{"id":"501a1d29-9aaf-430f-89d5-b55b1bcb1582","arxiv_id":"2412.04021","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A U-Net restores foreground-removed Fourier modes in simulated 21cm intensity maps, preserves BAO reconstruction performance, and transfers from coarse to fine resolutions.","lead":"Using a U-Net neural network, this paper fills in the Fourier modes of simulated 21cm hydrogen maps that were removed because of foreground contamination. The restored maps stay close to the true signal, and BAO reconstruction, which measures cosmic distances, works on them about as well as on complete maps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No-AI baseline missing: BAO reconstruction on the foreground-masked map is never shown, so the abstract's claim that AI restoration 'proves effectiveness' for BAO is unsupported.","rationale":"The reader's weakest_assumption focuses on simulation-to-real transferability. That is a legitimate and self-acknowledged limitation, but it concerns external validity. A more immediate problem is internal: the paper never runs its own BAO pipeline on the foreground-masked map without AI, so the central effectiveness claim is not tested against the very baseline it is meant to beat. The correlation ratio of ~0.9 shows the U-Net can mimic true maps, but the abstract's claim about BAO reconstruction ('minimal impact ... proving effectiveness') needs a control. Without it, even perfect transferability to real data would not establish the claim. The train/test leakage via rotated copies is also a concern for the quantitative correlation numbers, but it is secondary to the missing BAO baseline and could be checked separately. Because the missing baseline is readily addable by rerunning existing code, a conditional verdict is appropriate; the reader's conditional verdict stands, though for a different primary reason than the one listed as weakest_assumption.","tokens_in":19327,"tokens_out":4284,"duration_ms":42222,"concrete_test":"Rerun the BAO reconstruction (§3.3) and MCMC Σ fit (§3.4) on the 'Observed Tb' map (foreground modes set to zero, no U-Net) at 256^3, using the same seed-matched no-wiggle and template pipeline, and add the resulting S_BAO(k) and Σ to Figure 6 panels (a)–(d) alongside the true and AI-restored results. If the observed (no-AI) Σ is not appreciably larger than the AI-restored Σ, the claimed benefit of AI restoration for BAO is not supported; if it is larger, the claim gains the missing control.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim — that U-Net restoration mitigates the impact of foreground contamination on BAO reconstruction — requires a comparison between BAO reconstruction applied to (i) the foreground-masked map without AI ('Observed Tb') and (ii) the AI-restored map. Section 4.2 and Figure 6 provide only true and AI-restored curves; the observed mode-missing field never enters the BAO pipeline. The dashed 'Observed Tb' curves in Figure 5 are correlation ratios only. Consequently, the abstract's statement that AI restoration 'has minimal impact on the performance of the linearized BAO signal' cannot be distinguished from 'the missing modes never mattered' or 'AI restoration neither helps nor hurts.' Panel (a) even processes the no-wiggle spectrum with the same AI/BAO pipeline, which cancels the very distortions that would reveal degradation. Without the no-AI baseline, the paper cannot demonstrate that lost Fourier modes negatively affect BAO reconstruction in this setup, nor that the U-Net recovers the relevant information. This is internal to the paper's own logic and must be settled before any transferability concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a U-Net-based method to restore Fourier modes of 21cm intensity mapping maps that are lost to foreground avoidance (intrinsic foreground k_parallel < 0.1 h/Mpc and the foreground wedge). The training data are COLA simulations with an empirical HI prescription and subgrid modeling, augmented by rotations, with an 80/10/10 train/validation/test split. The authors evaluate the restored maps by visual inspection, power spectra, and the cross-correlation ratio r(k), reporting r ~ 0.9 at k ~ 1 h/Mpc for the finest grids, and they further study the impact on particle-based BAO reconstruction by comparing BAO signatures and fitting the smearing parameter Sigma from an MCMC analysis. The paper also claims that a model trained at 256^3 can be applied to 320^3 with improved correlation, attributing this to scale invariance of nonlinear mode