{"id":"c5982ce0-c0ff-481c-8fcd-2d71aa2bf72f","arxiv_id":"2502.07169","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 3D machine-learning data assimilation framework tracks synthetic CO2 plumes with conditional normalizing flows on 128-cubed volumes.","lead":"This paper extends a 2D Digital Shadow monitoring framework for CO2 storage to 3D, using machine learning to estimate the CO2 plume from time-lapse seismic images. The demonstration is a single synthetic test case and claims calibrated uncertainty, but it lacks quantitative metrics and field validation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central claims of improved plume estimation and well-calibrated uncertainty are not supported by the reported evidence because the results rest on a single in-distribution synthetic test with no quantitative metrics or baselines.","rationale":"The reader's conditional verdict is appropriate and matches my reading of the paper. The central assertion is that the 3D Digital Shadow provides accurate and well-calibrated posterior plume estimates from 4D seismic data. The evidence, however, is a single synthetic test with no quantitative accuracy or calibration measures, no baseline comparison, and a test plume drawn from the same forward model used to create the training set. My main stress-test concern is the absence of any numerical evaluation; without that, 'significant improvements' and 'well-calibrated uncertainty' are not established. The reader's weakest assumption additionally identifies the in-distribution nature of the test, which I see as a related but slightly different issue: even strong quantitative metrics on the same simulator would not demonstrate transfer to a real site unless model robustness is tested. I therefore agree with the conditional verdict and propose a concrete numerical test that would settle the quantitative gap and probe sensitivity to model misspecification.","tokens_in":4286,"tokens_out":4998,"duration_ms":45577,"concrete_test":"Compute on at least 20 held-out plume realizations, drawn from the same simulator and from a perturbed forward model (e.g., velocity-permeability transform changed by ±20%, or a different colored-noise covariance), the posterior mean NRMSE and the empirical coverage of the 95% credible interval, and compare with a forecast-only baseline. If coverage is far below nominal or the posterior does not beat the baseline, the central claims fail; if metrics are strong on the same-model test but degrade on the perturbed model, the method is valid only for the exact assumed generative model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4, results are shown for one 'unseen' plume (Figure 1b). The evaluation is purely visual: no error norms, coverage statistics, or interval scores are reported, and no comparison is made to a forecast-only prior or any other baseline. Moreover, the test plume and all 128 training pairs are generated by the same flow simulator (JutulDarcy), the same seismic modeling package (JUDI), and the same Patchy saturation observation model, so the experiment only shows the network can invert the authors' own forward model on a sample from the same generative process. The 'well-calibrated' statement in Section 5 is based on a claimed visual correlation between Figures 1e and 1f, which is not a calibration metric. Consequently, the conclusions of 'significant improvements' and 'well-calibrated uncertainty quantification' are broader than the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 3D Digital Shadow framework for monitoring CO2 plumes in geological carbon storage, using conditional normalizing flows to approximate the posterior distribution of CO2 saturation given 4D seismic data. The authors simulate plume dynamics with JutulDarcy, model 4D seismic responses with JUDI and the Patchy saturation model, and train a 3D conditional normalizing flow in InvertibleNetworks.jl on 128 synthetic ensemble pairs. The method is tested on a single unseen synthetic plume from the Compass model, with results shown in Figure 1. The authors conclude that assimilating 4D seismic data significantly improves plume estimates and that the framework provides well-calibrated uncertainty quantification.","tokens_in":4490,"tokens_out":3639,"duration_ms":32603,"significance":"If fully supported, the work would be a valuable advance: it extends a 2D digital shadow to realistically sized 3D volumes, uses open-source and reproducible components (JutulDarcy, JUDI, InvertibleNetworks.jl), and addresses an application of societal importance. The stated novelty of a scalable 3D conditional normalizing flow and the potential for practical GCS monitoring are noteworthy. However, as presented, the evidence is insufficient to substantiate the headline claims of improved accuracy and well-calibrated uncertainty, so the significance is conditional on a substantially stronger evaluation.","major_comments":[{"comment":"The central claim of improved plume estimates rests on a single visual comparison. No quantitative accuracy metric (e.g., RMSE, structural similarity, or pixel-wise error norm) is reported for the inferred plume in Figure 1d against the ground truth in Figure 1b, and there is no comparison against a forecast-only prior or an alternative assimilation method such as an ensemble Kalman filter. The statement in Section 5 that incorporating 4D seismic data produced 'significant improvements' is therefore not supported by the reported evidence; please add quantitative error metrics over multiple test plumes and a baseline.","section":"Section 4, Figure 1"},{"comment":"The evaluation is in-distribution: the 128 training pairs and the single test plume are all generated by the same flow simulator (JutulDarcy), the same seismic simulation package (JUDI), and the same Patchy saturation observation model, and the permeability prior is produced by the authors' own full-waveform inversion approach. This experiment only demonstrates that the network can invert the exact forward model used to create the data. To support generalization claims, please test on out-of-distribution scenarios, such as different geological realizations, different injection rates or times, or perturbations to the observation model.","section":"Section 3 and Figure 1b"},{"comment":"The claim of 'well-calibrated uncertainty quantification, with strong correlations between inferred uncertainty and ground-truth errors' is based on a visual similarity between Figure 1e and Figure 1f. A calibration claim requires quantitative verification, such as empirical coverage of credible intervals, interval scores, or a correlation coefficient between uncertainty and absolute error computed across the volume or across an ensemble of test cases. Please include such diagnostics.","section":"Section 5, Conclusions"}],"minor_comments":[{"comment":"The L2 norm notation in the training objective, written as '||...|| 2 2 2', is garbled and should be cleaned up to a standard norm notation for a normalizing flow negative log-likelihood.","section":"Equation (3)"},{"comment":"The caption for panel (d) says 'A posterior sample', while Section 4 states that the mean of 128 posterior samples is shown; clarify which quantity is displayed in panel (d).","section":"Figure 1 caption"},{"comment":"The text says 'This abstract extends...' but the document is a full preprint; change 'abstract' to 'paper' or 'work'.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is very short and reads like an extended abstract, yet its conclusions include strong claims of significant improvement and well-calibrated uncertainty. The central evaluation is a single in-distribution synthetic test with visual inspection only, so the claims are not yet supportable. I recommend a major revision requiring quantitative metrics, baselines, and out-of-distribution tests."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a technically real extension of the authors' prior 2D Digital Shadow work to 3D, and the open-source stack is a clear plus. But the evaluation does not carry the weight of the conclusions. The central claims of improved plume estimates and well-calibrated uncertainty come from one synthetic test, with no quantitative error, no calibration metric, and no baseline comparison.\n\nThe genuinely new thing is scaling conditional normalizing flows to 128x128x128 volumes, which required memory-frugal invertible layers. That is a legitimate engineering contribution, and the use of JutulDarcy, JUDI, and InvertibleNetworks.jl means others can reproduce the pipeline. If you work on Bayesian inverse problems for CO2 monitoring, this is a useful data point that 3D amortized inference is feasible.\n\nThe soft spots are mostly in Section 4. The test is a single unseen plume from the same simulator family used to generate the training pairs. The 'well-calibrated' statement is based on a visual correlation between error and uncertainty, not a proper calibration score like coverage or interval score. There is no comparison to a forecast-only prior, an ensemble Kalman filter, or even the 2D version. And the seismic noise level, the velocity-permeability transform, and the 3% injection cap are all fixed choices; we do not know how sensitive the result is to them. So the paper demonstrates proof-of-concept, not 'significant improvements' in a general sense.\n\nThe circularity concern is worth taking seriously, but I would phrase it as in-distribution evaluation rather than a fatal flaw. The network is learning to invert the authors' own forward model, which is a legitimate first step, but it means the calibration claims are only about the simulator, not about a real site.