{"id":"d0b7a728-a8d5-44dc-8b4a-bf72d3a0ca4d","arxiv_id":"2506.02957","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A physics-informed neural network (SPINN) solves the Schrödinger-Poisson equations for fuzzy dark matter collapse in 1D and 3D, matching a spectral solver on a sinusoidal test case.","lead":"This paper uses physics-informed neural networks to simulate the gravitational collapse of fuzzy dark matter, a hypothetical wave-like form of dark matter. The authors show their method reproduces standard numerical solutions for simple test cases, hinting at a possible new way to run dark matter simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The accuracy claim rests on qualitative visual comparisons; no quantitative error norms are reported, so the validation of SPINN against the spectral benchmark is not yet established.","rationale":"The reader's CONDITIONAL verdict is appropriate, and my pass does not move it. The independent spectral comparison and the unsupervised training setup are genuine strengths; I see no internal inconsistency that would require rejection of the proof-of-concept. The weakest point, however, is not only the representativeness of the test case (the reader's chosen concern) but that even this single test case is never scored quantitatively. The abstract promises accurate predictions of mass conservation, density profiles, and structure suppression, yet the paper reports no error norms, no loss values, and no mass-drift tolerance. This makes the central accuracy claim hard to verify from the text alone. The proposed recomputation is inexpensive and decisive: if the errors are already large on the simple sinusoidal test, then the representativeness question is moot; if the errors are small, the remaining limitation is exactly the one the reader identified, namely extrapolation to complex cosmological initial conditions and long evolution times. The manuscript's own Section 5 limitation statement supports the conditional verdict, but the missing quantitative validation should be fixed in revision rather than left implicit in figures.","tokens_in":19466,"tokens_out":4485,"duration_ms":59808,"concrete_test":"Reproduce the 1D and 3D runs of Section 4 with the provided code and compute, at every reported output time, (i) the relative L2 density error e2(t) = ||rho_SPINN(t) - rho_spectral(t)||_2 / ||rho_spectral(t)||_2, (ii) the maximum pointwise density error, and (iii) the maximum absolute deviation of the total mass, |int |Psi|^2 dx - 1|, over t in [0,3]. Report these as a table and repeat the training with 2-3 different collocation-point random seeds. If e2(t) remains below 10% and the mass deviation below 1% for both 1D and 3D, the accuracy claim is quantitatively supported; otherwise the paper should be revised to state the achieved accuracy explicitly and to soften the headline claims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is that SPINN 'accurately predicts' mass conservation, density profiles, and structure suppression (abstract; Section 4.2). The supporting evidence is visual agreement with a spectral solver (Figures 2, 4-6) and a normalization-versus-time curve (Figure 3). No numerical error metric, such as a relative L2 error, a maximum pointwise error, or a mass-drift tolerance, appears anywhere in the paper. This matters because the PINN loss in Eq. (26) measures PDE residuals at collocation points, not the solution error; small residual values can coexist with large solution errors in under-resolved or poorly conditioned settings. Mass conservation is also not hard-enforced: Section 3.1 explicitly states that no normalization constraint is included, so Figure 3's 'approximately one' needs a quantitative tolerance to support the conservation claim. Furthermore, the 3D comparison in Figure 4 uses only the angle-averaged radial density, which can wash out localized errors. The missing quantitative validation is especially consequential because Section 5 limits the test to short evolution times and a single sinusoidal initial condition; without accuracy numbers even on this simple case, the paper's broader scalability and cosmological-relevance statements are unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces SPINN, a physics-informed neural network solver for the nonlinear Schrödinger-Poisson equations describing fuzzy dark matter, and applies it to gravitational collapse from a sinusoidal perturbation in one and three spatial dimensions. The network outputs the real and imaginary parts of the wavefunction together with the gravitational potential, and is trained by minimizing PDE, initial, and boundary residuals