{"id":"ee4ffb8e-114f-4931-a8ed-e1a05950396c","arxiv_id":"2507.05249","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual implicit neural network with a physics-guided convolution kernel separates single-magnon signals from heterogeneous background in 4D neutron scattering data without labels.","lead":"This paper trains two neural networks together to separate a clean physical signal from a noisy background in 4D neutron scattering data, using a physics simulation to anchor the signal part. It is a label-free cleanup method for experiments where background overwhelms the signal of interest.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthetic validation is circular: ground truth is generated by the same convolution-plus-background model being tested, so model mismatch in the experimental data remains untested.","rationale":"The reader's weakest-assumption identification (model mismatch in the convolutional form and smoothness of the background) is real and technically accurate, but the more load-bearing issue is that the paper's synthetic validation cannot detect that mismatch because the synthetic ground truth is generated by the same model family. The construction in Section 3.2 explicitly uses the reconstructed background from the experimental run and the same S_sim/kernel form, so the benchmark tests self-consistency rather than scientific fidelity. This does not make the method wrong, but it means the strongest experimental claim is supported only by plausibility and by hyperparameter-based arguments. The chi-squared p-value anomaly and the lack of baselines reinforce this but are secondary. A conditional verdict is therefore appropriate: accept only after an independent synthetic benchmark with model mismatch and ideally code release. The reader's verdict is already CONDITIONAL, so no change is required, though the justification should emphasize the circularity more sharply.","tokens_in":14305,"tokens_out":2789,"duration_ms":43501,"concrete_test":"Build a synthetic benchmark where the ground-truth signal is generated by an independent forward model not equal to the pre-trained S_sim used in the pipeline, for example a linear spin-wave calculation with different J/J_p values or one augmented with a Lorentzian lifetime broadening, and where the background is generated by an independent stochastic process that includes sharp elastic and Bragg-like features. Run the pipeline with fixed hyperparameters on this benchmark and quantify MSLE-sig and MAE-bkg as in Table 1. If the separation errors remain comparable to Table 1, the circularity concern is mitigated; if errors degrade substantially, the experimental decomposition cannot be certified without independent validation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the decomposition into a single-magnon signal and a background is physically correct on experimental data. The paper's only quantitative support is the synthetic experiment of Section 3.2, but that experiment generates its ground truth from the very same model family used for separation. Equation (8) constructs S_syn using the same S_sim, the same learnable-kernel convolution form (3), and a background B_hat obtained by training the pipeline on the experimental data. The test therefore verifies only that the dual-INR can invert its own generative process when the model is exactly correct. It does not test what happens when the real single-magnon response deviates from the pre-trained S_sim (e.g., due to neglected interactions, an incorrect spin-wave Hamiltonian, or non-Gaussian lifetime broadening), nor whether a non-smooth background (sharp elastic peaks, Bragg tails) can be absorbed by the signal pathway. Since the experimental decomposition has no ground truth, the claim that the extracted component is the physically meaningful single-magnon signal is not established. Additional supporting concerns include Table 2's anomalous chi-squared p-value of 1.0 for the supposedly optimal configuration, which suggests an incorrect statistical computation, and the absence of any quantitative baseline comparison or code release. None of these individually disprove the method, but together they place the central claim on a circular validation loop.