{"id":"db99f8e2-1cd6-4d14-bc43-e9cdd9a2ad4a","arxiv_id":"2506.11200","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A CNN plus normalizing-flow pipeline recovers calibrated 4D posteriors for planet mass, viscosity, scale height and Stokes number from ALMA disc images, and infers mostly sub-Jupiter candidates in 49 observed gaps.","lead":"DBNets2.0 is a publicly available deep-learning tool that reads a protoplanetary disc image and returns a probability distribution for the planet mass plus three disc properties, exposing how different physical settings can produce the same gap. The authors test it on synthetic ALMA-like observations and then apply it to 49 observed gaps, finding mostly sub-Jupiter planets, which matches why direct imaging has not detected them.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The real-data population results rest on the single-planet assumption, yet DBNets2.0's own confidence score flags none of the two-planet simulations (Fig. A.4); Sect. 6's 49-gap masses are therefore unprotected against the most likely excluded scenario.","rationale":"The reader's weakest_assumption already identifies the single-planet restriction, and I agree it is the decisive issue. The reason it is load-bearing rather than a standard 'future work' caveat is quantitative: the paper's own Fig. A.4 shows the confidence score has zero power against two-planet systems, and Table E.1 shows all 49 real gaps pass the recommended threshold. A reader cannot tell which of the 49 inferences are in-distribution single-planet-like and which are multi-planet morphologies that the model would happily fit. The methodological core is nevertheless independently supported: TARP calibration on a held-out test set, posterior predictive checks, and public code are real evidence that, conditional on the forward model, the posteriors are reliable. The r2 formula typo in Eq. (5) and the incomplete release of training data or weights are real but secondary issues; neither undermines the method itself. I therefore do not move the reader's CONDITIONAL verdict; the condition should remain, and the concrete test above would determine whether the real-data section needs to be presented even more cautiously.","tokens_in":32797,"tokens_out":8528,"duration_ms":98862,"concrete_test":"For a grid of two-planet FARGO3D simulations spanning the training parameter ranges, run DBNets2.0 exactly as in the real-data pipeline: apply it to each gap individually, record the median and 16th–84th percentiles of the inferred primary-planet mass posterior, and compare with the true mass of the planet responsible for that gap. Quantify the fraction of cases in which the true mass falls outside the credible interval or the inferred mass is off by more than 0.5 dex. If that fraction is large, the single-planet assumption is a dominant systematic and the Sect. 6 population results should be downgraded to scenario-conditional illustrations; if the intervals cover the true masses in most cases, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central methodological claim — a TARP-calibrated SBI estimator for the four parameters under the training forward model — is well supported. The load-bearing step is the transfer to real discs: Sect. 6 interprets each of 49 observed gaps as produced by one embedded planet, yielding a population with 83% sub-Jupiter candidates. Every element of this transfer depends on the single-planet, locally-isothermal, no-migration, no-feedback model of Sect. 3.1. The pipeline's own robustness metric, the confidence score, was tested on two-planet simulations (Appendix A.1, Fig. A.4), where all inputs scored above the recommended 0.6 threshold; the same is true of every real observation in Table E.1. Thus the metric cannot distinguish the most likely excluded scenario (multiple planets) from in-distribution data, and Sect. 3.4 concedes that a high score cannot exclude different models. Since several analysed discs (AS 209, HD 163296, HL Tau) contain multiple gaps that are treated independently, the inferred low masses and low α values could be artefacts of forcing a single-planet morphology. This does not invalidate the simulator-based inference framework; it invalidates any claim that the observed-gap population results are protected against model misspecification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents DBNets2.0, a simulation-based inference pipeline that compresses ALMA-like dust continuum images with a CNN and fits the four-dimensional posterior p(Mp, alpha, h, St | x) with normalizing flows. The training and test data are single-planet FARGO3D