\n\nBottom line: worth a serious referee, but the referee should push for quantitative metrics, baselines, and an out-of-distribution or partially field-like test. I would not cite it yet in my own work, but I would read the revised version.","headline":"A real but incremental 3D extension of the authors' 2D Digital Shadow; however, the single synthetic test with no quantitative metrics or baselines is too thin to support the claims of well-calibrated uncertainty.","tokens_in":5020,"tokens_out":2073,"would_cite":false,"duration_ms":18071,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["86A22","68T07","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A conditional normalizing flow trained on simulated 3D reservoir and seismic data estimates CO2 plume state and uncertainty from 4D seismic measurements, marking the first generative model applied to realistically sized 3D volumes.","keywords":["geological carbon storage","CO2 plume monitoring","4D seismic","conditional normalizing flows","Bayesian inverse problems","uncertainty quantification","digital shadow","time-lapse seismic"],"falsifier":"Take the trained 3D normalizing flow and apply it to a plume produced by a different multiphase flow simulator, or to a permeability field drawn from a prior distribution that lies outside the training range, and compare the posterior mean and uncertainty to the ground truth. A simpler and stronger test would be to condition the network on field 4D seismic data from an operating CO2 storage site, such as Sleipner, and check whether the posterior uncertainty brackets the independently known plume extent. If the error-vs-uncertainty correlation breaks in either test, the calibration claim would be falsified.","tokens_in":4091,"feed_emoji":"💨","tokens_out":7937,"duration_ms":61836,"temperature":0.7,"pith_summary":"This paper addresses a practical need in geological carbon storage: knowing, with quantified uncertainty, where injected CO2 is spreading underground. The authors extend an existing 2D 'digital shadow' monitoring framework to three dimensions, replacing a 2D generative model with a conditional normalizing flow that maps 4D seismic images to a posterior distribution over a 128x128x128 CO2 saturation volume. The network is trained entirely on simulated data--128 pairs of forecasts from multiphase flow simulations and the corresponding synthetic time-lapse seismic surveys--and then tested on an unseen simulated plume. The paper claims that the assimilated posterior mean matches the true plume's shape and size, and that the posterior uncertainty is well calibrated, meaning the model's error is high where its uncertainty is high. If this approach transfers to real field data, it would provide a practical tool for monitoring CO2 storage sites and certifying containment.","feed_headline":"3D generative model maps CO2 plumes from seismic data","feed_subtitle":"A normalizing-flow 'digital shadow' delivers plume shape, size, and calibrated uncertainty in realistic 3D volumes.","key_machinery":"The load-bearing object is the conditional normalizing flow: an invertible neural network that transforms a latent Gaussian distribution into the plume state while conditioning on the seismic observation. Its memory-frugal invertible layers are what make a 128x128x128 generative model tractable, and its exact log-density enables direct maximum-likelihood training on the simulated pairs. The surrounding simulation-based Bayesian filtering loop generates the training data: multiphase flow simulation advances the plume state, time-lapse seismic simulation produces observations with colored noise, and the flow is trained to approximate the posterior p(plume | seismic). The permeability fields, which drive the forecast ensemble, come from a probabilistic full-waveform inversion of a baseline survey, converted to permeability through an empirical relationship.","core_discovery":"The central claim is that a 3D conditional normalizing flow, trained on simulation pairs, solves the Bayesian inverse problem of reconstructing CO2 saturation from 4D seismic data at reservoir scale. Each training pair couples a forecasted saturation plume, produced by multiphase flow simulation under randomly sampled permeability fields, with the corresponding synthetic 4D seismic image at 24 Hz with added colored noise. After training, the network turns a seismic observation into a full posterior distribution over the plume volume, from which a mean map and a standard-deviation map are computed. On the unseen test plume, the posterior mean closely tracks the ground truth, and the posterior standard deviation is strongly correlated with the absolute error, which the authors take as evidence of well-calibrated uncertainty. This constitutes, in the paper's words, the first known application of generative modeling to realistically sized 3D volumes.","pith_inferences":["An unreported stress test would be to train on permeability samples from one geological trend and evaluate on samples from a different trend; this would reveal how quickly the posterior calibration degrades when the prior is misspecified.","The fixed, sparse acquisition geometry used in the synthetic surveys (16 sources, 16,000 receivers) leaves open whether the learned posterior remains calibrated when the acquisition geometry changes between surveys, a common field condition that could be tested with simulated geometry variations.","The fact that 128 training pairs suffice for a 3D