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a self-supervised source-separation method based on a dual implicit neural representation (INR) architecture. One INR represents the signal of interest as a spatially varying convolution of a pre-trained simulated spectrum with a learnable kernel, while a second INR represents the background; the two are trained jointly by minimizing a reconstruction loss with a background-magnitude penalty. The method is demonstrated on four-dimensional inelastic neutron scattering (INS) data of La2NiO4, with the goal of extracting the single-magnon signal from a heterogeneous background. The authors also propose an analytic estimator for the regularization parameter and report that the trained INRs provide a 792:1 compression of the experimental data.","tokens_in":14472,"tokens_out":5162,"duration_ms":64392,"significance":"If the central claim is correct, the method would be a practically useful tool for extracting weak, physically meaningful signals from high-dimensional scattering data without labels or manual background subtraction. The combination of a physics-guided pre-trained forward model with a dual-INR decomposition is a sensible and potentially generalizable idea, and the analytic regularization-parameter estimate is a useful practical contribution. However, the current evidence does not establish the central claim: the synthetic validation is circular because the ground truth is generated from the same model family and even the same reconstructed background used by the method, and the experimental demonstration lacks any quantitative ground truth. The absence of baseline comparisons further weakens the claim of superiority over conventional approaches. The paper also misses an opportunity to test robustness to model mismatch, which is the main risk for real experimental use.","major_comments":[{"comment":"The synthetic validation is circular. In Eq. (8), the synthetic background is the reconstructed background B_hat obtained by training the method on the experimental data, the signal is the same S_sim used in the forward model, and the distortion is generated by the same learnable-kernel convolution as Eq. (3). The benchmark therefore verifies only that the dual-INR can invert its own generative process when the model is exactly correct. It does not test whether the decomposition remains accurate when the true single-magnon response deviates from S_sim (for example, due to an incomplete spin-wave Hamiltonian, neglected interactions, or non-Gaussian broadening), nor whether a non-smooth background can be absorbed by the signal pathway. I recommend adding a synthetic experiment in which the ground truth is generated from an independent forward model with known perturbations that are not available to the method, and reporting the separation error against that independent ground truth.","section":"Section 3.2, Eq. (8)"},{"comment":"The central physical claim on experimental data rests on the assumption that the true distorted single-magnon signal is exactly a convolution of the simulated signal S_sim with a spatially varying learnable kernel. The paper itself acknowledges in Section 4 that the work relies on simulated signals S_sim, but it never analyzes the bias induced if this convolutional model is wrong. For example, if the real single-magnon response has an additional component not representable by the parametric spin-wave model, the background INR could absorb it, or the signal pathway could be forced to fit a biased decomposition. The authors should provide a concrete robustness test, such as injecting a known non-convolutional distortion or an added spectral mode into a synthetic background and measuring how well the signal pathway recovers the known component.","section":"Section 2.2, Eq. (3); Section 4"},{"comment":"The reported chi-squared p-value of 1.0 for the supposedly optimal configuration (r = 2, lambda = 0.0005, chi-squared = 6828.510) is statistically implausible unless the number of degrees of freedom is incorrectly specified or the statistic is computed in a nonstandard way. Since Table 2 is used to support the claim that this configuration is optimal according to 'statistical analysis,' this anomaly undermines the hyperparameter-selection evidence. The authors should specify the exact chi-squared statistic, the number of degrees of freedom, and the fitting procedure used, or replace this criterion with a calibrated test such as reduced chi-squared.","section":"Table 2"},{"comment":"The hyperparameter optimization is not internally consistent. In Table 1, the configuration {r = 2, lambda = 0.0005} has the best reconstruction loss (RMSE, PSNR, SSIM), while {r = 3, lambda = 0.005} has the best signal-component recovery (MSLE-sig = 0.001). The analytic lambda estimator in Eq. (7) yields lambda* = 0.005, which matches one optimum but not the other. The paper does not explain which criterion should be used for the physical separation task, nor does it reconcile the discrepancy. This weakens the claim that the analytic estimator provides a reliable default choice.","section":"Table 1 and Section 3.4.1"},{"comment":"The paper criticizes conventional source-separation methods but provides no quantitative comparison against any baseline, such as global background subtraction, ICA, RPCA, Double-DIP, or a simple deconvolution with a Gaussian kernel. Without at least one baseline comparison on the synthetic data with known ground truth, the claim that the proposed method 'more effectively' handles heterogeneous backgrounds and unknown distortions is unsupported. Adding such a comparison