{"id":"1c02c0ae-8434-4a67-902b-53c10c5ab0e9","arxiv_id":"2607.28604","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Physics-informed generative U-Nets evolve and super-resolve fuzzy dark matter fields under Schrödinger–Poisson constraints with far less supervised data than pure data-driven baselines.","lead":"A U-Net trained with Schrödinger–Poisson residuals can evolve and super-resolve fuzzy-dark-matter fields in a tiny box while cutting generative artifacts. It matters because full wave dark-matter simulations are extremely expensive, so physically consistent emulators could unlock larger statistical studies.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Physics loss only enforces local SP/KDK consistency inside thin target-a windows; long-range IC→field maps and cross-IC phase/small-scale fidelity stay under-constrained.","rationale":"The reader’s weakest assumption is exactly the load-bearing joint: narrow a-windows + residual SP/KDK + MSE on R/I/V are assumed to fix an ill-posed map (including phase) well enough for the abstract’s broader emulator claims. Full-text evidence supports useful, scoped gains—physics loss lowers residuals and helps sparse-label interpolation inside the trained window (Fig. 2); SR on full-SP cubes with PS tests is new (Figs. 5–7)—but also documents the soft spot (App. B phase degeneracy and high-k failure; App. A edge degradation; §5.1 IC caveats; tiny L and Δa). No stronger internal contradiction appears (e.g. no mis-derived SP residual). CONDITIONAL with tightened generalization wording remains the right verdict; this pass sharpens the same concern rather than replacing it. Concrete test isolates whether L_physics constrains dynamics off the supervised a-slice, which is what would have to be true for “evolution from ICs with 20% data” to mean more than local regularization at the target epoch.","tokens_in":22256,"tokens_out":775,"duration_ms":57814,"concrete_test":"Retrain the evolution model with L_data unchanged (20% labels only in a∈[0.149,0.157]) but evaluate L_physics on a disjoint intermediate band (e.g. a∈[0.08,0.10]) never used for supervision; report SP residuals, projected-field errors, and P(k) (especially k≳k_J) on that band and on App. B unseen ICs. If errors rise materially above Figs. 2, 4, 8 levels, the physics term is not supplying the long-range/phase constraint the central claim needs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim—that an explicit SP (or linearized KDK) physics loss lets a generative U-Net accurately evolve FDM fields from ICs with only ~20% temporal labels and generalize across realizations—depends on L_physics actually pinning the dynamical map, including wavefunction phase. In the paper it does not do that over the claimed domain. Per §3.2–3.3 and Eqs. (37)–(41), both data and physics losses are evaluated only inside extremely narrow scale-factor intervals (evolution: a∈[0.149,0.157]; SR: a∈[0.0798,0.0803]). Evolution is a direct residual map from z=127 ICs to fields near a≈0.15 (Eq. 45 / Fig. 4), while SP/KDK residuals are formed from generator queries at a and a+Δa drawn solely from that same thin target window—not along the path from a0. Appendix B then states that multi-realization training recovers large-scale morphology but fails on high-k power, and that “the physics-informed loss alone admits multiple solutions that satisfy the governing equations but differ in phase.” Abstract/§6 language (“arbitrary cosmological scale factor,” “generalizes effectively,” first full-SP generative SR as a path to large-volume emulators) therefore rests on local regularization plus MSE inside a 1 h⁻¹ Mpc, 64³/128³, Δa≲0.01 regime, not on demonstrated SP-constrained forecasting or phase-correct transfer.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript introduces Cosmo-SPINN, a U-Net generative framework with a Schrödinger–Poisson (SP) physics-informed loss for fuzzy dark matter in a 1 h⁻¹ Mpc box. It addresses (i) mapping initial conditions at z=127 to fields near a target scale factor and (ii) 32³→128³ super-resolution at fixed epoch. Physics losses are either SP PDE residuals (Eqs. 37–40) or a first-order kick–drift–kick residual (Eqs. 41–42), combined with MSE on normalized (R,I,V). For evolution, adding the SP loss after a data-only warm-up improves residuals and data fidelity when only ~20% of snapshots in a∈[0.149,0.157] are labeled. For SR, models trained on full-SP Jaxion simulations recover power