus a 5% window splice of the analytic logarithmic peak. The 5% is empirically chosen, and there is no sensitivity study showing that the window width is robust across impact parameter and lens mass. For the SIS series, the n=500 truncation has no convergence check; the 1e-4 calibration statement is asserted, not demonstrated. If the reference carries a systematic error near the peak or at high frequencies, the reported mismatches could understate the true error against the actual lensed waveform. I think that risk is moderate, not disqualifying, because the analytic anchors keep the reference from being wildly wrong, but it should be shown rather than asserted. Second, no code or data is shipped. For a methods paper that is a genuine gap; the training data and evaluation scripts would make the work reproducible and much more useful. Third, the SIS surrogate is compared only against the numerical solver, not against the existing time-domain lookup table (Ref 95). That is the closer baseline and its absence weakens the speed claim.\n\nWho this is for: anyone working on microlensed GW parameter estimation with next-generation detectors. It is a practical tool paper, not a physics breakthrough.\n\nMy recommendation: send it out for peer review. The referee should ask for the code and data, a convergence check on the SIS series, and a short study varying the 5% matching window. With those, this becomes a solid methods contribution.","headline":"A solid proof-of-principle that surrogates can speed up evaluation of the time-domain microlensing amplification factor; the main caveat is that the training truth itself is a reconstructed quantity, though it is anchored to analytic PML and SIS results.","tokens_in":22822,"tokens_out":2243,"would_cite":true,"duration_ms":21260,"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":"Surrogate modeling reproduces microlensed gravitational-wave amplification factors for point-mass and singular isothermal sphere lenses, with mismatches below $5\\times10^{-4}$ and evaluation near 100 ms.","keywords":["gravitational wave microlensing","surrogate modeling","amplification factor","wave optics","time-delay curves","point-mass lens","singular isothermal sphere","reduced basis"],"falsifier":"Take the point-mass-lens surrogate and compare its frequency-domain amplification factor directly with the exact analytic expression of Eq. (A6) at the upper end of the next-generation detector band (for example $M_L=10^3\\,M_\\odot$ and $y=0.1$), instead of comparing with the reconstructed numerical reference. If the mismatch there exceeds the reported $\\sim5\\times10^{-4}$, the peak-reconstructed training reference is biased at high dimensionless frequency and the surrogate's advertised accuracy is not the true error. The same check can be run for the SIS surrogate against the high-order summation form of Eq. (A13).","tokens_in":21772,"feed_emoji":"🔭","tokens_out":11601,"duration_ms":97471,"temperature":0.7,"pith_summary":"Gravitational-wave microlensing imprints wave-optics features on a signal through an amplification factor, but most realistic lenses have no closed-form expression for it, and numerical evaluation is too slow for the large Bayesian searches that a detection will require. This paper establishes that surrogate modeling, an interpolation method already used for binary-black-hole waveforms, can instead be applied to the time-domain amplification factor $\\widetilde{F}(t)$, which is smooth enough to interpolate after the logarithmic saddle-image peak is reconstructed analytically and then regularized. Building surrogates for the point-mass lens and the singular isothermal sphere, the paper reports microlensed waveforms that match the underlying numerical lens models with mismatches below about $5\\times10^{-4}$ and evaluate in about 100 ms, between 5 and 1000 times faster than the original models. If this holds, parameter estimation of microlensed gravitational-wave events becomes practical rather than prohibitive.","feed_headline":"Surrogate models make microlensed GW waveforms 1000x faster","feed_subtitle":"An interpolant of the time-domain amplification factor hits 0.05% mismatch for two lens models.","key_machinery":"Two constructions do the work. First, the amplification factor is computed and modeled in the time domain: $\\widetilde F(t)=dS/dt$, the rate of change of the area on the lens plane