would also help calibrate the practical significance of the reported reconstruction metrics.","section":"Section 3 and Section 1"}],"minor_comments":[{"comment":"The caption labels the background panel as P_{L,omega} hat{S}^{(1)}_{sig}, but based on the text this should be P_{L,omega} hat{S}^{(2)}_{sig}; the same mislabel appears to affect the caption for panel (b).","section":"Figure 3 caption"},{"comment":"The typesetting of Eq. (7) is malformed: the denominator lacks explicit norm bars, and the symbols Omega, Omega^c, f_approx, and LPF are not defined in the immediately surrounding text. These should be defined to make the estimator reproducible.","section":"Equation (7)"},{"comment":"The stated approximation lambda approximately ||N||^2 / ||S^{(2)}_{sig}||^2 is asserted without derivation. Since this expression is the basis for the analytic estimator, a short derivation or a reference is needed.","section":"Section 2.3"},{"comment":"There is a typo in 'Voigt profiles' (rendered as 'V oigt'), and the sentence beginning 'Furthermore, in many experimental scenarios...' is a run-on that should be split.","section":"Section 3.1"},{"comment":"The SIREN frequency scaling w0 = 30 and the use of Softmax on the kernel output are not discussed in the main text; a brief sensitivity analysis or justification for these choices would improve reproducibility.","section":"Appendix B, Table B1"},{"comment":"No code or trained model is released. Given the reliance on pre-trained components from Ref. [5] and the detailed hyperparameter choices, releasing the pipeline would substantially strengthen the reproducibility of the results.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the experimental application is interesting, but the validation strategy is the main weakness. The circular synthetic benchmark and the absence of a baseline comparison are fixable within the manuscript's scope by adding independent synthetic tests and at least one baseline. I therefore recommend major revision rather than rejection. The novelty relative to the authors' previous INR works [5,6] is moderate, and the paper should more clearly distinguish what is new here from those prior contributions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zach,\n\nThis one is a classic 'good architecture, weak evidence' situation. The idea is genuinely worth a look: a dual-INR decomposition where one branch applies a learnable, coordinate-dependent convolution kernel to a pre-trained spin-wave simulation, and the other learns a smooth background. That specific configuration is not in Double-DIP or Roddenberry et al., and it is a sensible way to inject physics into the source-separation prior. The analytic lambda estimate in Section 2.3 and the reported 792:1 compression are also nice side results.\n\nWhat I cannot get past is the circular synthetic validation. In Section 3.2 the synthetic ground truth is generated as B_hat + S_sim * kappa, where B_hat is the background network trained on the experimental data and kappa is the learned kernel from that same pipeline. So the test only confirms the network can invert its own generative process when the model is exactly correct. It says nothing about what happens when the real signal deviates from the spin-wave simulation or when the background has sharp elastic or Bragg features that the signal pathway could absorb. The experimental demonstration has no ground truth, and the hyperparameter selection there is post-hoc (physical consistency plus a rule of thumb), so the central claim that the method 'successfully separates' the single-magnon signal is not established.\n\nTwo smaller problems: there are no quantitative baselines, and the chi-squared p-value of 1.0 for the supposedly optimal configuration in Table 2 is anomalous enough that I suspect an incorrect statistical computation. The paper would also be much stronger with code release.\n\nIf I were the editor, I would send it to a serious referee. The architecture is a real contribution and the experimental dataset is a good test bed, but the validation needs a major rework: an independent synthetic benchmark with a different generative model (e.g., an unmodeled distortion or a non-smooth background component), a comparison against Double-DIP or a simple Gaussian-deconvolution plus smooth-background fit, and a corrected chi-squared analysis. The paper is honest about being a proof of concept, and it reads like a competent group's serious attempt, so it deserves the full review process rather than a desk reject.","headline":"Promising physics-guided INR decomposition, but the synthetic validation is circular and the experimental claim lacks independent support.","tokens_in":15112,"tokens_out":2720,"would_cite":false,"duration_ms":30805,"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":"A self-supervised dual implicit-neural-network method separates the single-magnon signal from background