simulations with a locally isothermal gas and pressureless dust, augmented with noise, beam convolution, and masking. The authors validate on a held-out synthetic test set with TARP coverage tests, rmse/r2 metrics, and three posterior predictive checks; they then apply the pipeline to 49 gaps in 34 observed discs and report generally low alpha, low h, and a population with 83% masses below 1 MJ.","tokens_in":33059,"tokens_out":10153,"duration_ms":112891,"significance":"The central methodological contribution is solid: the 4D TARP curve is well calibrated (ks-pval 0.999), the planet-mass rmse is 0.07 in normalized units, the code is public, and the paper is honest about many limitations. If the validation transfers to real observations, the tool will be useful for interpreting dust substructures and for Bayesian integration of external constraints such as direct imaging limits. The main limitation is that this transfer is not demonstrated: the confidence-score metric fails to flag the paper's own two-planet simulations, and the observational population results depend on a single-planet model that real discs may violate.","major_comments":[{"comment":"The two-planet test in Fig. A.4 shows that every two-planet image receives a confidence score above the recommended 0.6 threshold, and Table E.1 shows that all 49 real observations also score above 0.6. Since Sect. 6 interprets each of these observations as a single embedded planet, the confidence score does not protect the results against the most likely excluded scenario, multiple planets per gap. I acknowledge the caveats in Sect. 3.4 and 8, but the abstract and Sect. 6 present the 83% sub-Jupiter population without this qualification. Please either add an OOD test that demonstrates measurable degradation on two-planet systems or explicitly rephrase the Sect. 6 population results as conditional on the single-planet model.","section":"Appendix A.1 / Sect. 6"},{"comment":"The test-set size is stated inconsistently: Sect. 3.1 reports 534 test observations, while Sect. 4.1 reports that 900 synthetic observations were selected for the test set. Please clarify that the 534 are the filtered subset of the 900 snapshots from the 300 held-out simulations, and state explicitly that no snapshot from a simulation contributing to the training or validation folds appears in the test set. Without this clarification, the reported TARP and rmse results cannot be independently verified against leakage between snapshots of the same simulation.","section":"Sect. 4.1 / Sect. 3.1"},{"comment":"The validation is entirely in-distribution: the test simulations use the same hydrodynamics code, the same single-planet and locally isothermal assumptions, and the same LHS parameter ranges as the training simulations. A calibrated posterior under this forward model does not by itself quantify the risk that real gaps were not generated by that model. The confidence score was intended to address this, but its two-planet test does not flag the excluded scenario. Please make the model-conditional nature of the Sect. 6 inferences explicit in the abstract and conclusions, or provide a quantitative OOD validation (e.g., coverage or posterior error on two-planet and non-planar simulations) that bounds the misspecification error for the real-data claims.","section":"Sect. 3.1 / Sect. 4.2"}],"minor_comments":[{"comment":"The r2-score definition is missing squared terms in the printed equation; as written it is 1 - Sum(theta - theta_hat)/Sum(theta - theta_bar), which is dimensionally inconsistent and not the standard coefficient of determination. Please correct it to 1 - Sum(theta - theta_hat)^2 / Sum(theta - theta_bar)^2.","section":"Sect. 4.2, Eq. (5)"},{"comment":"The description of TARP curve shapes appears reversed: an overconfident posterior typically gives ECP below the diagonal for all alpha_TARP, whereas an underconfident posterior gives ECP above the diagonal at low alpha_TARP. Please check the wording.","section":"Sect. 4.2"},{"comment":"The stopping criterion for the normalizing-flow training ('no longer improve') should specify the patience and the validation metric used, to make the procedure reproducible.","section":"Sect. 3.3"},{"comment":"The simulated external priors in the constraint-integration test are unbiased by construction (their means are drawn around the true value of the constrained property), so the reported improvements are optimistic relative to real external constraints that may be biased; please state this caveat explicitly.","section":"Sect. 7.2"},{"comment":"The confidence score uses linear interpolation on the training images