{"id":"c56b7664-6300-442d-8e76-0c4582563abe","arxiv_id":"2501.14055","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An invertible recurrent inference machine recovers simulated Sgr A* structure down to about 5 microarcseconds from scattered images, below the roughly 24 microarcsecond nominal resolution of the Event Horizon Telescope.","lead":"A machine learning model trained on synthetic images can strip away the blur and noise that interstellar gas imprints on radio images of the black hole at our galaxy's center, recovering simulated detail at 5 microarcseconds. If it transfers to real observations, interstellar scattering would no longer be the hard resolution limit for Event Horizon Telescope imaging of Sgr A*.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 5 µas recovery claim is demonstrated only on fully sampled, noise-free scattered images; real EHT (u,v) coverage and thermal noise are deferred, so the claim does not yet apply to actual EHT observations.","rationale":"The paper is a clean proof-of-concept: an IRIM trained on phenomenological Kolmogorov Gaussians mitigates simulated interstellar scattering on out-of-distribution GRMHD images, maintaining Stokes-averaged NXCORR above 0.95 down to about 4–5 µas. The GRMHD tests are genuinely independent on the source side, and the authors are transparent in Section 5 about deferred complications. The most load-bearing gap is not the scattering spectrum index, though that is a real concern; it is the observation operator. The evaluation uses fully sampled, noise-free scattered images, meaning the model has access to spatial frequencies that EHT cannot measure at 1.3 mm. The abstract's 'well below the nominal instrumental resolution of EHT' framing therefore overstates what has been demonstrated: the experiment shows scattering mitigation on full-resolution synthetic images, not on EHT-like data. If real EHT (u,v) coverage and thermal noise are inserted, the effective resolution of the descattered product could degrade substantially, because the information needed to reconstruct 5 µas structure is not present in the measured visibilities. This does not invalidate the idealized capability claim, but it does mean the practical conclusion in Section 5 is conditional on a pipeline that has not yet been tested. The reader's CONDITIONAL verdict is therefore appropriate, and no verdict change is needed.","tokens_in":12813,"tokens_out":6425,"duration_ms":63853,"concrete_test":"Repeat the Section 4.3 GRMHD evaluation with a realistic EHT observation operator: sample the scattered Stokes maps onto the April 2017 EHT (u,v) coverage, add thermal noise at the achieved sensitivities, reconstruct images with the same imaging pipeline (e.g., eht-imaging) used to produce EHT images, and then apply the trained IRIM descattering model. Recompute the Stokes-averaged NXCORR threshold of Figure 7. If the scale at which NXCORR drops below 0.95 moves to ~12–24 µas or above, the 5 µas claim does not transfer to real EHT observations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 constructs 'observed' images with complete (u,v) coverage, and Section 4.3 evaluates IRIM on those fully sampled scattered GRMHD images. The abstract's 5 µas claim therefore compares descattered truth with descattered estimate at pixel scales far finer than EHT can measure: 24 µas corresponds to roughly 10 Gλ at 1.3 mm, while 5 µas corresponds to baselines several times the Earth's diameter. The model may be exploiting spatial frequencies that simply do not exist in EHT data. Section 5 explicitly defers instrument resolution, thermal noise, and imaging specifics, so the practical conclusion that scattering is not a substantial impediment for ground-based VLBI is not supported by the experiments shown. The scattering-index mismatch (α=1.38 vs 5/3) is a real but secondary uncertainty; the missing observation operator is a difference in the forward model itself, not just a parameter value.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using an invertible recurrent inference machine (IRIM) to mitigate interstellar scattering of Sgr A* at 1.3 mm. The model is trained on synthetic 'Kolmogorov Gaussian' images with a simulated scattering screen and then evaluated on GRMHD simulation images. The central quantitative claim is that the Stokes-averaged NXCORR between the descattered estimate and the truth exceeds 0.95 for all angular scales down to about 4 microarcseconds, well below the EHT nominal resolution of 24 microarcseconds. The authors conclude that sufficient information exists in scattered images to remove both diffractive and refractive scattering at resolutions relevant for ground-based VLBI.","tokens_in":12864,"tokens_out":6477,"duration_ms":58786,"significance":"If the result held for realistic EHT observations, it would be a valuable advance for high-resolution imaging of Sgr A*. The paper has clear strengths: the GRMHD test set is held out and out-of-distribution relative to the training set, the use of non-birefringence to couple the four Stokes maps is physically motivated, and the quantitative comparison against a deblurring baseline is a useful benchmark. However, the missing interferometric observation operator means the experiments support