coupling.","tokens_in":19604,"tokens_out":9971,"duration_ms":88084,"significance":"The potential significance is real: if the method robustly recovers the information needed for BAO reconstruction, it would offer a practical way to use foreground-avoided 21cm data without discarding the wedge. The paper has several strengths: it uses a held-out test set, presents quantitative 2D and 1D correlation metrics, applies a realistic BAO reconstruction pipeline, and explicitly discusses limitations of the HI modeling and the absence of beam/noise/residual foregrounds in the training. However, the central claim regarding 'mitigating the impact of foreground contamination' on BAO reconstruction is not supported by the current analysis because the BAO pipeline is never applied to the foreground-masked map without AI restoration. The scale-transfer claim is also partially confounded by the two-stage retraining. These issues are addressable in revision, so the paper warrants major revision rather than rejection.","major_comments":[{"comment":"The central claim that AI restoration 'mitigates the impact of foreground contamination' on BAO reconstruction requires a comparison between BAO reconstruction applied to the foreground-masked map (Observed Tb) and the AI-restored map. None of the panels in Figure 6 shows the observed mode-missing field entering the BAO pipeline; dashed 'Observed Tb' curves appear only in the correlation-ratio plots of Figure 5. Without this no-AI baseline, the result cannot distinguish 'the missing modes never mattered' from 'AI restoration helps' or even 'AI restoration hurts.' Please run the same particle-based BAO reconstruction and Sigma-fitting on the mode-missing observed map and present it alongside the true and AI-restored results.","section":"Section 4.2, Figure 6"},{"comment":"The fitted smearing parameter is quoted as Sigma = 8.33 without BAO reconstruction and Sigma = 3.20 with BAO reconstruction, but the text states that the one-sigma error on Sigma is quite large, and the corner plots show non-Gaussianity. The claim that BAO reconstruction reduces nonlinear smearing and that AI restoration has minimal impact needs quantitative error bars on Sigma, ideally from the MCMC chains, so the reader can judge whether the differences are significant.","section":"Section 4.2, Figure 6(d) and Appendix A"},{"comment":"The chi-squared fitting is said to use 'the diagonal components of the covariance matrix ... computed using our original 37 simulated training samples.' If the field used for the BAO analysis (e.g., the 256^3 field in Figure 6) is part of the same 37 realizations, the covariance is not independent of the data being fit. Please clarify whether the covariance is computed from the training split only, and if not, recompute it excluding the test field.","section":"Section 3.4, covariance estimate"},{"comment":"The abstract's claim that a model trained on coarser fields can be effectively applied to finer fields is partially confounded by the two-stage training procedure. The 256^3 and 320^3 results are obtained with the stage-two model retrained at 256^3, not with the original stage-one model trained at 128^3, and the text notes that the stage-one model showed amplitude deviations on higher-resolution maps. Please show explicitly which model is applied to each resolution, and if possible include the stage-one model's performance at 256^3 to separate retraining gains from genuine scale transfer.","section":"Sections 3.2 and 4.1, Figures 2 and 5"}],"minor_comments":[{"comment":"'Proving the effectiveness' is too strong for an idealized simulation with no beam, noise, or residual foregrounds; consider 'demonstrating' or 'indicating'.","section":"Abstract and Section 5"},{"comment":"The choice theta_FOV = pi/2 covers an extremely wide primary beam; please justify this value or discuss its effect on the size of the wedge and the difficulty of the restoration task.","section":"Section 2, Eq. (2.3)"},{"comment":"'Viralized' should be 'virialized'.","section":"Section 3.1, Eq. (3.2)"},{"comment":"'Cloud-in-Cloud algorithm' is likely a typo for 'Cloud-in-Cell' (CIC) mass assignment.","section":"Section 3.3"},{"comment":"The red solid line is described as showing that BAO reconstruction alone introduces significantly greater shape deviations; the text also says the sky-blue line in panel (c) is identical to panel (b), which is not immediately clear from the figure. Please make the repeated curves explicit in the caption.","section":"Section 4.2, Figure 6(c) caption"},{"comment":"References [39] and [73] are the same paper (Schmittfull, Baldauf, Zaldarriaga 2017); please merge or cite once.","section":"References [39] and [73]"},{"comment":"The 'Observed Tb(mode missing)' entries are repeated four times; simplify the legend to avoid clutter.","section":"Figure 5 legend"},{"comment":"'There are many systematic effects could potential affect' should be '...could potentially affect'.