between curves of constant time delay, which avoids the rapidly oscillating Kirchhoff integrand. Second, the surrogate itself is a reduced-basis and empirical-interpolation model: a greedy algorithm selects a small set of representative waveforms, empirical interpolation chooses sparse time nodes at which the waveform must be known, and Gaussian-process regressions fit the dependence on the impact parameter $y$ at those nodes. Two preprocessing steps make interpolation accurate near the saddle-image peak: the peak is reconstructed by splicing a rescaled analytic logarithmic approximation over an empirically chosen $\\pm5\\%$ window around $t_{\\mathrm{peak}}$, and the regularization $\\bar F(t)=1-\\exp[-\\pi(\\widetilde F(t)-1)/\\sqrt{\\mu_-}]$ removes the logarithmic divergence so the interpolation target is smooth. The lens mass $M_L$ only sets the time scale, so each model is a single-parameter surrogate in $y$.","core_discovery":"The paper's central claim is that the time-domain amplification factor $\\widetilde F(t)$, defined by $\\widetilde F(t)=\\int d^2x\\,\\delta(\\tau_d(x,y)-t)$ and computed as the area swept out between neighboring constant-time-delay curves, is a suitable surrogate-modeling target even though the frequency-domain factor $F(w)$ is highly oscillatory. For a point-mass lens over impact parameters $y\\in[0.1,2]$ and a singular isothermal sphere over $y\\in[0.01,0.95]$ (the two-image regime), the authors first reconstruct the logarithmic peak contributed by the saddle image using the analytic near-peak form of Ref. [60], then apply the regularization $\\bar F=1-\\exp[-\\pi(\\widetilde F-1)/\\sqrt{\\mu_-}]$, and build reduced-basis surrogates with 27 and 15 basis vectors at a basis tolerance of $5\\times10^{-4}$. After inverse regularizing and Fourier transforming, the surrogate amplification factor is multiplied into unlensed binary-black-hole waveforms; the resulting mismatches against the numerical reference are below $5\\times10^{-4}$ across the tested $M_L$-$y$ grid and for binary total masses from $20$ to $100\\,M_\\odot$, with frequency-domain waveform evaluation times of order $10^{-2}$-$10^{-1}$ s. The claim is that this makes surrogate modeling a viable route to fast, accurate microlensed templates in regimes where no analytic amplification factor exists.","pith_inferences":["A direct high-frequency validation against the analytic point-mass result, rather than the reconstructed reference, would sharpen the accuracy claim, since the paper's own discussion identifies peak resolution as the main high-frequency error source.","Because the lens mass only rescales time, the single-parameter-in-$y$ surrogates can be extended to a two-parameter $(M_L,y)$ model simply by rescaling the time axis, which would make them directly usable in parameter estimation.","For asymmetric or multi-lens configurations the paper notes that caustics split parameter space into regions with different image counts; separate surrogates per region are a plausible path, but the caustic geometry itself would need to be modeled.","The rate-determining step is the Fourier transform, not the surrogate evaluation, so faster FFT algorithms or direct frequency-domain interpolation would push the total cost well below 100 ms."],"forward_implications":["Large-scale Bayesian parameter estimation of microlensed gravitational-wave signals becomes computationally feasible, since surrogate templates evaluate in about 100 ms rather than seconds.","The same time-domain prescription, compute $\\widetilde F(t)$, reconstruct and regularize the peak, and build a reduced-basis surrogate, extends to any lensing potential for which numerical time-delay solutions can be produced, without needing an analytic $F(f)$.","For the point-mass lens the surrogate is faster than the existing lookup-table interpolation used in current searches while remaining fully wave-optics based.","The singular-isothermal-sphere surrogate demonstrates that a logarithmic divergence in $\\widetilde F(t)$ does not block accurate interpolation when the peak is handled analytically before surrogate construction."],"supporting_citations":[{"why":"Supplies the time-domain route to the amplification factor as area between constant-time-delay curves, the function the surrogate interpolates, and the analytic near-peak form used for reconstruction.","marker":"[60]"},{"why":"Provides the