as a surrogate for simulations; this makes the score sensitive to training-set density and interpolation details. A sentence describing the interpolation scheme and its limitations should be added.","section":"Appendix A"},{"comment":"The caption of Fig. 16 refers to 'HD14266', which does not match the object 'HD 142666' in Table E.1; please correct the typo.","section":"Fig. 16 / Table E.1"}],"recommendation":"major_revision","confidential_remarks":"The central SBI validation is sound and the paper is within scope. The main revision should focus on making the model-conditional nature of the Sect. 6 population results unambiguous and on clarifying the train/test split; I do not see a need to redo the full inference."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the methodological core is solid and worth taking seriously, but the real-data conclusions carry a caveat the paper acknowledges and then partly pushes past. DBNets2.0 genuinely improves on the earlier DBNets by inferring a full 4D joint posterior (planet mass, alpha, scale height, Stokes number) via CNN summaries plus normalizing flows, and it removes the resolution and outer-boundary sensitivities that plagued the old tool. The TARP-calibrated held-out test set (ks-pval 0.999) and the resolution-robustness tests are convincing, and the code is public. That is the right way to build a population-inference tool.\n\nThe main soft spot is that the validation is entirely in-distribution. The test images come from the same FARGO3D model, same physics, same parameter ranges, so calibration only shows the posterior is accurate within that training family. The authors state this in Sect. 3.4, and say a high confidence score cannot exclude other models. But then Appendix A.1 shows the confidence score leaves all of their two-planet simulations above the 0.6 threshold. That matters because the most natural alternative to a single planet is multiple planets, and Sect. 6 treats each of the 49 gaps as an independent single-planet system, including multiple gaps in AS 209, HD 163296, and HL Tau. If those morphologies are multi-planet or non-planar in origin, the low masses and low alpha values in the population results could be artefacts of forcing the wrong model. This does not invalidate the SBI pipeline; it limits the strength of the population claims.\n\nMinor issues: the r2 formula in Eq. 5 appears to be missing squares; the full training data and trained weights are not clearly released despite the code being public; and the comparison with direct-imaging non-detections is described qualitatively.\n\nThe paper is useful, well-structured, and honest about model dependence. The Sect. 6 interpretation, however, goes further than the validation supports. For whom: anyone working on planet-disc interaction or using SBI forward models in astronomy. Recommendation: send to peer review. A serious referee should ask the authors to temper or re-frame the population results, or to demonstrate the confidence score against multi-planet and non-planet simulations. I would cite the pipeline.","headline":"A genuinely useful SBI pipeline with well-calibrated posteriors on synthetic data; the population-level planet masses rest on a single-planet model that the paper's own confidence score cannot flag.","tokens_in":33601,"tokens_out":2893,"would_cite":true,"duration_ms":30960,"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 single dust-continuum image of a protoplanetary disc gap can be mapped to a calibrated four-dimensional posterior over planet mass, disc viscosity, scale height, and dust Stokes number.","keywords":["simulation-based inference","neural posterior estimation","normalizing flows","protoplanetary discs","planet-disc interaction","dust continuum","convolutional neural networks","planet mass inference"],"falsifier":"Create two-planet synthetic observations with the same FARGO3D setup and known masses, run DBNets2.0, and check the confidence scores: the paper already reports that all such images score above the recommended 0.6 threshold, so a corrected confidence metric that actually rejects these out-of-model images — or, alternatively, a PDS 70 run whose planet-mass posterior brackets the directly measured companion masses — would settle whether the reliability claims hold.","tokens_in":32577,"feed_emoji":"🪐","tokens_out":5573,"duration_ms":64295,"temperature":0.7,"pith_summary":"The authors are trying to establish that the morphology of a dust gap in a protoplanetary disc carries enough information to recover not just the mass of an embedded planet, but also the disc's viscosity, scale height, and