without labeled simulation data, with a spectral method used only for validation. The authors report density evolution, mass conservation, extrapolation beyond the training time, and comparisons with Madelung-formulation, Fourier-potential, and semi-supervised variants. The central claim is that SPINN accurately predicts key metrics and offers a scalable path toward FDM cosmological simulations.","tokens_in":19676,"tokens_out":4008,"duration_ms":43817,"significance":"If the accuracy claims are substantiated, this would be a useful proof-of-concept for applying PINNs to self-gravitating Schrödinger-Poisson systems, a setting where traditional solvers face multi-scale resolution challenges. The paper has several strengths: the unsupervised formulation avoids circularity, since the spectral benchmark enters only in validation and not in the loss; the code is publicly available; and the appendices contain useful ablations (Madelung versus SP formulation, external Poisson solver, activation functions, and hard conditioning). At present, however, the accuracy and scalability claims are not yet supported by quantitative measurements, so the significance is primarily methodological promise rather than demonstrated cosmological capability.","major_comments":[{"comment":"The abstract and Section 4.2 claim 'accurate predictions' of density profiles and structure suppression, but no quantitative error norm (relative L2, maximum pointwise error, or similar) is reported for the 1D or 3D comparisons; the 'error' panels are evaluated only visually. Since Eq. (26) minimizes PDE residuals at collocation points rather than solution error, small loss values do not by themselves establish accuracy. Please add numerical errors for the density (and ideally for R, I, V) at the shown times, including the radial profiles in Figure 4 and the slices/cubes in Figures 5 and 6.","section":"Section 4.2, Figures 2, 4-6"},{"comment":"Mass conservation is claimed on the basis of Figure 3, where the normalization 'remains approximately one,' but no numerical deviations or tolerance are given, and Section 3.1 explicitly states that no normalization constraint is included. Because normalization is not hard-enforced, the conservation statement needs a quantitative bound, such as max |∫|Ψ|² - 1| over the evolution, and should separate the SPINN error from the spectral reference drift. Without that, 'approximately one' is too weak to support the conservation claim in the abstract.","section":"Section 3.1 and Figure 3"},{"comment":"The paper's broader significance rests on efficiency and scalability ('efficient, possibly scalable modeling'), but no wall-clock times, GPU/memory usage, or scaling with domain size or training duration are reported, and Section 5 states that the training time was not long enough and the initial condition not complex enough to assess scalability. The sinusoidal, short-time tests alone cannot support the scalability statements; please either add concrete scaling measurements or explicitly restrict the claims to the studied regime.","section":"Section 5 and Abstract"},{"comment":"The extrapolation claim is supported only by visual inspection of heatmaps; the text says the solutions agree 'uptil t = 4' and then deviate, but no error measure or tolerance is provided. Please quantify the extrapolation error over time (for example, the L2 error relative to the spectral solution in the extrapolated interval) so that the reader can judge how much of the solution is usable beyond the training domain.","section":"Section 4.3, Figure 7"}],"minor_comments":[{"comment":"There are several typos and formatting issues: 'sucessfully' in the Introduction, 'demosntrates' in Section 5, 'T able 1' in the Table 1 caption, 'Inital' in the Section 4.1 heading, and 'The numerical solutions is' in Section 4.2.","section":"Throughout"},{"comment":"The boundary residual for V is written as |Vθ - Vb| without a square, while the R and I terms are squared; this should presumably be |Vθ - Vb|² for consistency with a mean-squared-error loss.","section":"Equation (27)"},{"comment":"The text says the errors are slightly higher 'as shown in Fig. (3),' but Figure 3 is the mass-conservation plot; this cross-reference appears to point to the wrong figure.","section":"Appendix C.2"},{"comment":"The boundary operator in Eq. (14) is written as B[NN(X,θ); ] with an empty argument, which is incomplete and should read B[NN(X,θ); Λ] or similar.","section":"Equation (14)"},{"comment":"The Code Availability section gives only '(DeepWaves-SPINN)' with no URL or DOI; a resolved link should be provided. Also, the first reference entry '2006, Line Search Methods' has no author listed.","section":"References and Code Availability"},{"comment":"The captions describe the time and slice locations as 'randomly chosen,' but randomness is not essential to the validation and the selection procedure is not described; please state how these frames were selected.","section":"Figures 5 and 6"}],"recommendation":"major_revision","confidential_remarks":"I see no concerns about circularity or data provenance; the spectral benchmark is used only for validation. The main issue is that the manuscript's headline claims outrun the quantitative evidence, but this is fixable within the paper's scope by adding error norms and tempering the scalability language. The GitHub link is currently a placeholder and should be resolved before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's real contribution is modest but real: it is the first test of a physics-informed neural network on the Schrödinger-Poisson system for fuzzy dark matter, using no training data. The core demonstration, that a PINN trained on PDE residuals can reproduce the 1D collapse of a sinusoidal perturbation and get the 3D radial density roughly right against a spectral benchmark, looks genuine. I also credit the authors for trying the Madelung formulation, showing why it fails to capture wave interference, and for comparing activation functions and semi-supervised training.\n\nThe soft spots are proportionate to the strength of the claims. The biggest one is that “accurate predictions” is supported only by visual agreement; there is no relative L2 error, no pointwise maximum, no convergence test. The mass conservation plot shows normalization hovering near one, but no tolerance is given, and the 3D comparison uses only the angle-averaged radial density, which can wash out localized errors. The test case is a single sinusoidal initial condition run to t=3, so the evidence is proof-of-concept, not scalability. To their credit, the authors admit this in the conclusion, but the abstract and introduction push a broader efficiency and cosmological-relevance narrative that the paper does not measure.\n\nA minor reproducibility issue: the code link is just a placeholder text, and the extrapolation result is only visual. These are fixable.\n\nDoes the central argument hold up? As a proof-of-concept, yes. The network genuinely solves the SP equations for a simple case, and the benchmark comparison is meaningful. But the paper should not claim accurate, scalable cosmological simulations from this evidence. It needs quantitative error metrics, a convergence study, and either a working code repository or a clear statement that code is available upon request.\n\nI would send this to peer review. It is a first step, not a finished framework, but it is a legitimate one that the FDM community should see, and the requested revisions are clear. A serious referee could make the paper much stronger.","headline":"A first proof-of-concept that PINNs can handle Schrödinger-Poisson for FDM, but the accuracy and scalability claims outrun the quantitative evidence.","tokens_in":20125,"tokens_out":2002,"would_cite":false,"duration_ms":25946,"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 physics-informed neural network called SPINN solves the nonlinear Schrödinger–Poisson equations of fuzzy dark matter in 1D and 3D, reproducing gravitational collapse and conserving total mass without a mesh.","keywords":["fuzzy dark matter","Schrödinger-Poisson equations","physics-informed neural networks","gravitational collapse","cosmological simulations","wave dark matter","mass conservation","spectral methods"],"falsifier":"Train SPINN on the same sinusoidal case but push the comparison to $t = 10$ instead of $t = 3$, and also on a multi-mode initial condition such as a Gaussian perturbation; the paper's own extrapolation result shows divergence beginning near $t = 4$, so measurable growth in the density error or drift of the normalization $\\int |\\Psi|^2\\,dV$ away from 1 over longer times would falsify the claim that the network has learned the physics rather than fitted this single test case.","tokens_in":19286,"feed_emoji":"🌌","tokens_out":11941,"duration_ms":115421,"temperature":0.7,"pith_summary":"This paper sets out to show that a physics-informed neural network can act as a mesh-free solver for the nonlinear Schrödinger–Poisson equations that govern fuzzy dark matter, reproducing gravitational collapse in one and three dimensions. The network, SPINN, is trained directly on the equations plus initial and boundary conditions, with no simulation data in the unsupervised case, and its density