in 4D neutron scattering data without labels.","keywords":["Self-supervised learning","Implicit Neural Representation","Image Decomposition","Source Separation","Signal Processing","Physics Data Analysis","Inelastic neutron scattering","La2NiO4"],"falsifier":"Generate a synthetic 4D inelastic-neutron dataset from the same spin-wave model but distort the signal with an operation that cannot be expressed as convolution with a local kernel, for example an intensity-dependent peak shift or an extra magnon branch absent from the simulation; train the dual INR, and compare the recovered signal to the known ground truth. A systematic failure to recover it would falsify the central claim.","tokens_in":14009,"feed_emoji":"🧲","tokens_out":7068,"duration_ms":76915,"temperature":0.7,"pith_summary":"This paper proposes a self-supervised source-separation method built from two implicit neural representations (INRs) that are trained jointly on raw measurements. One INR learns a coordinate-dependent convolution kernel that maps a physics-simulated spin-wave spectrum to the distorted single-magnon signal actually measured, while the other INR represents all remaining contributions as a smooth background. The authors demonstrate on four-dimensional inelastic neutron scattering data from La2NiO4 that this dual-INR decomposition recovers the single-magnon component without labels or predefined dictionaries, and they give an analytical rule for setting the regularization strength. If the claim holds, the approach gives a general recipe for pulling a physically meaningful signal out of heterogeneous, structured backgrounds in high-dimensional experimental data.","feed_headline":"Two neural nets separate signal from background in 4D neutron data","feed_subtitle":"Self-supervised method recovers the single-magnon spectrum of La2NiO4 without labels, using a learnable distortion kernel.","key_machinery":"The central object is the physics-guided dual-INR decomposition with a learnable convolutional pathway. A pre-trained network evaluates the simulated spin-wave spectrum $S_\\mathrm{sim}$ at neighboring coordinates; a kernel network (a sinusoidal-activation INR) outputs a localized, $Q,\\omega$-dependent kernel $\\kappa_{r,\\hat\\phi}$; and the distorted signal of interest is the convolution $\\hat{S}^{(1)}_\\mathrm{sig} = S_\\mathrm{sim} * \\kappa_{\\hat\\phi}$. A second INR, the background network $B_{\\hat\\theta}$, represents everything else. The objective $L = \\|S^*_\\mathrm{expt} - \\bar{S}^\\mathrm{pred}_\\mathrm{expt}\\|^2 + \\lambda\\|B_{\\hat\\theta}\\|^2$ is minimized jointly, and the inductive bias that each INR is efficient for its own component but inefficient for the other, together with the injected physics model, produces the separation.","core_discovery":"The central claim is that the dual-INR framework successfully separates physically meaningful signals from a complex or structured background even when signal characteristics vary across all four dimensions of the parameter space. In the demonstration, the signal of interest is the single-magnon excitation spectrum of the square-lattice spin-1 antiferromagnet La2NiO4. It is reconstructed as the convolution of a simulated spin-wave spectrum $S_\\mathrm{sim}(Q,\\omega;J,J_p)$ with a learnable, spatially varying kernel $\\kappa_{r,\\hat\\phi}$, while a separate background INR $B_{\\hat\\theta}$ captures all other sample-related signals. Training minimizes a reconstruction loss plus a $\\lambda$-weighted penalty on the background magnitude, with no labels and no pre-defined dictionaries. The paper reports quantitative recovery on synthetic data and physically consistent separation on experimental data, and identifies the analytically derived $\\lambda^*$ as the setting that best recovers the signal component.","pith_inferences":["An implication the authors leave implicit is that the same scheme could separate other excitation channels, such as the multi-magnon continuum, by choosing a different simulated spectrum for the signal pathway; whether the background INR then absorbs the targeted channel is an open question.","The paper's analytical $\\lambda$ recipe assumes the background is smooth and the signal has finite support; in settings with spatially varying noise or overlapping spectral peaks, a data-driven $\\lambda$ schedule or a spatially weighted loss would be a natural testable extension.","If the assumption of a local convolutional distortion fails, the kernel window $r$ would have to grow with the correlation length of the distortion, which could make the method expensive; replacing the local kernel with a nonlocal or conditional kernel is a direct extension.","The success in 4D suggests the INR basis could be especially effective for other high-dimensional scientific measurements with low intrinsic dimensionality, but the paper only demonstrates one instance; applying the method to a different spectroscopy