{"id":"a68fe76b-38d0-420f-9a7c-5d29c974adad","arxiv_id":"2504.14405","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A conditional normalizing flow digital shadow that conditions on the Brie rock physics exponent enables real-time sensitivity analysis of CO2 plume forecasts without retraining.","lead":"This paper adds a 'context' switch to a machine-learning monitoring system for underground CO2 storage, letting operators ask what-if questions about rock physics assumptions without retraining. The method is demonstrated on a synthetic North Sea reservoir model, where conditioning on the correct rock physics model improves plume forecasts.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The what-if claim rests on untested interpolation: the CNF is trained on only six discrete Brie exponents (§3.2–3.3) but must generalize to arbitrary e in (1,10); §4 evaluates only e=4 and e=9.","rationale":"The reader's weakest assumption identifies the same gap: the conditional flow is trained on six discrete contexts despite the continuous e ~ U(1,10) sampling stated in §3.3, and the experiments evaluate only two contexts, both likely seen during training. This is not merely a missing baseline; it is the load-bearing condition for the central 'what-if without retraining' claim. If the network does not interpolate reliably to unseen Brie exponents, the method reduces to a memorized lookup table over six contexts, and the advertised sensitivity analysis over a continuous range of rock physics models is unsupported. The paper is otherwise internally consistent and transparent about being a proof of concept, so the CONDITIONAL verdict remains appropriate. The proposed test would settle the concern directly by evaluating held-out contexts against a reference model. No additional objections to the core methodology were identified.","tokens_in":6345,"tokens_out":4546,"duration_ms":44441,"concrete_test":"Hold out e=5 and e=7 from the six training contexts (report the full set), retrain the CNF on the remaining contexts, and generate observed seismic data with e=5. Condition the trained network on e=5 and compare the resulting posterior mean, SSIM, and RMSE against (i) a reference CNF trained only on e=5 and (ii) the data-residual minimum. If the held-out context achieves SSIM comparable to the reference and the residual is minimized near e=5, the amortization over unseen contexts is supported; if not, the what-if claim must be restricted to trained contexts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a single network can be conditioned at inference on any rock-physics context, enabling sensitivity analysis without retraining. For this to hold, the conditional normalizing flow must interpolate (and perhaps extrapolate) across the Brie-exponent range. In §3.2 the training set is built from exactly six discrete exponent values, and §3.3 states e ~ U(1,10) while the outer expectation over e is implemented only over these six contexts. The six values are not reported, and §4 tests only e=4 and e=9, both presumably in the training set. The observed SSIM/RMSE improvement at e=4 therefore demonstrates self-consistency on seen contexts, not the advertised what-if capability on unseen ones. If the network merely memorizes context-specific posterior means, then conditioning on an unseen exponent would produce an unreliable posterior and the main contribution collapses to a discrete lookup table rather than a continuous sensitivity tool. This is the weakest load-bearing assumption: the paper's conclusion 'without requiring retraining' depends entirely on this untested generalization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a context-aware Digital Shadow (DS) for monitoring underground energy storage, specifically tracking CO2 saturation from time-lapse seismic data. The key idea, borrowed from sensitivity-aware amortized Bayesian inference, is to train a conditional normalizing flow (CNF) that takes the rock-physics model as an additional conditioning variable, represented by the Brie saturation exponent e. The network is trained on a synthetic 2D Compass-based reservoir model with 128 flow realizations and six different rock-physics contexts, yielding 768 augmented training samples. At inference time the network can be conditioned on a specific e, enabling 'what if' sensitivity analysis without retraining. Results on synthetic data show that conditioning on the correct exponent e=4 gives higher SSIM, lower RMSE, and lower posterior standard deviation than conditioning on the incorrect exponent e=9, and the data-residual curve has a minimum at e=4.","tokens_in":6533,"tokens_out":4152,"duration_ms":37381,"significance":"If the central claim is validated, the paper would make a useful practical contribution: it would allow real-time sensitivity analysis of CO2 plume forecasts with respect to rock-physics assumptions, avoiding both the averaging effect of a marginal posterior and the cost of retraining a network for each rock-physics model. The authors correctly identify a real limitation of the earlier marginalizing DS approach and connect it to a principled amortized-inference framework. The synthetic experiments use open-source simulation tools (JutulDarcy.jl, JUDI.jl, InvertibleNetworks.jl), which is a strength for reproducibility. The main weakness is that the evaluation only tests the two training contexts e=4 and e=9, so the advertised 'what if' capability for arbitrary rock-physics models is not actually demonstrated. The paper would be strengthened substantially by held-out-context experiments and a baseline comparison with the marginalizing DS from prior work.","major_comments":[{"comment":"The manuscript repeatedly refers to '6 different contexts' but never reports the six Brie exponent values or how they were chosen. Since the outer