{"id":"68575a5c-3251-4ea2-bcd1-5370026f0c04","arxiv_id":"2508.13088","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Bayesian system that samples and visualizes the distribution of simulation parameters consistent with a user-specified output feature, using a density prior and Hamiltonian Monte Carlo.","lead":"This paper shows how to display every simulation setting that produces a chosen output feature, not just a single best match. It combines a neural surrogate, a training-data-aware prior, and fast Hamiltonian Monte Carlo sampling so scientists can explore parameter distributions interactively.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ground-truth validation: posterior samples are judged only by the surrogate that defines them, so the 'plausible parameters' claim rests on an untested density-prior proxy.","rationale":"The paper makes a valuable and clearly presented contribution: it casts inverse surrogate exploration as posterior sampling, with a pragmatic density-based prior, and provides HMC diagnostics (R-hat, MMD convergence, NLL quality) that convincingly show the sampler converges to the intended posterior. The visualization design and case studies demonstrate utility for qualitative parameter-feature analysis. These are real strengths. However, the central claim is stronger than the evidence. 'Plausible parameter configurations that generate a target output feature' is a statement about the underlying simulation, not about the surrogate. Every quantitative quality signal in the paper is computed from the surrogate output that also defines the likelihood; this would be acceptable only if the prior could be trusted to eliminate surrogate-error regions. The reader's weakest assumption—that the density prior is a reliable proxy for surrogate approximation error—is exactly the load-bearing assumption. I agree with the reader that this is the central risk. The proposed concrete test (running the true simulator on posterior samples) is straightforward and would settle whether the concern lands. If it passes, the method's claims are substantiated; if it fails, the posterior is an artifact of the surrogate and the visualization could mislead. Because the missing validation is addressable and the method is otherwise well-engineered, I do not see grounds to change the reader's CONDITIONAL verdict.","tokens_in":11964,"tokens_out":8811,"duration_ms":85843,"concrete_test":"Run the original simulation solver (e.g., the CFD code used to generate the NACA and Rayleigh-Taylor training ensembles) at 50-100 random posterior samples for a small set of feature specifications. For each sample, extract the same spatial-patch feature from the true output and compute the feature error with the same distance used in the likelihood (Eq. 10). Report the fraction of samples with true-feature error below the threshold used in the likelihood, and compare this fraction to that of uniform random parameter draws. If the posterior's true-feature match rate is not markedly higher than uniform and not above a reasonable threshold (e.g., 80%), the density prior has failed to confine sampling to reliable surrogate regions, and the central plausibility claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that the method 'gives us an efficient way to sample plausible parameter configurations that generate a target output feature.' Here 'plausible' refers to the true simulation, yet no experiment in the paper verifies that sampled parameters, when run through the simulator that generated the training data, actually produce the specified feature. The posterior is defined by a likelihood computed from the surrogate output and a prior derived from closeness to training parameters (Sec. 3.4). All feature-quality evaluations—negative log-likelihood (Fig. 14), comparisons to uniform sampling (Fig. 15), and the case-study observations—use the same SIREN surrogate that defines the likelihood, so the evaluation is self-consistent by construction. The density prior is the only component intended to protect against surrogate approximation error, but its reliability is supported only by a qualitative illustration (Appendix B) and an asserted comparability to deep ensembles; no quantitative measure of how well the density estimate predicts actual surrogate error on held-out simulations is provided. If the Fisher-information distance misorders parameter neighborhoods, or if the nearest-neighbor bandwidths are miscalibrated, the prior will place high density in regions where the surrogate is wrong, and HMC will confidently produce parameters whose true outputs do not match the requested feature. The Discussion mentions surrogate reliability as a general concern but offers no test. This untested proxy assumption is the single point on which the central claim depends.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian framework for sampling distributions of simulation parameters conditioned on a user-specified output feature in neural-field surrogates. The prior is a density estimate over training parameters based on Fisher information and Euclidean distances, and the likelihood is derived from the surrogate's negative log-likelihood between the predicted feature patch and the target feature (Eqs. 9-10). Hamiltonian Monte Carlo draws samples from the resulting posterior, and a scatterplot-matrix interface displays the marginal distributions. The paper evaluates HMC convergence via R-hat and MMD statistics and assesses feature-match quality using the same NLL, then demonstrates the workflow on three simulation datasets.","tokens_in":12283,"tokens_out":4283,"duration_ms":46261,"significance":"If the posterior samples were validated against ground-truth simulations, the method would be a practical and useful contribution to surrogate-based inverse exploration in visualization and scientific