generative model suggests the simulator encodes a strong physical prior; this could be exploited to build site-specific digital shadows with even smaller ensembles, or to fine-tune from a pre-trained base model.","The stated path from 'digital shadow' to 'digital twin' requires a control loop that adjusts injection operations based on the posteriors; the paper demonstrates the statistical input to that loop, but not the loop itself."],"forward_implications":["Operators could turn each repeated 3D seismic survey into a volumetric CO2 saturation map with explicit uncertainty, directly supporting leak detection and storage certification.","The demonstration that normalizing flows scale to realistically sized 3D volumes opens the same conditional generative approach to other 3D subsurface inverse problems, such as reservoir characterization or monitoring of other injected fluids.","If the uncertainty is genuinely well calibrated, the posterior standard deviation can be used as a quantitative error estimate, which is a prerequisite for risk-based decisions in carbon storage.","Because the training data are purely simulated, the framework can in principle be rebuilt for a new storage site by re-running the same simulation pipeline on that site's geological prior, without needing labeled field data."],"supporting_citations":[{"why":"Supplies the 2D Digital Shadow formulation and the Patchy saturation model that the 3D approach extends.","marker":"Gahlot, Orozco, et al. 2024"},{"why":"The multiphase flow simulator that generates the 128 training plumes.","marker":"Møyner, Bruer, and Yin 2023"},{"why":"The seismic simulation and imaging software that produces the 4D observations used in training and testing.","marker":"Witte et al. 2019; Louboutin et al. 2023"},{"why":"The package whose memory-frugal invertible layers enable the first scalable 3D conditional normalizing flow.","marker":"Orozco, Witte, et al. 2024"},{"why":"Provides the conditional normalizing flow methodology used for neural posterior density estimation.","marker":"Papamakarios et al. 2021"},{"why":"Supplies the nonlinear ensemble Bayesian filtering framework into which the neural posterior estimation is inserted.","marker":"Spantini, Baptista, and Marzouk 2022"},{"why":"Provides the Compass model from which the synthetic 3D North Sea reservoir subset is taken.","marker":"E. Jones et al. 2012"},{"why":"The probabilistic full-waveform inference used to initialize the permeability distribution for the ensemble.","marker":"Orozco, Siahkoohi, et al. 2024"}],"fun_headline_variants":["AI turns seismic data into 3D CO2 plume maps","Normalizing flow tracks CO2 in 3D from seismic","Digital shadow pinpoints CO2 plumes in 3D","Generative AI maps carbon storage plumes in 3D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The validation plume is generated by the same simulation pipeline (flow simulator, seismic simulator, and observation model) used to create the 128 training pairs, so the demonstrated accuracy and calibration are only guaranteed with respect to the authors' own synthetic world; if the simulator or the underlying velocity-permeability transform misrepresents a real storage site, the performance claims would not transfer.","fun_headline_variants_meta":{"raw":{"variants":["AI turns seismic data into 3D CO2 plume maps","Normalizing flow tracks CO2 in 3D from seismic","Digital shadow pinpoints CO2 plumes in 3D","Generative AI maps carbon storage plumes in 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1368,"prompt_tokens":861,"completion_tokens":507,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":436}},"tokens_in":477,"tokens_out":507,"duration_ms":4181,"temperature":1.0,"reasoning_tokens":436,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:34:26.136889+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained 3D normalizing flow and apply it to a plume produced by a different multiphase flow simulator, or to a permeability field drawn from a prior distribution that lies outside the training range, and compare the posterior mean and uncertainty to the ground truth. A simpler and stronger test would be to condition the network on field 4D seismic data from an operating CO2 storage site, such as Sleipner, and check whether the posterior uncertainty brackets the independently known plume extent. If the error-vs-uncertainty correlation breaks in either test, the calibration claim would be falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The multiphase flow simulator that generates the 128 training plumes."},{"cited_title":"Gorman, and Felix J","cited_arxiv_id":null,"evidence_quote":"The seismic simulation and imaging software that produces the 4D observations used in training and testing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the conditional normalizing flow methodology used for neural posterior density estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the nonlinear ensemble Bayesian filtering framework into which the neural posterior estimation is inserted."}],"review_version":1}