evolution matches a second-order spectral scheme while total mass stays conserved without any explicit conservation constraint. The motivation is concrete: conventional fuzzy-dark-matter simulations must resolve the de Broglie wavelength throughout the box, which limits them to volumes of a few megaparsecs, whereas a trained PINN no longer needs the mesh at all. The paper presents SPINN as a first step toward scalable cosmological fuzzy-dark-matter simulation, demonstrating modest time extrapolation and an accuracy gain when a small amount of simulation data is added.","feed_headline":"SPINN neural net solves fuzzy dark matter collapse in 1D and 3D","feed_subtitle":"Physics-informed network matches spectral solvers on dark-matter collapse and conserves mass without a grid.","key_machinery":"The load-bearing object is the network's output choice: SPINN predicts the real part $R$ and imaginary part $I$ of the wavefunction $\\Psi$ along with the gravitational potential $V$ as an auxiliary output, instead of the density and phase used in the Madelung fluid formulation. Writing the Schrödinger–Poisson equations in terms of $R$ and $I$ yields coupled real equations, $\\partial_t R = -\\tfrac{1}{2}\\nabla^2 I + V I$ and $\\partial_t I = \\tfrac{1}{2}\\nabla^2 R - V R$, together with the Poisson constraint $\\nabla^2 V = R^2 + I^2 - 1$; a network that minimizes these residuals respects the complex-conjugate structure of the Schrödinger equation, which is why mass conservation follows without an explicit normalization loss. Predicting $V$ directly sidesteps an FFT-based Poisson solve at each training step, whose $\\mathcal{O}(N^2\\log N)$ cost and memory footprint would otherwise cap resolution. The 1D network uses two hidden layers of 32 neurons with sine activations; the 3D network uses five layers of 16 neurons with a sine first layer and wavelet activations; optimization combines Adam pretraining with L-BFGS, and the loss is a mean-squared sum of PDE, boundary, and initial residuals at randomly sampled collocation points.","core_discovery":"The paper's claim is that a physics-informed neural network can solve the full nonlinear Schrödinger–Poisson system governing fuzzy dark matter directly from the equations, with no training data and no mesh, and reproduce gravitational collapse in both one and three dimensions. With a single sinusoidal Fourier mode as the initial perturbation, a standard test case for Vlasov–Poisson solvers, the network SPINN matches the density evolution computed by a second-order unitary spectral scheme, keeps the wavefunction normalization (total mass) at approximately one with no explicit conservation constraint, and tracks the growth of overdensities in the three-dimensional radial density profile. The paper further claims that the network extrapolates modestly beyond its training time, and that predicting the gravitational potential as a network output avoids the memory and compute cost of solving Poisson's equation by Fourier transform at every training step. The authors present this as a first step toward efficient, possibly scalable cosmological simulation of fuzzy dark matter, explicitly deferring the question of whether the framework scales to realistic initial conditions and long times.","pith_inferences":["The modest extrapolation horizon (accuracy degrading near $t = 4$ after training to $t \\approx 3$) suggests time-window chaining or domain-decomposition, mentioned but not implemented in the paper, as the natural route to long cosmological evolutions; whether the horizon grows with training time is directly testable.","The semi-supervised result points to a concrete pipeline the paper proposes but does not build: take existing large cold-dark-matter simulations as the data constraint for large scales and let the Schrödinger–Poisson residuals paint in the small-scale wave features of fuzzy dark matter.","The failure of the density–phase (Madelung) variant to produce wave splitting is a prediction about any fluid-formulation neural solver: without enforcing the quantization condition that makes the fluid equations equivalent to Schrödinger–Poisson, a network trained on those equations should also miss interference features.","A stricter test than the paper's own: compare the density power spectrum, not just radial profiles, against the spectral solution, so that structure suppression, the key fuzzy-dark-matter observable, is quantified directly."],"forward_implications":["A PINN trained only on equation residuals and initial/boundary conditions reproduces 1D and 3D gravitational collapse of fuzzy dark matter, matching a spectral method, so the basic