analysis. The paper has clear strengths: it ships source code, provides explicit HMC diagnostics (R-hat and MMD convergence), and demonstrates a progressive sampling scheme that supports interactive visualization. However, the central claim that sampled parameters are 'plausible' and 'generate a target output feature' is not yet established, because the evaluation is self-consistent with the surrogate that defines the posterior. The density prior is an interesting idea, but its reliability as a proxy for surrogate approximation error is asserted rather than demonstrated.","major_comments":[{"comment":"The central claim that sampled parameters are plausible and generate a target output feature is validated only against the surrogate's negative log-likelihood that defines the posterior. Figures 14 and 15 evaluate feature-match quality using the same NLL (Eq. 9) that appears in the likelihood, making the evaluation circular. No experiment runs the HMC-sampled parameters through the actual simulator that generated the training data to verify that the requested feature is present. I recommend adding ground-truth validation: for held-out target features, sample from the posterior, run the simulator, and measure the feature-match error, comparing against uniform random sampling and a baseline optimizer.","section":"Section 5 and Appendix C, Eqs. (9)-(10)"},{"comment":"The density prior is the only mechanism protecting against surrogate approximation error, yet its validity as a proxy is not quantitatively established. The paper asserts in Section 1 that the density estimate is 'comparable in being predictive of surrogate error' to standard uncertainty estimates, but no experiment directly measures the correlation between the density and actual surrogate error on held-out inputs. Appendix B provides only a qualitative illustration for the Circle simulation. I recommend a quantitative study: compute the density and the true surrogate error over a held-out set of parameter configurations, report their correlation, and test sensitivity to the bandwidths sigma_f and sigma_s.","section":"Section 3.4 and Appendix B"},{"comment":"The HMC diagnostics (R-hat, MMD) demonstrate convergence to the surrogate-defined posterior, but they do not validate the correctness of that posterior relative to the true simulation. The free parameters C (feature likelihood scale), sigma_f, sigma_s, step size, and leapfrog steps are set heuristically, and no sensitivity analysis is reported. Without such an analysis, it is unclear how robust the displayed distributions are to reasonable variations in these choices. Please add a sensitivity study over C, sigma_f, sigma_s, and HMC step size, and report how the posterior marginals change.","section":"Section 3.5 and Appendix C"}],"minor_comments":[{"comment":"In several places the text uses 'c.f.' where the standard abbreviation is 'cf.' (e.g., Section 3.5 and Section 4); this is a minor style issue.","section":"Throughout"},{"comment":"The paragraph beginning 'Last, throughout the paper...' should use 'Finally' and could be split into two paragraphs for readability.","section":"Section 6"},{"comment":"The case-study observations are presented as user-made findings, but no user study is reported; the text should clarify that these are illustrative scenario walkthroughs rather than empirical evaluations.","section":"Section 5.2"},{"comment":"The R-hat histograms would be easier to interpret if the text stated the specific threshold used to judge mixing (e.g., R-hat below 1.1), rather than only saying 'quite close to 1.'","section":"Section 5.1 and Figure 12"}],"recommendation":"major_revision","confidential_remarks":"This is a well-written visualization-systems paper, but the validation gap is substantial: the main claim depends on a density-prior proxy that is not tested against ground-truth simulation outputs. I believe the missing experiments are within the manuscript's scope and the existing code and diagnostics make them feasible, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is worth a look: it offers a genuinely new way to explore the full distribution of inverse parameters in a neural surrogate, and the HMC implementation is handled with care. But the central claim that the sampled parameters are 'plausible' rests on an assumption that is not directly tested: that the density prior reliably tracks where the surrogate is accurate.\n\nWhat's new: instead of returning a point match or a whole-field inverse, the authors define a posterior over parameters given a local output feature, using a training-density prior (a Fisher-information-based distance plus Euclidean distance) and a feature likelihood, then sample with HMC. That combination is not in Surroflow or Explorable INR. The visualization interface, with binned heatmaps of posterior marginals and output variance, is a practical addition. The HMC diagnostics (R-hat and MMD convergence curves) are thorough and support that the sampler is doing its job. Source code is provided, which is a plus.\n\nThe soft spots: the evaluation of feature-match quality uses the same negative log-likelihood that defines the posterior, so the results are self-consistent by construction. There is no test on real simulator outputs to see whether the sampled parameters actually produce the requested feature. The density prior is the only thing protecting against surrogate error, but its predictive power is only illustrated qualitatively in Appendix B, not measured on held-out simulations. The bandwidths (sigma_f, sigma_s, C) are heuristics and could be miscalibrated, though the case studies suggest they work. These are genuine limitations, but they are addressable and partially confessed in the Discussion.