dust grain size — and that these can be recovered together as a full joint posterior, with the degeneracies between them made explicit. This matters because planet mass alone cannot be inferred from a gap without knowing the disc conditions; a tool that exposes the degeneracy allows astronomers to combine independent measurements and actually constrain the planet population. The paper builds a simulation-based inference pipeline, validates it on held-out synthetic observations, and then applies it to 49 real gaps in 34 discs. The resulting population is mostly low-mass planets, consistent with the failure of direct imaging surveys to detect them.","feed_headline":"Dust gap images now yield planet masses and disc physics together","feed_subtitle":"Simulation-based tool maps one ALMA image to a four-parameter posterior, and 83% of 49 observed gaps come out sub-Jupiter.","key_machinery":"The pipeline is a CNN-plus-normalizing-flow architecture: a convolutional network turns the input image (with its resolution as an extra conditioning input) into low-dimensional summary statistics, and masked autoregressive flows perform neural posterior estimation on those summaries. The CNN is trained with augmentation that randomizes beam size and outer disc boundary, which is what makes the tool's accuracy independent of observational resolution; Monte Carlo dropout generates 1500 summary-statistic samples per image, and a confidence score built on Fourier-domain comparison with the training set provides a recommended rejection threshold of 0.6.","core_discovery":"The central claim is that DBNets2.0, a two-stage simulation-based inference pipeline, produces well-calibrated posteriors for the planet-to-star mass ratio $M_p/M_\\star$, the disc $\\alpha$-viscosity, the aspect ratio $h$, and the dust Stokes number $St$ from a single ALMA dust continuum image of a disc with substructure. The first stage is a CNN that compresses each image, together with its beam size, into a set of summary statistics via Monte Carlo dropout; the second stage is an ensemble of masked autoregressive normalizing flows that learn the posterior $p(M_p, \\alpha, h, St \\mid x, b)$. On a held-out synthetic test set the full four-dimensional posteriors pass a TARP coverage test, with the planet mass recovered most precisely. Applied to 49 observed gaps in 34 discs, the tool infers generally low viscosities and scale heights, and planet masses below one Jupiter mass in 83% of cases.","pith_inferences":["The systematic offset the paper finds between local $\\alpha$ inferred from gap morphology and global accretion-based viscous timescales, if it survives selection effects, is evidence for an additional angular-momentum-loss channel such as disc winds; the paper notes this possibility but stops short of making it a quantitative claim.","The confidence score's failure to reject two-planet synthetic images — all scores stay above the recommended 0.6 threshold — suggests that the tool as designed cannot detect the most likely alternative to the single-planet model, so a multi-planet extension or a joint posterior over multiple gaps is needed before multi-gap discs are fitted independently.","A natural validation that the paper does not perform explicitly is to run DBNets2.0 on PDS 70, whose two embedded planets have direct mass measurements, and check whether the inferred mass posteriors bracket those measurements; the paper's published PDS 70 estimate is consistent, but a dedicated validation with updated masses would tighten the test.","The same summary-statistics-plus-normalizing-flows recipe could be applied to other disc observables, such as gas kinematics or scattered-light polarization, wherever a forward simulator exists; that template is implicit in the paper's design but not pursued."],"forward_implications":["A single ALMA image of a gap can be converted into a calibrated four-dimensional posterior, so degeneracies such as the planet-mass–viscosity correlation are exposed rather than hidden inside a point estimate.","Independent constraints on one disc property, such as a measured $\\alpha$, can be folded in as a prior; the paper demonstrates a roughly 15% reduction in planet-mass error and uncertainty when $\\alpha$ is constrained this way.","The 49-gap survey implies low local viscosities and long viscous timescales, and a planet population that is 83% sub-Jupiter, which explains why direct imaging surveys have mostly failed to detect these putative planets.","Because the tool is public and returns full posteriors rather than point estimates, it can be re-run with different priors or applied to new