only an idealized, fully sampled image-domain proof of concept, not the abstract's practical conclusion about EHT resolution. With additional experiments or a more carefully scoped claim, the work would be suitable for publication.","major_comments":[{"comment":"The 4–5 microarcsecond recovery claim is established only on fully sampled, noise-free scattered images. The 'observed' images in Section 3.2 are generated with complete (u,v)-coverage, and Section 4.3 evaluates NXCORR on those fully sampled images. Real EHT data consist of sparse visibilities with thermal noise, and 5 microarcsecond features at 1.3 mm correspond to baselines several times the Earth's diameter, which EHT cannot measure. Section 5 explicitly defers instrument resolution, thermal noise, and imaging specifics. Consequently, the abstract's statement that scattering mitigation is possible 'well below the nominal instrumental resolution of EHT' is not supported by the experiments as presented. The claim should be restricted to the full-information image-domain problem, or an experiment with realistic coverage and noise should be added.","section":"Section 3.2, 4.3, 5"},{"comment":"The forward model and training data use a Kolmogorov phase structure function with alpha = 5/3, while the paper itself notes that observations of Sgr A* favor alpha near 1.38 (Johnson et al. 2018). Because the model is trained to suppress refractive substructures whose statistics depend on alpha, the reported performance may be specific to alpha = 5/3. No robustness test at alpha = 1.38 is reported. A concrete test would be to generate scattered images with alpha = 1.38 and evaluate the already-trained model, and also to retrain on alpha = 1.38 images, quantifying the change in the NXCORR crossing scale.","section":"Section 2.1, 3.2"},{"comment":"The training image generation uses I proportional to g exp(n0) and then approximates this by g |n0 + 1|. For a unit-variance n0, exp(n0) is not well approximated by |n0 + 1|; the approximation changes the amplitude distribution and the power spectrum of the simulated intrinsic fluctuations. This makes the exact training distribution unclear and hampers reproducibility. Please either justify the approximation in a valid small-|n0| regime or generate training images as g exp(n0) directly.","section":"Section 3.2, Eq. (16)"},{"comment":"The deblurred baseline is computed with a Fourier-domain floor of K_tilde = 0.1 beyond 10 G lambda, whereas the IRIM model is applied to and evaluated at all scales in the fully sampled image. Because the two methods receive different information, the NXCORR gap at small scales partly reflects this asymmetry rather than the intrinsic merits of IRIM. The comparison should be made under the same band limit, for example by applying the 10 G lambda cutoff to both the input and output of IRIM or by evaluating both methods only on angular scales accessible to EHT.","section":"Section 4.1"}],"minor_comments":[{"comment":"The text says 'trained the model over one interaction of this set'; this should be 'one iteration'.","section":"Section 3.3"},{"comment":"The caption contains 'Irim' and mixes 'descattered' and 'deblurred' labels; please correct the spelling and clarify which image corresponds to which method.","section":"Figure 6 caption"},{"comment":"The phrase 'this process both diffractive blurs and adds stochastic refractive substructures that limits' should be 'that limit' for grammatical agreement.","section":"Abstract"},{"comment":"There is a typesetting issue: 'r + r2 F ∇phi(r)' should be 'r + r_F^2 ∇phi(r)'.","section":"Equation (2)"},{"comment":"The choice of rho = 0.95 as the fidelity threshold is arbitrary; the 4 microarcsecond crossing scale depends on this choice. Reporting the sensitivity of the crossing scale to the threshold would help the reader assess the robustness of the headline number.","section":"Section 4.1"},{"comment":"The Porth et al. (2019) reference appears twice; please consolidate.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is best read as a proof of concept for an idealized image-domain descattering problem. The main gap is the absence of the interferometric observation operator; this is fixable either by adding realistic sparse-coverage and noise simulations or by clearly reframing the claims as applying only to fully sampled images. The paper fits the journal's scope if presented as a proof of concept rather than as an EHT-ready result."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, you should know this paper is a solid simulation proof-of-concept that gets oversold in the abstract. The new thing is applying an invertible recurrent inference machine (IRIM) to descatter Sgr A* images, leveraging the non-birefringence of the scattering screen across four Stokes channels. The architecture and scattering formalism are prior work, but the combination, plus the explicit claim of recovering structure below 5 microarcseconds on GRMHD test images, is new. What is genuinely good: the out-of-distribution evaluation. The model is trained on phenomenological Kolmogorov Gaussians and tested on GRMHD simulations it never saw, and it beats the deblurred and scattered baselines at all reported scales. The paper also does a fair job of comparing against the standard deblurring approach and is honest in Section 5 about the missing complications. That honesty is worth crediting.