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The missing no-AI BAO baseline is the key blocker; the paper should be sent back for major revision. The scale-invariance claim also needs the training confound addressed. The paper is otherwise within the scope of JCAP and the technical content is largely sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid simulation-only demonstration of U-Net-based recovery of foreground-wedge modes, and the coarse-to-fine transfer result is genuinely new and worth taking seriously. But the headline claim that AI restoration mitigates the impact of foreground contamination on BAO reconstruction is not supported by the analysis as presented, because the observed mode-missing map is never run through BAO reconstruction.\n\nWhat the paper does well: the U-Net architecture is sensible, the two-stage coarse-to-fine training is a pragmatic response to GPU limits, and the cross-correlation ratios (Figure 5) show the restored fields track the true fields well, especially at finer resolution. The scale-invariance argument is plausible and the transfer result is the most interesting part. The BAO extraction and fitting follow standard practice, and the limitations section is candid about HI modeling, subgrid treatment, RSD, and missing noise/beam effects.\n\nThe soft spots: the abstract says subsequent BAO reconstruction 'proves the effectiveness' of the machine learning approach to mitigate foreground impact. But there is no curve in Figure 6 for BAO reconstruction applied to the observed (mode-missing) map. The correct baseline is: run reconstruction on the observed map without AI, show the BAO signal is degraded, then show the AI-restored map recovers it. Instead, the paper only shows the AI-restored map behaves like the true map, which is consistent with either 'AI helps' or 'the missing modes never mattered for this BAO algorithm.' Panel (a) normalizes both BAO and no-wiggle spectra through the same AI/BAO pipeline, which cancels the very distortions that would reveal degradation. This is a load-bearing gap for the paper's main claim.\n\nSeparately, the train/test split may leak: rotating each of 37 realizations gives 148 maps, and if the split is done after rotation, rotated copies of the same realization can appear in both training and test. That would inflate the correlation numbers. The split should be at the realization level. Also, the correlation ratios have no error bars, and no code or data is released.\n\nFor whom is this? Read it for the coarse-to-fine transfer result and for the practical pipeline question of whether to inpaint wedge modes before BAO reconstruction. The paper deserves a serious referee, but only after adding the no-AI baseline and clarifying the split. I'd send it to review with a request for major revision rather than desk-reject.","headline":"Useful U-Net mode-restoration study, but the abstract's BAO claim lacks the no-AI baseline needed to support it.","tokens_in":20152,"tokens_out":3262,"would_cite":false,"duration_ms":28124,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that a U-Net trained on simulated 21cm intensity maps can restore Fourier modes lost to foreground contamination, reaching about 0.9 cross-correlation with the true field at $k \\sim 1\\,h/\\mathrm{Mpc}$, and that this…","keywords":["21cm intensity mapping","foreground wedge","mode recovery","deep learning","U-Net","baryon acoustic oscillations","BAO reconstruction","scale invariance"],"falsifier":"Feed an observed or hydrodynamically simulated 21cm datacube that includes beam, thermal noise, and residual foregrounds through the trained U-Net, mask the same wedge, and compare the restored modes against independently measured modes from a cross-correlation tracer; if the cross-correlation ratio in the wedge falls well below 0.9 or the recovered BAO peak shifts by more than the statistical error, the central claim fails.","tokens_in":1744,"feed_emoji":"📡","tokens_out":1822,"duration_ms":77121,"temperature":0.7,"pith_summary":"The paper addresses a practical obstacle in 21cm intensity mapping: foreground contamination removes a wedge-shaped region of Fourier