with known ground truth would test the generality."],"forward_implications":["The single-magnon signal in La2NiO4 inelastic neutron scattering data can be extracted without manual background subtraction, revealing features in momentum-energy space that are otherwise hidden.","The analytical formula for $\\lambda^*$ lets practitioners set the regularization weight from data outside the signal support instead of relying on cross-validation.","Because the decomposition uses compact INRs, it simultaneously denoises the data and compresses it; the paper reports a 792:1 compression ratio for the La2NiO4 dataset.","The same dual-INR structure, with the signal pathway replaced by an appropriate forward model, is proposed as a general tool for source separation in astronomy, biomedical imaging, and other experimental fields.","The framework naturally extends to more than two sources by adding further INRs, allowing multi-component decompositions of the measured signal."],"supporting_citations":[{"why":"Supplies the pre-trained neural representation of the simulated spin-wave spectrum and the INR approach to experimental steering that the dual-INR pipeline builds on.","marker":"[5]"},{"why":"Establishes implicit neural representations for capturing dynamical correlations in the same type of inelastic neutron scattering measurement and motivates the smooth-background INR.","marker":"[6]"},{"why":"Provides the La2NiO4 experimental dataset, data folding, and the spin-wave model parameters used for the simulations.","marker":"[15]"},{"why":"Supports the assertion that neural networks act as smooth priors, which is the inductive bias that lets the background INR suppress noise and avoid absorbing the signal.","marker":"[1]"},{"why":"Demonstrates a prior dual-INR signal decomposition into a smooth part and an auxiliary component, the template the authors extend to a physics-guided convolutional pathway.","marker":"[17]"},{"why":"Supplies the sinusoidal-activation SIREN architecture used for the kernel and background networks.","marker":"[19]"}],"fun_headline_variants":["No labels: dual neural nets split 4D neutron spectra","Self-supervised nets unmix signal and background in 4D neutron data","Dual neural representations isolate signal from background in 4D","Physics-guided AI separates signal from background in 4D","Neural net pair cleans 4D neutron scattering without labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the measured data can be written as a convolution of the simulated spin-wave spectrum with a spatially varying learnable kernel plus a smooth network background; if the true single-magnon response departs from this form, the separation will be biased.","fun_headline_variants_meta":{"raw":{"variants":["No labels: dual neural nets split 4D neutron spectra","Self-supervised nets unmix signal and background in 4D neutron data","Dual neural representations isolate signal from background in 4D","Physics-guided AI separates signal from background in 4D","Neural net pair cleans 4D neutron scattering without labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1376,"prompt_tokens":951,"completion_tokens":425,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":337}},"tokens_in":567,"tokens_out":425,"duration_ms":5066,"temperature":1.0,"reasoning_tokens":337,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:29:32.014401+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a synthetic 4D inelastic-neutron dataset from the same spin-wave model but distort the signal with an operation that cannot be expressed as convolution with a local kernel, for example an intensity-dependent peak shift or an extra magnon branch absent from the simulation; train the dual INR, and compare the recovered signal to the known ground truth. A systematic failure to recover it would falsify the central claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained neural representation of the simulated spin-wave spectrum and the INR approach to experimental steering that the dual-INR pipeline builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes implicit neural representations for capturing dynamical correlations in the same type of inelastic neutron scattering measurement and motivates the smooth-background INR."},{"cited_title":"Petsch, N","cited_arxiv_id":null,"evidence_quote":"Provides the La2NiO4 experimental dataset, data folding, and the spin-wave model parameters used for the simulations."},{"cited_title":"Berg and K","cited_arxiv_id":null,"evidence_quote":"Supports the assertion that neural networks act as smooth priors, which is the inductive bias that lets the background INR suppress noise and avoid absorbing the signal."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates a prior dual-INR signal decomposition into a smooth part and an auxiliary component, the template the authors extend to a physics-guided convolutional pathway."}],"review_version":1}