{"id":"6c3f5ebd-5bc8-4053-8de0-ed7e121c5d60","arxiv_id":"2606.07215","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":3.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Comparative benchmark finds U-FNO most accurate for CO2 saturation and FNO for pressure predictions in 2D heterogeneous reservoir models, with performance varying by PDE type.","lead":"This paper benchmarks deep learning models including U-Net, FNO, and U-FNO as surrogate models to speed up simulations of CO2 injection into underground reservoirs. A smart generalist might read it to see which AI tools could enable faster planning and risk assessment for carbon storage projects.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Performance difference attributed to PDE type (hyperbolic saturation vs elliptic pressure) without isolating PDE class from other field properties in the coupled system.","rationale":"Reader correctly flags limited representativeness of the 2-D benchmark for real 3-D GCS, but that affects external validity more than internal support for the PDE-type dependence claim. The proposed concrete test directly checks whether the causal link asserted in the strongest claim holds when PDE class is isolated.","tokens_in":1784,"tokens_out":317,"duration_ms":13946,"concrete_test":"Generate two additional datasets on the same 2-D heterogeneous domain: (1) steady single-phase Darcy flow (pure elliptic) and (2) linear advection with fixed velocity (pure hyperbolic). Retrain FNO and U-FNO with identical hyperparameters and data volume; if the relative accuracy ordering does not match the original claim, the PDE-type attribution is not supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim states that surrogate performance 'is strongly dependent on the underlying PDE type'. In the single-wellbore CO₂ injection benchmark both fields are generated by the same multiphase system; saturation fronts are discontinuous while pressure is smoother. No experiments decouple the equations (e.g., pure Darcy flow or pure advection) or hold field statistics fixed while changing only the operator type. Consequently the observed ranking (U-FNO for saturation, FNO for pressure) could arise from output regularity, discontinuity handling, or loss weighting rather than the hyperbolic/elliptic distinction itself.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.3","summary":"The manuscript performs a comparative evaluation of deep learning surrogate models (U-Net, V-Net, Temporal Convolutional Networks, FNO, and U-FNO) for accelerating numerical simulations of geological carbon sequestration. Using a 2D single-wellbore CO₂ injection benchmark with heterogeneous anisotropic permeability/porosity fields, the study reports relative performance in accuracy, training speed, memory use, robustness, and generalization for predicting transient pressure build-up (elliptic) and CO₂ saturation (hyperbolic) fields, concluding that model ranking depends strongly on PDE type with U-FNO best for saturation and FNO best for pressure.","tokens_in":1900,"tokens_out":536,"duration_ms":12342,"significance":"If the empirical rankings prove robust under proper statistical validation and controlled isolation of factors, the work would offer practical guidance for selecting operator-learning architectures in subsurface flow surrogates, directly supporting real-time optimization and uncertainty quantification in GCS digital twins. The explicit inclusion of computational costs (memory, training time) alongside accuracy is a positive feature that strengthens applicability.","major_comments":[{"comment":"Abstract: The central claim that 'surrogate model performance is strongly dependent on the underlying PDE type (i.e., hyperbolic vs. elliptic)' is not load-bearing supported by the described experiments. Both fields are generated by the same coupled multiphase system; no ablation holds field statistics fixed while varying only the operator class (e.g., pure Darcy vs. pure advection problems), so the observed U-FNO vs. FNO ranking could equally reflect output regularity, discontinuity handling, or loss weighting rather than the hyperbolic/elliptic distinction.","section":"Abstract"},{"comment":"Abstract / benchmark description: The 2D single-wellbore injection problem is presented as representative of 'high-dimensional transient subsurface flow problems,' yet no scaling studies or comparisons to 3D heterogeneous cases are referenced to substantiate this; this assumption directly underpins the generalization claims for real GCS applications.","section":"Abstract"}],"minor_comments":[{"comment":"Abstract: Dataset sizes, number of training realizations, error metrics (e.g., relative L2, MAE), statistical significance tests, and validation procedures are not reported, making it impossible to assess whether the stated performance differences are statistically meaningful.","section":"Abstract"},{"comment":"The manuscript should clarify whether the same loss weighting and hyperparameter tuning protocol was used across all architectures when comparing saturation versus pressure predictions.