\n\nGiven it's a visualization paper, the circularity is not fatal; the tool is meant for exploratory analysis where the surrogate is a stand-in. Still, one experiment that runs a handful of sampled parameters through the simulator would make the 'plausible' claim much stronger.\n\nWho's it for: visualization researchers working on surrogate-based parameter exploration, and simulation analysts who want to see all the ways to hit a target feature. It's a solid contribution to that community, and it deserves a serious referee.\n\nRecommendation: send it to peer review. The authors should be asked to add ground-truth validation for a subset of samples and a quantitative comparison of the density prior against surrogate error.","headline":"Useful and novel distributional inverse-design paper; the main reservation is that 'plausible' parameters are never validated against the true simulator.","tokens_in":12761,"tokens_out":3374,"would_cite":false,"duration_ms":33389,"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":"The paper proposes sampling from a Bayesian posterior to recover the full distribution of parameters that produce a user-specified simulation feature, rather than searching for a few optimal matches.","keywords":["parameter space exploration","neural surrogates","inverse problems","Bayesian inference","Hamiltonian Monte Carlo","density estimation","ensemble visualization","neural fields"],"falsifier":"Take a feature located in a region where the surrogate is known to be inaccurate, for instance between two training configurations where the density map in Appendix B shows a low-density streak; run the full posterior sampler, then evaluate the high-density samples on the ground-truth simulator. If those samples frequently fail to reproduce the specified feature, the density prior has not captured surrogate error. A cheaper check is to compare the HMC sample distribution against a dense-grid posterior for the one-dimensional viscosity example and verify that the sample density matches the true posterior where the prior is peaked.","tokens_in":11795,"feed_emoji":"🎯","tokens_out":10872,"duration_ms":99846,"temperature":0.7,"pith_summary":"This paper tackles a question inverse-design tools usually skip: given a target feature in a simulated field, what is the full set of input parameters that could produce it, not just one best match? The paper claims that the answer is a probability distribution over parameters, formed by combining a density-based prior with a likelihood on the feature and then sampling from the posterior with Hamiltonian Monte Carlo. The prior is designed to report high density only where the surrogate is close to the training data, so the sampled configurations are both plausible matches and likely to be regions where the surrogate is accurate. The authors demonstrate the approach through a visualization interface on three simulation datasets, showing pairwise marginal distributions of the posterior that let a user compare features and see how parameters trade off against each other. If the claim holds, interactive parameter-space exploration can shift from 'find the optimum' to 'see the whole family of solutions.'","feed_headline":"Sampling replaces search to map every parameter behind a feature","feed_subtitle":"A density prior keeps samples near trustworthy surrogate regions, so inverse queries show the whole plausible set, not one optimum.","key_machinery":"At the center is a neural field surrogate $g(\\mathbf{x}, \\boldsymbol{\\theta})$ that predicts field values at spatial locations $\\mathbf{x}$ given simulation parameters $\\boldsymbol{\\theta}$, so the same network serves both forward evaluation and feature-based queries. The prior is a kernel density estimate whose bandwidths come from nearest-neighbor distances measured with a Fisher-information metric on the surrogate's output, which makes the density drop in regions where small parameter changes produce large field changes; this is the mechanism that encodes surrogate error into the prior. The feature likelihood compares the surrogate's predicted field over the user's selected region to the specified feature, and Hamiltonian Monte Carlo with leapfrog integration samples the resulting posterior. Pairwise binned heatmaps of the collected samples then approximate all two-dimensional marginals of the posterior for visualization.","core_discovery":"The central claim is that the inverse problem 'which input parameters generate this output feature?' should be solved as Bayesian posterior sampling rather than as optimization. The paper defines a prior over parameters via a density estimator that rewards configurations close to training points in both input and output space, and a likelihood that measures how well the surrogate's field over a user-selected neighborhood matches the target feature. The posterior is then proportional to the product of this likelihood and prior, and Hamiltonian Monte Carlo is used to draw samples whose density approximates the posterior. The paper argues this procedure is fast enough for progressive, interactive visualization and that the resulting sample distributions reveal the full set of plausible inverse solutions, including approximate matches and the correlations among parameters that produce them.","pith_inferences":["The paper does not pursue this, but the same density estimate could serve as an acquisition function: configurations with low prior density yet high feature likelihood mark regions where the surrogate most needs additional training simulations.","A natural extension, which the paper leaves to future work, is to decouple the surrogate-accuracy prior from the feature likelihood so that model uncertainty can be inspected separately from feature-based querying.","The framework should transfer to time-dependent or gradient-derived features whenever the feature can be written as a differentiable function of the neural field; the paper notes this but