expectation in Eq. (3) is over e ~ U(1,10) but is implemented with only these six samples, the effective training distribution is a six-point discrete distribution, not a continuous uniform range. The authors should report the six values and justify that they adequately cover [1,10]; otherwise the claim of amortization over a continuous range is unsupported.","section":"§3.2–3.3"},{"comment":"The evaluation only considers e=4 and e=9, and both appear to be among the six training contexts. The paper's central promise is 'what if' sensitivity testing at inference time for arbitrary rock-physics assumptions 'without requiring retraining'. To substantiate this, the authors need to evaluate held-out exponents (for example, e=2.5, e=6, or e=8) and report the resulting SSIM, RMSE, and posterior standard deviation. If performance degrades sharply for interpolated values, the method reduces to a lookup table over six discrete contexts rather than a continuous sensitivity tool.","section":"§4, Figures 3–6"},{"comment":"The observed time-lapse data are generated with e=4 using the same simulation family that produced the training contexts, so the finding that the residual minimum occurs at e=4 is partly a self-consistency check. The authors should also test with observed data generated under a held-out rock-physics model or with an exponent not included in the six training contexts. In addition, a direct comparison with the marginalizing DS of Gahlot and Herrmann (2025) is needed to show that conditioning on the correct context yields systematically better predictive performance rather than merely a different posterior mean.","section":"§4, Figure 3"}],"minor_comments":[{"comment":"The Brie saturation model is not defined; since the exponent e is the central context variable, the authors should include the Brie mixing formula or provide the exact equations used to convert CO2 saturation and pressure to elastic properties, with a specific reference to the equation numbers in Avseth et al. (2010).","section":"§3.2"},{"comment":"The loss expression in Eq. (3) has typographical issues (triple vertical bars around the flow mapping) and ambiguous notation; please define f_phi explicitly as the forward or inverse transformation and write the standard change-of-variables negative log-likelihood in a cleaner form, stating which network outputs form the affine coupling layers.","section":"§3.3, Eq. (3)"},{"comment":"The text says that 'the captions of Figure 5 and figure 6 show' the RMSE and standard deviation values, but the corresponding numerical values are not given in the text; please report the exact SSIM, RMSE, and mean standard deviation values for both the correct and incorrect exponent cases in the main text.","section":"§4"},{"comment":"The discussion of Figure 1 and Figure 2 is confusing: the text refers to 'the right plot of figure 2' when describing the ground-truth plume and to the left/right in Figure 1; please clarify the figure references and ensure the captions explicitly identify which panel corresponds to the uniform and patchy rock-physics models.","section":"§1"},{"comment":"The phrase '28 dB SNR colored Gaussian noise' should be rephrased as 'additive colored Gaussian noise with 28 dB SNR', and the coloring (e.g., correlation length or frequency content) should be specified for reproducibility.","section":"§3.2"},{"comment":"The statement 'The overall readability is enhanced using ChatGPT 4' is unconventional; if the journal requires generative-AI disclosure, please format it as a standard AI-use statement and place it in the appropriate section according to the journal's policy.","section":"Acknowledgments"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the proposed application is timely, but the novelty relative to Elsemüller et al. (2024) and Gahlot et al. (2024) should be made explicit: the contribution is an application of sensitivity-aware amortized Bayesian inference to the Digital Shadow framework. The most important technical risk is that the trained CNF may have memorized the six discrete training contexts; the authors must add held-out-context experiments to rule this out. I would also encourage the editor to require a comparison with the marginalizing DS from Gahlot and Herrmann (2025), since without that baseline it is difficult to assess whether the extra complexity of the context-aware approach is justified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this paper gives the DS community a clean way to condition a trained network on the Brie rock-physics exponent, so you can ask 'what if the saturation model were patchy?' without retraining. The idea is a direct and sensible extension of Elsemüller et al.'s sensitivity-aware amortized inference and of Gahlot & Herrmann's marginalizing DS. What it doesn't yet do is show the what-if claim actually works for arbitrary exponents.\n\nThe paper does a few things well. The motivation is clear: marginalizing over rock physics blurs uncertainty sources, and retraining per model is expensive. Conditioning the CNF by adding a sinusoidal-embedded e to each activation layer is a reasonable architectural choice. The synthetic test uses realistic tools (Compass model, JutulDarcy, JUDI, InvertibleNetworks) and the paper is properly transparent about being a proof of concept. Citations are appropriate.\n\nThe soft spots are concentrated in validation. The six training contexts are never listed. The evaluation only compares e=4 and e=9, both very likely in the training set. So the observed SSIM/RMSE improvement at e=4 is a self-consistency check: the network is told the context that generated the data, and it does better than when told a wrong context. That is reassuring but does not support the conclusion 'without requiring retraining' for unseen contexts. The network may be memorizing context-specific posterior means rather than interpolating smoothly, and the paper gives us no way to tell. This is the load-bearing gap, and it matches the stress-test note.