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on our comparative study of deep learning surrogates for geological carbon sequestration. We address the two major comments below and will make revisions to strengthen the manuscript.","responses":[{"response":"We agree that the experiments are performed on the coupled multiphase flow system and do not include controlled ablations that isolate the PDE type while holding other factors fixed. The observed performance differences between pressure and saturation predictions could indeed arise from other characteristics of the output fields. We will revise the abstract and discussion sections to remove the strong causal claim of dependence on PDE type and instead report the empirical finding that U-FNO performed best on saturation while FNO performed best on pressure, with a note that further work would be needed to isolate the underlying cause.","revision_made":"yes","referee_comment":"[Abstract] Abstract: The central claim that 'surrogate model performance is strongly dependent on the underlying PDE type (i.e., hyperbolic vs. elliptic)' is not load-bearing supported by the described experiments. Both fields are generated by the same coupled multiphase system; no ablation holds field statistics fixed while varying only the operator class (e.g., pure Darcy vs. pure advection problems), so the observed U-FNO vs. FNO ranking could equally reflect output regularity, discontinuity handling, or loss weighting rather than the hyperbolic/elliptic distinction."},{"response":"The study is conducted on a 2D benchmark with heterogeneous fields to enable controlled comparison of architectures under varying permeability, porosity, and injection conditions. We acknowledge that no scaling studies or 3D results are presented. We will revise the abstract and introduction to describe the setup explicitly as a 2D heterogeneous reservoir benchmark and qualify any generalization statements accordingly, noting extension to 3D as future work.","revision_made":"yes","referee_comment":"[Abstract] Abstract / benchmark description: The 2D single-wellbore injection problem is presented as representative of 'high-dimensional transient subsurface flow problems,' yet no scaling studies or comparisons to 3D heterogeneous cases are referenced to substantiate this; this assumption directly underpins the generalization claims for real GCS applications."}],"tokens_in":1474,"tokens_out":465,"duration_ms":14224,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper runs U-Net, V-Net, TCN, FNO, and U-FNO on a 2D single-well CO2 injection benchmark with heterogeneous anisotropic permeability and porosity. It reports U-FNO strongest on saturation fields and FNO strongest on pressure, then links the difference to the underlying PDE character.\n\nNothing here is new in method or theory. The architectures are taken as-is from prior work, and the task is a standard multiphase flow surrogate problem. The value, if any, is in the side-by-side numbers on this specific setup.\n\nThe comparison itself is the part that could be useful. Practitioners choosing a starting model for similar reservoir surrogates might get a quick sense of which architecture to try first for saturation versus pressure outputs. If the full text supplies the missing dataset sizes, training protocols, error metrics, and statistical checks that the abstract leaves out, the results could serve as a reference point.\n\nThe main weakness is the interpretation. The claim that performance depends strongly on PDE type rests on one coupled system where saturation fronts are sharp and pressure is smoother. No runs isolate the operator class while holding field statistics and coupling fixed, so the ranking could stem from discontinuity handling, loss weighting, or output regularity instead. The 2D single-well case is also narrow compared with the high-dimensional transient problems the abstract mentions, which limits how far the ranking can be trusted.\n\nThis is for groups already working on DL surrogates for subsurface flow who need a quick empirical check on these architectures. It does not advance the methods, but a well-documented benchmark can still be worth referee time if the data and code are released.