does not implement it.","One implicit empirical claim worth testing is that Fisher-information distance orders parameter-space proximity better than Euclidean distance for surrogate error; a direct comparison of the two metrics on the same datasets would settle it."],"forward_implications":["A user can specify a local feature, such as a vortex or rotation region, and immediately see every parameter configuration that approximately reproduces it, rather than a single optimized configuration.","Because the prior density is high only near training data, the collected samples are restricted to regions where the surrogate is likely to be accurate, so the visualization does not present extrapolated matches as credible.","Comparing the posterior for two different features highlights which parameter combinations distinguish one feature from the other, supporting feature-driven hypothesis generation.","The density prior gives a surrogate-accuracy signal that is cheaper than standard deep-ensemble or dropout uncertainty estimates while remaining predictive of error, enabling lower-latency interaction.","Because the sampler updates progressively after burn-in, the interface can show a coarse posterior quickly and refine it, making high-dimensional exploration practical in a visual loop."],"supporting_citations":[{"why":"Supplies the periodic-activation neural field architecture used as the surrogate model.","marker":"[39]"},{"why":"Frames inverse design as an optimization toward target outputs, which this paper contrasts with distributional posterior sampling.","marker":"[45]"},{"why":"Normalizing-flow surrogate that treats inversion as a field-level constraint, the approach this paper generalizes to local features.","marker":"[36]"},{"why":"INR-based surrogate inverse method that propagates uncertainty but returns only candidate matches, another direct point of contrast.","marker":"[7]"},{"why":"Provides the deep-ensemble uncertainty estimates that this paper compares its density estimate against.","marker":"[23]"},{"why":"Supplies the Hamiltonian Monte Carlo algorithm used to sample the posterior.","marker":"[29]"},{"why":"Supplies the Bayesian inference background and the R-hat diagnostic used to validate mixing.","marker":"[13]"},{"why":"Gives maximum mean discrepancy, the metric used to test HMC convergence against a long reference chain.","marker":"[16]"}],"fun_headline_variants":["Posterior sampling reveals all plausible inputs for a desired feature","Instead of one inverse answer, sample the whole plausible set","From one inverse solution to a distribution of plausible inputs","Parameter distributions beat single points for inverse feature queries","Sampling the posterior maps every plausible input for a given output"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that closeness to training parameters, measured with a Fisher-information distance (a distance that reflects how quickly simulated outputs change as parameters move) and nearest-neighbor bandwidths, is a reliable proxy for where the surrogate's predictions can be trusted; if that correspondence fails, the posterior will confidently sample parameters whose surrogate outputs do not match the true simulation.","fun_headline_variants_meta":{"raw":{"variants":["Posterior sampling reveals all plausible inputs for a desired feature","Instead of one inverse answer, sample the whole plausible set","From one inverse solution to a distribution of plausible inputs","Parameter distributions beat single points for inverse feature queries","Sampling the posterior maps every plausible input for a given output"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00096,"raw_usage":{"total_tokens":4095,"prompt_tokens":959,"completion_tokens":3136,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":3057}},"tokens_in":575,"tokens_out":3136,"duration_ms":19118,"temperature":1.0,"reasoning_tokens":3057,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:15:54.007627+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a feature located in a region where the surrogate is known to be inaccurate, for instance between two training configurations where the density map in Appendix B shows a low-density streak; run the full posterior sampler, then evaluate the high-density samples on the ground-truth simulator. If those samples frequently fail to reproduce the specified feature, the density prior has not captured surrogate error. A cheaper check is to compare the HMC sample distribution against a dense-grid posterior for the one-dimensional viscosity example and verify that the sample density matches the true posterior where the prior is peaked.","supporting_citations":[{"cited_title":"Sitzmann, J","cited_arxiv_id":null,"evidence_quote":"Supplies the periodic-activation neural field architecture used as the surrogate model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Normalizing-flow surrogate that treats inversion as a field-level constraint, the approach this paper generalizes to local features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"INR-based surrogate inverse method that propagates uncertainty but returns only candidate matches, another direct point of contrast."},{"cited_title":"Lakshminarayanan, A","cited_arxiv_id":null,"evidence_quote":"Provides the deep-ensemble uncertainty estimates that this paper compares its density estimate against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Hamiltonian Monte Carlo algorithm used to sample the posterior."},{"cited_title":"Gelman, J","cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian inference background and the R-hat diagnostic used to validate mixing."},{"cited_title":"Gretton, K","cited_arxiv_id":null,"evidence_quote":"Gives maximum mean discrepancy, the metric used to test HMC convergence against a long reference chain."}],"review_version":2}