spectra to ~100 h Mpc⁻¹ with mean fractional errors typically ≲20% on held-out realizations, with SP vs KDK ablations and data-only baselines reported.","tokens_in":22605,"tokens_out":1553,"duration_ms":29401,"significance":"If the claims hold in the demonstrated regime, this is a useful step for FDM emulation: full-SP-trained generative SR is new relative to prior FDM GANs that did not use full quantum-pressure simulations, and the evolution results show clear data-efficiency gains from an explicit SP residual versus data-only training (Figs. 2–3, 5–7). Public PyTorch code and systematic SP/KDK ablations with power-spectrum scatter are strengths. The work is relevant to Epoch-of-Reionization-scale FDM studies where SP simulations are expensive. Significance is currently limited by the narrow scale-factor windows, small box, and incomplete phase/small-scale transfer across realizations, so the path to large-volume emulators remains prospective rather than demonstrated.","major_comments":[{"comment":"Abstract and §1/§6 claim evolution “to an arbitrary cosmological scale factor” and effective generalization across unseen ICs. Training and physics losses are confined to a∈[0.149,0.157] (evolution) and a∈[0.0798,0.0803] (SR) (§3.2; Eqs. 37–41). The evolution map is a residual from z=127 ICs to fields inside that thin window (Eq. 45), with SP/KDK residuals formed only at a and a+Δa drawn from the same window—not along the IC→target path. Appendix A already shows degraded power-spectrum error at interval edges. Please revise abstract/conclusion language to the demonstrated Δa≲0.01 regime, or add experiments that sample physics residuals and targets over a substantially broader a range.","section":"Abstract; §3.2–3.3; Eqs. (37)–(45); Appendix A"},{"comment":"Appendix B states that multi-realization evolution recovers large-scale morphology but fails on high-k modes, and that “the physics-informed loss alone admits multiple solutions that satisfy the governing equations but differ in phase.” This undercuts the Abstract claim that the framework “generalizes effectively across previously unseen realizations.” The single-realization 20%-snapshot result (Fig. 4) should not be conflated with cross-IC generalization. Either qualify the generalization claim to large-scale morphology only, or strengthen phase constraints (e.g., explicit phase/velocity losses, denser temporal supervision) and report quantitative high-k metrics for multi-IC evolution comparable to Fig. 7.","section":"Abstract; Appendix B; Fig. 9"},{"comment":"For SR, ensemble statistics (Fig. 7) show the SP-residual model has larger high-k bias and realization scatter than KDK, while selected examples (Figs. 5–6) favor SP at high-k. §4.2.2 and §6 still lean toward SP as improving small-scale fidelity. Please reconcile example-level vs ensemble conclusions explicitly (e.g., preferred loss by metric and k-range), and state which checkpoint/selection rule (min physics test loss) is recommended for downstream use so the central SR claim is not overstated.","section":"§4.2.2; Figs. 5–7; §6"},{"comment":"§5.1 acknowledges that particle-based 2LPT/CIC ICs are suboptimal for field-based FDM (mesh error; CDM-suited velocities) yet all results use this setup. Given that phase learning is already identified as the weak point (Appendix B), the IC choice is load-bearing for claimed physical consistency. At minimum, quantify sensitivity (e.g., one comparison with field-consistent ICs as in Luu et al. 2025) or clearly limit claims to “standard particle-initialized SP suites” rather than general FDM dynamics.","section":"§5.1; §3.1"}],"minor_comments":[{"comment":"Equation numbering is inconsistent or duplicated in places (e.g., Poisson/mean density and timestep constraints reuse numbers such as (16), (34) in §2). Please renumber uniquely through the manuscript.","section":"§2.1–2.2"},{"comment":"Fig. 2 caption and §4.1 refer to a∈[0.0798,0.0803] when discussing the evolution model’s snapshot count; evolution training is a∈[0.149,0.157] (§3.2). Correct the interval in the evolution discussion.","section":"§4.1; Fig. 2"},{"comment":"Title and running text alternate “F uzzy” / “Schr¨ odinger” spacing artifacts from LaTeX; clean typography throughout.","section":"Title; §2"},{"comment":"State explicitly whether test realizations in Figs. 5–6 are drawn from the 20% held-out IC set and whether any hyperparameter (α_max, λ_w, residual scale 0.1) was tuned on test data.","section":"§3.4; §4.2"},{"comment":"Cite and briefly contrast related physics-informed or operator-learning cosmology emulators beyond the FDM GAN of Sipp et al. (2023) to clarify novelty boundaries.