\n\nTwo smaller gaps: no baseline comparison to the marginalizing DS from the authors' own prior work, so we don't see what conditioning buys over simple averaging. And no code or data release, though the open-source stack makes reproduction feasible in principle.\n\nNone of this is fatal. The method is plausible and the gap is fixable: report the six exponents, test on at least one held-out value (say e=6 or e=2), add the marginal-DS baseline, and consider a small ablation on the embedding. I'd send this to peer review with those requests. A serious referee can push the authors to close the loop; the core idea deserves that push.\n\nMy verdict: worth reviewing, not yet worth citing as validated. Bring it to reading group if you want a concrete case study of the gap between what-if framing and interpolation evidence.","headline":"A clean but under-tested extension: conditioning the Digital Shadow on the Brie exponent is a real idea, but the current experiments only show self-consistency on seen contexts, not the advertised what-if generalization.","tokens_in":7080,"tokens_out":3623,"would_cite":false,"duration_ms":33318,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single trained Digital Shadow can be conditioned on any assumed rock physics model at inference time, enabling real-time what-if testing for CO2 storage monitoring.","keywords":["Digital Shadow","CO2 storage monitoring","amortized Bayesian inference","conditional normalizing flows","Brie saturation model","rock physics uncertainty","time-lapse seismic","sensitivity analysis"],"falsifier":"Generate a test seismic image with an untrained Brie exponent, for example e=6.5 or e=2, condition the trained Digital Shadow on that same exponent, and compare the posterior mean and spread against a reference network retrained specifically at that exponent; if the context-conditioned result is no closer to the ground-truth plume than a wrong-context run, or if its uncertainty is not calibrated, the generalization claim fails.","tokens_in":6129,"feed_emoji":"🛢️","tokens_out":6677,"duration_ms":54741,"temperature":0.7,"pith_summary":"Time-lapse seismic monitoring of underground CO2 storage relies on a rock physics model to translate fluid-flow states into seismic images, and predictions break down when that model is wrong. This paper's proposal is to fold the rock physics model itself into the training of a Digital Shadow—a neural-network Bayesian data-assimilation system—as an explicit context variable, so the network learns a joint mapping between seismic data, CO2 saturation, and the rock physics model. At inference time, the same trained network can then be asked what would happen if the Brie exponent were 4 versus 9, and it produces a posterior that is conditioned on that choice rather than averaged over all possible choices. The paper argues this gives real-time sensitivity analysis for underground energy storage without retraining, and demonstrates on synthetic reservoir models that conditioning on the correct exponent yields a more accurate and less uncertain CO2 plume forecast.","feed_headline":"One network runs CO2 storage what-if tests without retraining","feed_subtitle":"Context-aware training lets the Digital Shadow adapt to any assumed rock physics model on the fly.","key_machinery":"The carrying mechanism is a conditional normalizing flow whose conditioning variable is not just the seismic image but also an embedded encoding of the rock physics context. The Brie saturation exponent e, which interpolates between uniform and patchy fluid mixing, is drawn from a uniform distribution, passed through a sinusoidal embedding followed by a dense layer, and added to the intermediate activations of the conditional arm of the network. Training minimizes a context-averaged negative log-posterior objective across e, so a single set of network weights is amortized over rock physics contexts. At inference the same weights are evaluated with a fixed e, which is what turns the posterior into a context-aware conditional distribution rather than a marginal one. The surrounding machinery is the Digital Shadow pipeline: multiphase flow simulation with random permeability fields produces an ensemble of CO2 saturation states, seismic simulation converts them to time-lapse images under each context, and the network is trained on those state, seismic, and context tuples.","core_discovery":"The central claim is that a Digital Shadow trained with a rock physics model as a context variable learns an explicit mapping between time-lapse seismic data, CO2 saturation, and the rock physics model, so that at inference time the network can be conditioned on a specific rock physics model and produce a context-aware posterior. The paper contrasts this with earlier augmentation that mixed multiple rock physics models into the training set, which produces a posterior marginalized over all models and thereby hides the source of uncertainty. Using the Brie saturation model family, with exponent e drawn uniformly from (1,10) and six discrete contexts used to augment 128 flow-simulation ensemble members, the network is trained as a conditional normalizing flow with a sinusoidal embedding of e added into the conditional arm. In the synthetic two-dimensional test, when the observed seismic data are generated with e=4, conditioning the trained network on e=4 yields a higher Structural Similarity Index, a lower root-mean-square error, and tighter posterior samples than conditioning on e=9. The paper reads this as evidence that the Digital Shadow can adapt to different rock physics assumptions dynamically and answer what-if questions without