\n\nI would send it to peer review. The core experiment is simple enough to evaluate once the details are in front of a referee, even if the PDE-type explanation needs more support.","headline":"A plain benchmark comparing off-the-shelf DL models on a 2D CO2 injection case, with an unisolated claim that performance tracks hyperbolic versus elliptic PDE type.","tokens_in":2356,"tokens_out":452,"would_cite":false,"duration_ms":13449,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Deep learning surrogate accuracy for carbon sequestration varies strongly with whether the target field follows a hyperbolic or elliptic PDE.","keywords":["deep learning","surrogate models","geological carbon sequestration","Fourier neural operators","subsurface flow","CO2 saturation","pressure build-up","U-FNO"],"falsifier":"Repeating the identical architecture comparison on a 3D heterogeneous reservoir model or on a multi-well injection scenario and checking whether U-FNO and FNO retain their respective top rankings for saturation and pressure.","tokens_in":2686,"feed_emoji":"🌍","tokens_out":713,"duration_ms":19592,"temperature":0.7,"pith_summary":"The paper compares five deep learning architectures as fast surrogates for expensive numerical reservoir simulations of CO2 injection. It runs the models on a 2D single-wellbore problem whose permeability, porosity, injection rate, and reservoir properties are varied across many realizations. Results show that no single architecture wins on both outputs: the U-FNO variant records the lowest error on CO2 saturation fields, while the plain FNO records the lowest error on pressure build-up. The authors attribute the split to the different mathematical character of the saturation transport equation versus the pressure equation. Because real-time optimization and uncertainty quantification for carbon storage require many repeated solves, knowing which architecture matches each PDE type directly affects which surrogate can be deployed at scale.","feed_headline":"U-FNO best for saturation, FNO best for pressure in CO2 surrogates","feed_subtitle":"Benchmark shows model accuracy splits along hyperbolic versus elliptic PDE type in 2D heterogeneous reservoir simulations.","key_machinery":"Head-to-head comparison of U-Net, V-Net, TCN, FNO and U-FNO trained to map heterogeneous permeability-porosity fields and injection parameters to transient saturation and pressure fields.","core_discovery":"In a controlled benchmark of U-Net, V-Net, temporal convolutional networks, Fourier neural operators, and U-FNO on the 2D CO2 injection problem, surrogate performance is strongly dependent on the underlying PDE type, with U-FNO achieving the highest accuracy for predicting CO2 saturation fields and FNO providing the best performance for pressure build-up prediction.","pith_inferences":["If the 2D ranking holds in 3D, digital-twin workflows for carbon storage could route saturation and pressure queries to different specialized surrogates rather than a single model.","The observed PDE-type dependence suggests that future operator-learning work should test whether separate Fourier or convolutional branches for hyperbolic and elliptic components improve accuracy on fully coupled multi-phase problems.","History-matching and optimization loops that alternate between saturation and pressure updates could achieve lower overall error by switching architectures mid-loop."],"forward_implications":["For saturation prediction tasks the U-FNO architecture should be selected over the other tested models.","For pressure build-up prediction the plain FNO should be selected over the other tested models.","Architecture selection for coupled flow problems must account for the hyperbolic versus elliptic character of each field.","Memory and training-time differences among the models become decision criteria once accuracy rankings are known."],"fun_headline_variants":["U-FNO for saturation, FNO for pressure in CO2 surrogates","Performance varies by PDE type in DL surrogates for CO2 flow","U-FNO saturation accuracy and FNO for pressure in 2D reservoir benchmark","Hyperbolic vs elliptic PDEs affect DL model results in carbon sequestration"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The 2D single-wellbore injection problem with anisotropic heterogeneous permeability and porosity is representative of the high-dimensional transient subsurface flow problems encountered in real geological carbon sequestration.","fun_headline_variants_meta":{"raw":{"variants":["U-FNO for saturation, FNO for pressure in CO2 surrogates","Performance varies by PDE type in DL surrogates for CO2 flow","U-FNO saturation accuracy and FNO for pressure in 2D reservoir benchmark","Hyperbolic vs elliptic PDEs affect DL model results in carbon sequestration"]},"model":"grok-4.3","cost_usd":0.007219,"raw_usage":{"total_tokens":3355,"prompt_tokens":720,"num_sources_used":0,"completion_tokens":78,"cost_in_usd_ticks":72187000,"prompt_tokens_details":{"text_tokens":720,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2557,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":720,"tokens_out":78,"duration_ms":17195,"temperature":1.0,"reasoning_tokens":2557,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T20:29:28.458309+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Repeating the identical architecture comparison on a 3D heterogeneous reservoir model or on a multi-well injection scenario and checking whether U-FNO and FNO retain their respective top rankings for saturation and pressure.","supporting_citations":[],"review_version":1}