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"The technical core (baselines, ablations, code) is real and publishable after claim-scope correction. The main risk is abstract/§6 overreach relative to Δa≲0.01, 1 h⁻¹ Mpc, and Appendix B phase failure. I would not reject on novelty grounds: full-SP FDM generative SR appears new. Fit for ApJ is reasonable if claims are narrowed; if the authors refuse to qualify “arbitrary scale factor” and cross-IC generalization, the paper would not meet the journal’s standard for supported conclusions."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a real methods step for wave-DM ML, not a cosmology rewrite. They put a conditional 3D U-Net on full Schrödinger–Poisson Jaxion cubes and add either SP residuals or a linearized KDK step as physics loss. The new piece that matters is generative super-resolution trained on actual full-SP FDM (not density-only / no-quantum-pressure proxies), with a clean SP-vs-KDK ablation and power-spectrum errors plus scatter on held-out realizations.\n\nWhat works: data-only baselines, two-stage training, and the evolution result that SP loss acts as a regularizer when only ~20% of temporal snapshots are labeled. Figures 2–7 support the local claims—physics residuals drop, SR recovers power to ~100 h Mpc⁻¹ with mean fractional errors often ~20%, SP sharper / noisier, KDK smoother / more stable. Code is public; citations to Sipp, CDM SR, and their SPINN paper are in the right places. No circular “rediscover Ωm” game—targets are independent pseudo-spectral runs.\n\nSoft spots, sized honestly: the stress-test mostly lands. Both data and physics losses live in very thin a intervals (evolution ~[0.149,0.157], SR ~[0.08]), and evolution is a residual map from z=127 ICs to near a≈0.15, not SP-constrained integration along the path. Appendix B already says multi-realization training gets large-scale morphology but loses high-k and that physics loss alone allows phase-degenerate solutions. So “arbitrary scale factor” and “generalizes effectively” in the abstract are ahead of the evidence. Box is 1 h⁻¹ Mpc at 64³/128³; particle-based ICs are conventional but imperfect, as they note. Free knobs (α ramp, λ_w, residual scale) are normal for this genre, not a hidden fit to cosmology.\n\nWho it’s for: people building FDM emulators or physics-informed generative SR. Worth a serious referee if claims are scoped to the demonstrated regime and the generalization wording is tightened. I’d engage the SR half; treat the long-range evolution story as provisional until they show physics along the trajectory and better phase control.","headline":"Useful first full-SP FDM generative SR plus a data-efficient evolution U-Net, but the physics loss only regularizes thin a-windows in a 1 h⁻¹ Mpc box—abstract generalization language outruns Appendix B.","tokens_in":23308,"tokens_out":587,"would_cite":true,"duration_ms":19214,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A physics-informed U-Net evolves and super-resolves fuzzy dark matter fields by enforcing Schrödinger–Poisson dynamics, matching full simulations from sparse training data.","keywords":["fuzzy dark matter","Schrödinger–Poisson","physics-informed neural networks","generative U-Net","super-resolution","cosmological simulations","wave dark matter"],"falsifier":"Train the same architecture with and without the physics loss on a held-out set of full Schrödinger–Poisson runs at a later redshift or larger box; if the physics-informed model no longer recovers the target power spectrum to within ~20 percent up to the Jeans scale, or if phase errors grow without bound, the central claim fails.","tokens_in":23024,"feed_emoji":"🌌","tokens_out":906,"duration_ms":16578,"temperature":0.7,"pith_summary":"Fuzzy dark matter is expensive to simulate because the wave equation must be resolved everywhere, so boxes stay small and high-resolution runs are scarce. This paper shows that a generative U-Net can do two jobs at once—evolve fields from initial conditions to a later scale factor, and super-resolve low-resolution cubes—if training adds an explicit loss that penalizes violations of the Schrödinger–Poisson equations. With that physics loss, the evolution model matches target 1 h⁻¹ Mpc simulations