retraining.","pith_inferences":["The same context-embedding design could be inverted at inference time: instead of conditioning on a known exponent, one could treat e as an unknown and infer it jointly with saturation, turning the network into a tool for identifying which rock physics model the seismic data actually support.","Generalization beyond the six training contexts is the central open risk; a natural stress test would be to train on the same six contexts but evaluate on several intermediate exponents and on values near the endpoints of (1,10) to map where the context encoding stops interpolating.","If the method holds up on field-like data, an operator could maintain one model plus a scalar context rather than a family of separate models, and the question of which rock physics model is right becomes an empirical test on the observed seismic response."],"forward_implications":["A single trained Digital Shadow can be evaluated under any supplied Brie exponent at inference time, giving operators a posterior for CO2 saturation that reflects that specific rock physics assumption.","The demonstrated gap between conditioning on e=4 and e=9 shows that the context variable materially changes the forecast, so using the wrong rock physics model remains detectable as poorer agreement instead of being silently absorbed into the posterior.","What-if scenario testing for rock physics assumptions can be run without training a new network, reducing the cost of sensitivity analysis relative to per-model refitting.","The method extends to any reservoir state estimated by the Digital Shadow framework, such as pressure perturbations, not only CO2 saturation.","Because the posterior is context-aware, uncertainty from rock physics choice can be separated from uncertainty due to permeability and seismic noise, aiding interpretation of monitoring results."],"supporting_citations":[{"why":"Supplies the coupling techniques for nonlinear ensemble filtering on which the Digital Shadow's Bayesian data assimilation is built.","marker":"Spantini, Baptista, and Marzouk 2022"},{"why":"Original uncertainty-aware Digital Shadow that contributes the training objective and simulation setup extended here with context conditioning.","marker":"Gahlot et al. 2024"},{"why":"The prior approach that augments training with multiple rock physics models, whose marginal posterior this paper identifies as limiting and replaces with context-aware conditioning.","marker":"Gahlot and Herrmann 2025"},{"why":"Source of sensitivity-aware amortized Bayesian inference with context variables, which motivates the context-conditioned posterior design.","marker":"Elsemüller et al. 2024"},{"why":"Defines the Brie saturation model family whose exponent e is the rock physics context used throughout the experiments.","marker":"Avseth, Mukerji, and Mavko 2010"},{"why":"Establishes time-lapse seismic monitoring of CO2 sequestration as the observation modality the Digital Shadow conditions on.","marker":"Lumley 2010"},{"why":"Provides the sinusoidal Fourier-feature embedding used to encode the Brie exponent into the conditional network.","marker":"Tancik et al. 2020"},{"why":"The flow simulator that generates the CO2 saturation ensemble used for training and testing.","marker":"Møyner, Bruer, and Yin 2023"},{"why":"The seismic modeling and imaging framework that turns saturations into time-lapse seismic images under each rock physics context.","marker":"Witte et al. 2019"}],"fun_headline_variants":["Digital Shadow runs CO2 what-if tests without retraining","Rock-physics-aware AI adapts on the fly for storage monitoring","Context-aware seismic AI answers what-if for underground storage","No retraining: Digital Shadow adapts to rock physics models","Seismic Digital Shadow adjusts to rock physics in real time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The Digital Shadow is trained on only six discrete values of the Brie exponent, the rock-physics parameter that links CO2 saturation to seismic response, and is trusted to work for any value in the claimed range (1,10), even though only two values are demonstrated.","fun_headline_variants_meta":{"raw":{"variants":["Digital Shadow runs CO2 what-if tests without retraining","Rock-physics-aware AI adapts on the fly for storage monitoring","Context-aware seismic AI answers what-if for underground storage","No retraining: Digital Shadow adapts to rock physics models","Seismic Digital Shadow adjusts to rock physics in real time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1266,"prompt_tokens":935,"completion_tokens":331,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":248}},"tokens_in":551,"tokens_out":331,"duration_ms":3584,"temperature":1.0,"reasoning_tokens":248,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:49:27.526327+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a test seismic image with an untrained Brie exponent, for example e=6.5 or e=2, condition the trained Digital Shadow on that same exponent, and compare the posterior mean and spread against a reference network retrained specifically at that exponent; if the context-conditioned result is no closer to the ground-truth plume than a wrong-context run, or if its uncertainty is not calibrated, the generalization claim fails.","supporting_citations":[{"cited_title":"Enhancing Robustness Of Digital Shadow For CO2 Storage Monitoring With Augmented Rock Physics Modeling","cited_arxiv_id":"2502.07171","evidence_quote":"The prior approach that augments training with multiple rock physics models, whose marginal posterior this paper identifies as limiting and replaces with context-aware conditioning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the sinusoidal Fourier-feature embedding used to encode the Brie exponent into the conditional network."}],"review_version":1}