modes, and those missing modes degrade the BAO reconstruction algorithms that would 'linearize' the acoustic signal. It proposes training a U-Net deep-learning model on simulated 21cm temperature cubes to regenerate the missing modes from the surviving modes, exploiting the non-linear mode coupling of structure formation. The paper reports that the AI-restored maps reach a cross-correlation ratio of roughly 0.9 with the true signal at $k \\sim 1\\,h/\\mathrm{Mpc}$, and that running BAO reconstruction on these restored maps yields a linearized BAO signal very close to that obtained with all modes present. It also finds that a model trained on coarse $128^3$ grids transfers successfully to finer $256^3$ and $320^3$ maps, with even higher correlation, a behavior the authors attribute to the scale invariance of non-linear mode coupling and U-Net's hierarchical architecture. A sympathetic reader would care because the method offers a way to salvage the information lost to the foreground wedge before performing precision cosmology with BAO.","feed_headline":"U-Net restores 21cm modes; BAO signal survives","feed_subtitle":"AI restores wedge-lost 21cm modes to ~0.9 correlation; BAO reconstruction is largely unchanged.","key_machinery":"The carrying mechanism is non-linear gravitational mode coupling: structure formation mixes Fourier modes, so the observed modes outside the foreground wedge contain information about the modes that are missing. The U-Net, a fully convolutional encoder-decoder with skip connections, is trained in configuration space, where the coupling is more local than in Fourier space, to map the masked temperature cube to the full temperature cube using a mean-squared-error loss; it is the instrument that learns that coupling. The BAO reconstruction is a particle-based iterative solver of the continuity equation that estimates the displacement field by successive Zel'dovich steps and produces a reconstructed density field whose BAO signal is closer to linear theory. The diagnostics are the cross-correlation ratio $r(k)$ between restored and true fields and the fitted smearing parameter $\\Sigma$, which measures how much of the BAO signal remains damped.","core_discovery":"The central claim is that missing Fourier modes caused by the intrinsic foreground and the foreground wedge in 21cm intensity mapping can be recovered by a U-Net trained in configuration space, and that the recovered field is faithful enough for BAO reconstruction to perform essentially as well as if no modes were missing. Quantitatively, the cross-correlation ratio between the AI-restored temperature field and the true field is about 0.9 at $k \\sim 1\\,h/\\mathrm{Mpc}$, exceeding the scale range where the BAO reconstruction algorithm is effective. After applying a particle-based iterative BAO reconstruction, the BAO signal extracted from the AI-restored map closely tracks the signal from the full-mode map, and the fitted nonlinear-smearing parameter $\\Sigma$ is reduced from about 8.33 without reconstruction to about 3.20 with it, showing that reconstruction still sharpens the BAO. A further claim is scale transfer: a U-Net trained on coarser grids generalizes to finer grids, with correlation improving at higher resolution, consistent with the scale invariance of the mode-coupling kernels of perturbation theory.","pith_inferences":["If the observed scale invariance holds more broadly, the same network could be applied across surveys with different pixel scales or at even finer resolution without retraining, potentially easing the demand for very large high-resolution simulations.","The same mode-restoration logic could be transferred to other Fourier-space data-loss problems in 21cm cosmology, such as the Epoch of Reionization window or radio frequency interference excision, where non-linear mode coupling also entangles lost and observed modes.","A direct observational test would be to compare the network's restored wedge modes with modes measured by an overlapping galaxy redshift survey; a high cross-correlation would validate phase recovery on real data rather than simulated data.","One could go beyond the paper's diagonal-covariance fit and use the restored maps to estimate a sound-horizon distance error bar, quantifying how much the AI restoration tightens BAO constraints relative to foreground avoidance alone."],"forward_implications":["Foreground avoidance plus U-Net restoration recovers the Fourier modes inside the wedge well enough that subsequent BAO reconstruction behaves like a full-mode map.","A