\n\nThe soft spots are in proportion. The most serious is the observation operator. Everything is done on fully sampled, noise-free images; the 5 microarcsecond crossing comes from comparing descattered truth and estimate at pixel scales far finer than EHT baselines can measure. Real EHT data has sparse (u,v) coverage and thermal noise, and the paper explicitly defers those. So the abstract's phrase 'demonstrate that it is possible to mitigate interstellar scattering' overreaches; the demonstration is conditional on a complete forward model. This is exactly the kind of gap a referee should push on. The scattering index mismatch (5/3 used, 1.38 favored by Sgr A* observations) is real but secondary; it is a parameter, not the forward model itself. Also, the headline numbers are single representative curves without ensemble statistics, so we don't know the variance across realizations. These are not fatal flaws; they are the conditions that define what has actually been shown.\n\nThe paper deserves a serious referee. For a reader working on EHT imaging or scattering, it is a useful reference and a plausible stepping stone. I would bring it to the reading group and would cite it if I were working on scattering mitigation. My recommendation: send it to peer review, but the referee should require either a softened conclusion or an additional experiment with realistic (u,v) coverage and noise. As it stands, the evidence supports a conditional capability claim, not the unconditional one in the abstract.","headline":"A capable simulation-only proof-of-concept that is oversold in the abstract; the 5 microarcsecond claim waits on realistic (u,v) coverage and noise.","tokens_in":13559,"tokens_out":3197,"would_cite":true,"duration_ms":26085,"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":"A recurrent neural network trained on simple synthetic blobs can remove interstellar scattering from 1.3 mm images of Sgr A* down to 5 microarcsecond scales.","keywords":["interstellar scattering","Sgr A*","Event Horizon Telescope","recurrent inference machine","inverse problem","deconvolution","polarimetry","very long baseline interferometry"],"falsifier":"Train the identical IRIM model on scattering screens generated with the observed phase-structure index $\\alpha = 1.38$ and evaluate NXCORR on the same GRMHD test set; if the descattered images no longer maintain $\\rho > 0.95$ at 4-5 microarcseconds, the specific 5 microarcsecond claim fails for the actual galactic-center screen. Complementary test: feed real EHT-like sparse $(u,v)$ data with thermal noise into the trained model and check whether the recovery threshold degrades.","tokens_in":12474,"feed_emoji":"🔭","tokens_out":6499,"duration_ms":49784,"temperature":0.7,"pith_summary":"This paper claims that a recurrent neural network called an Inverse Recurrent Inference Machine (IRIM) can remove both the diffractive blur and the stochastic refractive substructures that interstellar scattering imprints on 1.3 mm images of Sagittarius A*. Trained only on synthetic 'Kolmogorov Gaussian' images, the network recovers structures down to about 5 microarcseconds, far below the Event Horizon Telescope's nominal 24 microarcsecond resolution. If true, scattering is not a fundamental limit for ground-based VLBI imaging of the galactic center; the deconvolution problem, though formally ill-posed, can be solved in practice.","feed_headline":"Neural net sees through Sgr A* scattering down to 5 microarcseconds","feed_subtitle":"Model trained on Gaussian blobs recovers black-hole images below EHT's 24 microarcsecond resolution.","key_machinery":"The machinery is the Inverse Recurrent Inference Machine (IRIM), an invertible recurrent neural network that solves the inverse problem $y = Ax + \\epsilon$ by iteratively refining an estimate with a learned update that combines a user-supplied likelihood gradient with a learned prior via a memory variable. In this application, the forward map is the thin-screen scattering relation: the scattered Stokes image is the diffractively blurred intrinsic image, locally remapped by the gradient of the random phase screen, and the refractive contribution plays the role of the noise. The network is trained on 200,000 Kolmogorov-Gaussian images with noise drawn from the assumed Kolmogorov phase power spectrum, and it exploits the non-birefringence of the screen: all Stokes parameters see the same phase realization, so each image provides a channel for the same corruption. The key estimator that carries the argument is the Stokes-averaged normalized cross-correlation, NXCORR, evaluated as a function of a Gaussian blur scale to define the effective resolution down to which mitigation succeeds.","core_discovery":"The central claim is that scattering mitigation at resolutions relevant to the EHT is possible without any strong prior on the intrinsic image. The authors demonstrate this by training IRIM on phenomenological images and testing on GRMHD simulations of Sgr A* that were never seen in training. On these test images the Stokes-averaged normalized cross-correlation between the descattered