specific method for computing areas between nearby time-delay curves used to produce the training data.","marker":"[51]"},{"why":"Establishes the reduced-basis and empirical-interpolation surrogate framework that the paper adapts to lensing amplification factors.","marker":"[61]"},{"why":"Describes the greedy training-set selection algorithm used to choose lens parameters for the validation loop.","marker":"[62]"},{"why":"Supplies the Gaussian-process-regression fitting procedure used at the empirical nodes.","marker":"[69]"},{"why":"Gives the exact analytic point-mass amplification factor used to calibrate the numerical reference.","marker":"[97]"},{"why":"Gives the summation-form amplification factor for the singular isothermal sphere used to calibrate the SIS numerical reference.","marker":"[98]"},{"why":"Provides the lookup-table interpolation baseline that the point-mass surrogate is benchmarked against for speed.","marker":"[37]"},{"why":"Generates the unlensed binary-black-hole waveforms that are multiplied by the surrogate amplification factor in the mismatch tests.","marker":"[93]"}],"fun_headline_variants":["Microlensed GW surrogates: 1000x faster, 0.05% mismatch","Surrogate models speed up microlensed GW parameter estimation","Fast surrogates for lensed gravitational waves hit 0.05% error","Time-domain surrogates make microlensed GW analysis practical"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reference waveforms used to train and test the surrogate are not direct solutions of the Kirchhoff diffraction integral: they come from the area-between-time-delay-curves formula of Eq. (7), with the logarithmic peak replaced by a rescaled analytic approximation over an empirically chosen $\\pm5\\%$ matching window. If that reconstructed reference is biased at high frequencies, or if the 5% window is not robust across parameter space, the reported mismatches understate the surrogate's true error against the actual lensed waveform.","fun_headline_variants_meta":{"raw":{"variants":["Microlensed GW surrogates: 1000x faster, 0.05% mismatch","Surrogate models speed up microlensed GW parameter estimation","Fast surrogates for lensed gravitational waves hit 0.05% error","Time-domain surrogates make microlensed GW analysis practical"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000429,"raw_usage":{"total_tokens":2300,"prompt_tokens":1162,"completion_tokens":1138,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":778,"completion_tokens_details":{"reasoning_tokens":1056}},"tokens_in":778,"tokens_out":1138,"duration_ms":11351,"temperature":1.0,"reasoning_tokens":1056,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:58:59.165963+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the point-mass-lens surrogate and compare its frequency-domain amplification factor directly with the exact analytic expression of Eq. (A6) at the upper end of the next-generation detector band (for example $M_L=10^3\\,M_\\odot$ and $y=0.1$), instead of comparing with the reconstructed numerical reference. If the mismatch there exceeds the reported $\\sim5\\times10^{-4}$, the peak-reconstructed training reference is biased at high dimensionless frequency and the surrogate's advertised accuracy is not the true error. The same check can be run for the SIS surrogate against the high-order summation form of Eq. (A13).","supporting_citations":[{"cited_title":"Gravitational wave lensing beyond general relativity: Birefringence, echoes, and shadows,","cited_arxiv_id":null,"evidence_quote":"Provides the specific method for computing areas between nearby time-delay curves used to produce the training data."},{"cited_title":"Search for gravitational lensing signa- tures in ligo-virgo binary black hole events,","cited_arxiv_id":null,"evidence_quote":"Establishes the reduced-basis and empirical-interpolation surrogate framework that the paper adapts to lensing amplification factors."},{"cited_title":"gwastro/pycbc: v2.0.4 release of pycbc,","cited_arxiv_id":null,"evidence_quote":"Gives the exact analytic point-mass amplification factor used to calibrate the numerical reference."},{"cited_title":"A general multipurpose interpolation procedure: the magic points,","cited_arxiv_id":null,"evidence_quote":"Generates the unlensed binary-black-hole waveforms that are multiplied by the surrogate amplification factor in the mismatch tests."}],"review_version":1}