U-Net trained at one resolution transfers to finer grids, so low-resolution simulations can serve higher-resolution observations, reducing the computational cost of training data.","The restored-field correlation ratio reaches about 0.9 at $k \\sim 1\\,h/\\mathrm{Mpc}$, beyond the roughly $0.6\\,h/\\mathrm{Mpc}$ range where BAO reconstruction is effective, so the restored modes are usable for BAO analysis.","AI restoration introduces only mild shape distortions, which the polynomial transfer function in the BAO template absorbs; BAO reconstruction itself introduces larger shape distortion.","Training on noiseless maps with a perfectly known mask means that applying the method to real observations will require retraining with beam, noise, and residual foregrounds included."],"supporting_citations":[{"why":"Supplies the halo mass-HI mass fitting recipe and redshift-dependent parameters used to generate the 21cm brightness temperature cubes.","marker":"[16]"},{"why":"Provides the COLA simulation code used to generate the dark matter distributions that seed the training data.","marker":"[67]"},{"why":"The U-Net architecture, with convolutional encoder-decoder and skip connections, is the model family used for mode restoration.","marker":"[70, 71]"},{"why":"Introduced the Zel'dovich-approximation BAO reconstruction algorithm that the paper's particle-based iterative scheme builds on.","marker":"[35]"},{"why":"Provides the iterative initial-condition reconstruction code whose particle-based scheme the paper adopts and compares with.","marker":"[73]"},{"why":"Quantifies how foreground-wedge mode loss degrades BAO distance measurements, motivating the recovery approach.","marker":"[43]"},{"why":"Supplies the standard template-fitting approach, including the no-wiggle template, polynomial transfer function, and damping parameter, used to extract the BAO signal.","marker":"[74]"},{"why":"Gives the perturbation-theory mode-coupling coefficients whose scale invariance the paper invokes to explain U-Net's transfer across resolutions.","marker":"[76]"}],"fun_headline_variants":["U-Net recovers lost 21cm modes, BAO reconstruction intact","AI fills wedge holes in 21cm maps, BAO signal unharmed","Deep learning restores 21cm modes, preserves BAO reconstruction","U-Net recovers wedge-lost 21cm modes with 0.9 fidelity","Scale-invariant AI fixes 21cm missing modes, BAO unaffected"],"cache_read_input_tokens":22272,"weakest_assumption_plain":"The whole result rests on the assumption that the simulated 21cm maps produced by COLA simulations plus the empirical HI prescription reproduce the real non-linear mode coupling of the 21cm field closely enough that the values learned on them transfer to observations.","fun_headline_variants_meta":{"raw":{"variants":["U-Net recovers lost 21cm modes, BAO reconstruction intact","AI fills wedge holes in 21cm maps, BAO signal unharmed","Deep learning restores 21cm modes, preserves BAO reconstruction","U-Net recovers wedge-lost 21cm modes with 0.9 fidelity","Scale-invariant AI fixes 21cm missing modes, BAO unaffected"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00056,"raw_usage":{"total_tokens":2708,"prompt_tokens":1038,"completion_tokens":1670,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":1579}},"tokens_in":654,"tokens_out":1670,"duration_ms":10251,"temperature":1.0,"reasoning_tokens":1579,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:51:13.583941+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed an observed or hydrodynamically simulated 21cm datacube that includes beam, thermal noise, and residual foregrounds through the trained U-Net, mask the same wedge, and compare the restored modes against independently measured modes from a cross-correlation tracer; if the cross-correlation ratio in the wedge falls well below 0.9 or the recovered BAO peak shifts by more than the statistical error, the central claim fails.","supporting_citations":[{"cited_title":"Villaescusa-Navarro, S","cited_arxiv_id":null,"evidence_quote":"Supplies the halo mass-HI mass fitting recipe and redshift-dependent parameters used to generate the 21cm brightness temperature cubes."},{"cited_title":"Eisenstein, H","cited_arxiv_id":null,"evidence_quote":"Introduced the Zel'dovich-approximation BAO reconstruction algorithm that the paper's particle-based iterative scheme builds on."},{"cited_title":"Seo and C.M","cited_arxiv_id":null,"evidence_quote":"Quantifies how foreground-wedge mode loss degrades BAO distance measurements, motivating the recovery approach."}],"review_version":1}