using only about 20 percent of the temporal snapshots and still generalizes to new initial conditions. For super-resolution, the same idea yields the first generative models trained on full Schrödinger–Poisson fuzzy-dark-matter runs, recovering missing small-scale power while cutting the usual generative artifacts. The practical claim is that modern generative modeling can stay faithful to wave dark-matter physics and therefore reduce both storage and compute for the simulations cosmologists need.","feed_headline":"Physics loss lets AI evolve and sharpen fuzzy dark matter","feed_subtitle":"A U-Net enforcing Schrödinger–Poisson dynamics matches full simulations from only 20% of the snapshots","key_machinery":"Cosmo-SPINN: a residual U-Net that outputs the real and imaginary parts of the wavefunction plus the gravitational potential, trained on the sum of a data MSE and a weighted physics loss built from Schrödinger–Poisson residuals (or a linearized kick–drift–kick step).","core_discovery":"Including an explicit Schrödinger–Poisson residual (or a first-order kick–drift–kick consistency) loss inside a generative U-Net substantially improves both the evolution and the super-resolution of fuzzy-dark-matter fields relative to data-only training, so that a 1 h⁻¹ Mpc box can be reproduced accurately from only ~20 percent of the temporal snapshots and, for the first time, generative super-resolution can be trained directly on full Schrödinger–Poisson simulations.","pith_inferences":["The same residual-plus-physics pattern should transfer to mixed cold-plus-fuzzy dark-matter models where only a fraction of the matter is ultralight.","Autoregressive or neural-operator versions of the same loss could remove the need for any high-resolution supervised targets, provided phase degeneracy is controlled.","If the method scales, observational forecasts that currently rely on a handful of expensive fuzzy-dark-matter boxes can be rerun with large ensembles of emulated realizations."],"forward_implications":["Dense intermediate snapshots of fuzzy-dark-matter simulations need not be stored once an evolution emulator is trained.","Low-resolution Schrödinger–Poisson runs can be upscaled to high resolution while remaining consistent with the governing equations.","Physics-informed generative models become a practical route to larger-volume or longer-time fuzzy-dark-matter statistics that pure spectral solvers cannot yet reach.","Comparing residual versus kick–drift–kick physics losses gives a concrete design choice for future wave-dark-matter emulators."],"fun_headline_variants":["Physics-informed U-Net evolves fuzzy dark matter from 20% data","SP residual loss sharpens FDM evolution and super-resolution","Generative U-Net enforces Schrödinger-Poisson for FDM fields","Physics loss cuts artifacts in fuzzy dark matter emulators","FDM super-resolution trained on full SP simulations"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That residual physics losses evaluated only on narrow windows of scale factor, together with ordinary field matching, are enough to fix the wavefunction phase and make the method work beyond the small boxes and short intervals used in the tests.","fun_headline_variants_meta":{"raw":{"variants":["Physics-informed U-Net evolves fuzzy dark matter from 20% data","SP residual loss sharpens FDM evolution and super-resolution","Generative U-Net enforces Schrödinger-Poisson for FDM fields","Physics loss cuts artifacts in fuzzy dark matter emulators","FDM super-resolution trained on full SP simulations"]},"model":"grok-4.5","effort":"low","cost_usd":0.003448,"raw_usage":{"total_tokens":1177,"prompt_tokens":849,"num_sources_used":0,"completion_tokens":70,"cost_in_usd_ticks":34484000,"prompt_tokens_details":{"text_tokens":849,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":258,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":849,"tokens_out":70,"duration_ms":4755,"temperature":1.0,"reasoning_tokens":258,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T02:24:48.612634+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the same architecture with and without the physics loss on a held-out set of full Schrödinger–Poisson runs at a later redshift or larger box; if the physics-informed model no longer recovers the target power spectrum to within ~20 percent up to the Jeans scale, or if phase errors grow without bound, the central claim fails.","supporting_citations":[],"review_version":1}