observations without retraining."],"supporting_citations":[{"why":"Supplies the hydrodynamic simulation dataset, the previous DBNets tool, and the observed-disc sample that this pipeline extends and compares against.","marker":"Ruzza et al. (2024)"},{"why":"Provides the TARP coverage test used to validate the four-dimensional posteriors and define the ks-pval and atc metrics.","marker":"Lemos et al. (2023a)"},{"why":"Supplies the CNN-summary-statistics plus normalizing-flow architecture pattern that DBNets2.0 adopts.","marker":"Lemos et al. (2023b)"},{"why":"Provides the sbi package implementing masked autoregressive flows for neural posterior estimation.","marker":"Tejero-Cantero et al. (2020)"},{"why":"Provides the FARGO3D code that produced the planet-disc simulations in the training set.","marker":"Benitez-Llambay & Masset (2016)"},{"why":"Justifies Monte Carlo dropout, which generates the summary-statistic samples fed to the normalizing flows.","marker":"Gal & Ghahramani (2015)"},{"why":"Gives the K-coefficient gap-depth scaling used to interpret the inferred degeneracies and the observed K distribution.","marker":"Kanagawa et al. (2015)"},{"why":"Supplies typical viscosity and planet-mass constraints used to compare with the 49-gap survey results.","marker":"Zhang et al. (2018)"},{"why":"Provides Hill-radius-based planet mass estimates compared directly with DBNets2.0 on the same observed gaps.","marker":"Lodato et al. (2019)"}],"fun_headline_variants":["Simulation-based inference maps dust gaps to planet and disc properties","Posterior from ALMA images: planet mass plus disc viscosity, height, Stokes number","DBNets2.0: one image yields planet mass and 3 disc parameters","Neural flows infer planet mass and disc parameters from dust substructure","Dust ring images now yield planet mass and disc physics posteriors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the training-set premise that each observed gap is carved by exactly one planet in a locally isothermal, viscous disc with pressureless dust and no feedback, self-gravity, migration or accretion; the paper's own test shows the confidence score cannot flag two-planet systems, so if real gaps are multi-planet or non-planar, the inferred masses and disc properties are not reliable.","fun_headline_variants_meta":{"raw":{"variants":["Simulation-based inference maps dust gaps to planet and disc properties","Posterior from ALMA images: planet mass plus disc viscosity, height, Stokes number","DBNets2.0: one image yields planet mass and 3 disc parameters","Neural flows infer planet mass and disc parameters from dust substructure","Dust ring images now yield planet mass and disc physics posteriors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1602,"prompt_tokens":1092,"completion_tokens":510,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":708,"completion_tokens_details":{"reasoning_tokens":412}},"tokens_in":708,"tokens_out":510,"duration_ms":5448,"temperature":1.0,"reasoning_tokens":412,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:13:02.485936+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Create two-planet synthetic observations with the same FARGO3D setup and known masses, run DBNets2.0, and check the confidence scores: the paper already reports that all such images score above the recommended 0.6 threshold, so a corrected confidence metric that actually rejects these out-of-model images — or, alternatively, a PDS 70 run whose planet-mass posterior brackets the directly measured companion masses — would settle whether the reliability claims hold.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the hydrodynamic simulation dataset, the previous DBNets tool, and the observed-disc sample that this pipeline extends and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the sbi package implementing masked autoregressive flows for neural posterior estimation."},{"cited_title":"& Masset, F","cited_arxiv_id":null,"evidence_quote":"Provides the FARGO3D code that produced the planet-disc simulations in the training set."},{"cited_title":"& Ghahramani, Z","cited_arxiv_id":null,"evidence_quote":"Justifies Monte Carlo dropout, which generates the summary-statistic samples fed to the normalizing flows."},{"cited_title":"2018, AJ, 869, L47","cited_arxiv_id":null,"evidence_quote":"Supplies typical viscosity and planet-mass constraints used to compare with the 49-gap survey results."},{"cited_title":"2019, MNRAS, 486, 453","cited_arxiv_id":null,"evidence_quote":"Provides Hill-radius-based planet mass estimates compared directly with DBNets2.0 on the same observed gaps."}],"review_version":1}