form of the method needs no labeled simulation data.","Because the network outputs the real and imaginary parts of the wavefunction, wavefunction normalization and hence total mass stay conserved without an explicit constraint term in the loss.","The trained network keeps reasonable accuracy beyond its training window, to about $t = 4$ when trained to $t \\approx 3$, so time extrapolation is possible, not just interpolation.","Adding a small amount of simulation data in semi-supervised training improves the accuracy of the predicted wavefunction phase relative to pure unsupervised training.","Predicting the gravitational potential as a network output avoids per-step Fourier Poisson solves, removing a memory and compute bottleneck that would otherwise force lower resolution."],"supporting_citations":[{"why":"Supplies the 2nd-order unitary spectral scheme used as the numerical benchmark that SPINN predictions are validated against in 1D and 3D.","marker":"Mocz et al. (2017)"},{"why":"Establishes the resolution bottleneck, SP fuzzy-dark-matter simulations limited to about 10 Mpc/h boxes, that motivates seeking mesh-free alternatives.","marker":"May & Springel (2021)"},{"why":"Provides the standard sinusoidal test case the 1D initial condition is chosen to match for comparability.","marker":"Mocz & Szasz (2021)"},{"why":"One of the benchmarks the 1D initial condition is chosen to be comparable with.","marker":"Cappelli et al. (2024)"},{"why":"Defines the physics-informed neural network method of minimizing PDE, boundary, and initial residuals that SPINN is built on.","marker":"Raissi et al. (2019)"},{"why":"Demonstrates that PINNs can extrapolate in time, the capability the paper verifies for SPINN without their specialized training tricks.","marker":"Kim et al. (2020)"},{"why":"Motiveates the sine activation functions adopted to represent oscillatory wave structure and preserve gradient quality.","marker":"Sitzmann et al. (2020)"},{"why":"Proves a quantization condition is required for Madelung fluid equations to match Schrödinger–Poisson, used to explain why the fluid-formulation PINN fails.","marker":"Wallstrom (1994)"},{"why":"Cited for the non-equivalence of the Schrödinger–Poisson and Madelung formulations that drives the choice to train on the wavefunction.","marker":"Zhang et al. (2018)"}],"fun_headline_variants":["AI cracks fuzzy dark matter collapse without a grid","No-mesh neural net models fuzzy dark matter collapse","Physics-informed AI solves dark matter collapse equations","SPINN neural network matches solvers on dark matter collapse"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single sinusoidal initial condition evolved for a short time ($t = 0$ to $3$) faithfully represents the multi-scale, long-duration dynamics of a cosmological fuzzy-dark-matter box; as the paper itself states in Section 5, the training time was not long enough nor the initial condition complex enough to assess scalability, and without that transfer the broader significance of the method collapses to this one test case.","fun_headline_variants_meta":{"raw":{"variants":["AI cracks fuzzy dark matter collapse without a grid","No-mesh neural net models fuzzy dark matter collapse","Physics-informed AI solves dark matter collapse equations","SPINN neural network matches solvers on dark matter collapse"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1440,"prompt_tokens":926,"completion_tokens":514,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":452}},"tokens_in":542,"tokens_out":514,"duration_ms":6390,"temperature":1.0,"reasoning_tokens":452,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:11:17.100648+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SPINN on the same sinusoidal case but push the comparison to $t = 10$ instead of $t = 3$, and also on a multi-mode initial condition such as a Gaussian perturbation; the paper's own extrapolation result shows divergence beginning near $t = 4$, so measurable growth in the density error or drift of the normalization $\\int |\\Psi|^2\\,dV$ away from 1 over longer times would falsify the claim that the network has learned the physics rather than fitted this single test case.","supporting_citations":[{"cited_title":"2021, , 910, 29, 10.3847/1538-4357/abe6ac","cited_arxiv_id":null,"evidence_quote":"Provides the standard sinusoidal test case the 1D initial condition is chosen to match for comparability."},{"cited_title":"2024, Physical Review Research, 6, 013282, 10.1103/PhysRevResearch.6.013282","cited_arxiv_id":null,"evidence_quote":"One of the benchmarks the 1D initial condition is chosen to be comparable with."}],"review_version":1}