estimate and the truth stays above 0.95 for all angular scales larger than roughly 4 microarcseconds, and the IRIM descattering outperforms simple deblurring with the diffractive kernel at every scale. The conclusion is that the information needed to undo both diffractive and refractive scattering is present in scattered images, so ill-posedness is not an obstacle to practical mitigation.","pith_inferences":["A critical robustness check the paper leaves implicit: repeat the training and evaluation with the observed Sgr A* phase-structure index (about 1.38) instead of Kolmogorov $\\alpha = 5/3$; the 5 microarcsecond claim currently rests on the Kolmogorov choice.","Because training uses fully sampled images, real EHT visibilities with sparse $(u,v)$ coverage and thermal noise may degrade the effective resolution; a natural next test is to inject the trained model into the EHT imaging pipeline and compare recovered images against the current deblurring baseline.","If the descattering generalizes, it could be applied to other strongly scattered Galactic-center sources, and the same IRIM formalism could be reused for chromatic deconvolution in multi-frequency VLBI.","The paper's success suggests that learned iterative inference with physically motivated forward models may outperform closed-form deconvolution for other stochastic point-spread-function problems, such as atmospheric or ionospheric phase errors."],"forward_implications":["If the result holds, EHT and next-generation arrays can treat scattering as a correctable nuisance rather than a floor on angular resolution, permitting studies of accretion-flow substructure at scales of a few microarcseconds.","The method's independence from ring priors means it can be applied to extended or asymmetric sources near the galactic center, not just horizon-ring morphology.","Since the same phase screen corrupts I, Q, U, and V, the model provides a route to high-fidelity polarimetric images of Sgr A* after mitigation.","The training-on-Gaussians, testing-on-GRMHD generalization suggests that scattering mitigation may transfer to sources whose morphology is poorly known in advance.","Potential use in future multi-wavelength VLBI campaigns: after mitigation at 1.3 mm, comparisons with 0.87 mm data become cleaner."],"supporting_citations":[{"why":"Supplies the invertible recurrent inference machine (IRIM) architecture and the invert-train procedure the paper adapts for descattering.","marker":"Putzky & Welling 2019"},{"why":"Introduces Recurrent Inference Machines, the iterative-inference framework IRIM extends with an invertible design.","marker":"Putzky & Welling 2017"},{"why":"Provides the measured Sgr A* scattering parameters and the about-1.38 phase-structure index against which the adopted Kolmogorov index is compared.","marker":"Johnson et al. 2018"},{"why":"Establishes that interstellar scattering is non-birefringent, justifying the joint multi-Stokes training.","marker":"Ni et al. 2022"},{"why":"Supplies the GRMHD simulation suite from which the unseen test images are drawn.","marker":"Porth et al. 2019"},{"why":"Defines the NXCORR fidelity statistic and the validation simulation sets used for the GRMHD tests.","marker":"Sgr A* Paper IV"},{"why":"Provides the polarimetric GRMHD snapshots and associated simulation parameters used as test images.","marker":"M87* Paper V"}],"fun_headline_variants":["AI descatters Sgr A* images to 5 microarcseconds","Neural network removes scattering blur from Sgr A* down to 5 μas","AI sees Sgr A* below EHT resolution: 5 μas","Deep learning descatters Sgr A* to 5 microarcseconds","Recurrent net descatters Sgr A* to 5 μas"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's training and tests assume that the simulated thin-screen scattering with a Kolmogorov ($\\alpha = 5/3$) phase structure function and fully sampled images faithfully represents the real EHT measurement process, whereas Sgr A* observations favor an index near 1.38 and real data have sparse coverage and thermal noise.","fun_headline_variants_meta":{"raw":{"variants":["AI descatters Sgr A* images to 5 microarcseconds","Neural network removes scattering blur from Sgr A* down to 5 μas","AI sees Sgr A* below EHT resolution: 5 μas","Deep learning descatters Sgr A* to 5 microarcseconds","Recurrent net descatters Sgr A* to 5 μas"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00094,"raw_usage":{"total_tokens":3964,"prompt_tokens":835,"completion_tokens":3129,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":3029}},"tokens_in":451,"tokens_out":3129,"duration_ms":21382,"temperature":1.0,"reasoning_tokens":3029,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:26:59.877096+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical IRIM model on scattering screens generated with the observed phase-structure index $\\alpha = 1.38$ and evaluate NXCORR on the same GRMHD test set; if the descattered images no longer maintain $\\rho > 0.95$ at 4-5 microarcseconds, the specific 5 microarcsecond claim fails for the actual galactic-center screen. Complementary test: feed real EHT-like sparse $(u,v)$ data with thermal noise into the trained model and check whether the recovery threshold degrades.","supporting_citations":[],"review_version":1}