{"id":"8f1fa47a-e7fa-4b98-80b7-c46c044bb88d","arxiv_id":"2412.05576","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"STONet, a neural operator combining DeepONet with transformer-style attention, predicts contaminant concentration fields in micro-cracked reservoirs with reported relative errors below 1% against finite-element simulations and about 100x speedup.","lead":"STONet is a machine-learning model that predicts how dissolved contamination travels through rock riddled with tiny cracks, trained on 500 computer simulations. It reports errors below 1% and roughly 100x faster predictions than the simulations, which could speed up groundwater risk assessment and cleanup planning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4) as written reduces the fluid density contrast by a factor of rho0 ≈ 10^3; if this equation was used to generate the FEM labels, the dataset contains no true density-driven flow, so the central <1% accuracy claim is against a different, easier problem.","rationale":"The reader's weakest_assumption correctly identifies Eq. (4) and the density–concentration relation as the point where the validity of the FEM labels could fail. My analysis confirms that the equation as printed is dimensionally inconsistent and quantitatively suppresses the density contrast by a factor of roughly 10^3. This is not a stylistic or numerical-precision issue; it changes the physical regime of every generated sample. The central claim in the abstract explicitly invokes 'balance of pressure-driven to density-driven flow' as part of the training distribution, so the surrogate's reported sub-1% errors are only evidence of competence for the problem the authors intend if the labels actually contain density-driven effects. Since the manuscript provides no separate validation of the FEM solver (e.g., reproducing a classical density-driven benchmark such as Henry or Elder) and no code or data is currently available, the concern remains unresolved. The appropriate verdict is CONDITIONAL: the paper should be published only if the authors either demonstrate that Eq. (4) is a typesetting error with the correct formula used in the solver, or regenerate the dataset with the corrected density relation and retrain. All other weaknesses noted by the reader—missing baselines, single geometry, absent external validation—are real but secondary; they affect generalizability claims, not the internal correctness of the surrogate against its stated ground truth. My recommendation therefore preserves the reader's CONDITIONAL verdict while making the density-relation verification the explicit condition.","tokens_in":12699,"tokens_out":2576,"duration_ms":25895,"concrete_test":"Run one training realization through the FEM pipeline twice: once with the published Eq. (4) (ρ(c) = ρ0 + (ρs−ρ0)/ρ0·c) and once with the physically correct linear relation ρ(c) = ρ0 + (ρs−ρ0)·c, keeping all stochastic sampling and discretization identical. Compare the concentration fields and the concentration-rate fields at the last recorded time step (t = 36 h). If the two solutions differ substantially—e.g., maximum concentration, plume position, or arrival time change by more than a few percent—then the published equation is not a harmless typo and the dataset must be regenerated. In addition, upon release of the GitHub repository, grep the solver source for the exact density expression; if Eq. (4) appears verbatim, the labels are unphysical, and the 1%-accuracy result is invalid for the claimed problem.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's headline claim—'relative errors typically below 1% compared with FEM simulations'—is only meaningful if the FEM ground truth is physical. Equation (4) defines the density as ρ(c) = ρ0 + (ρs − ρ0)/ρ0 c. With ρ0 = 998.2 and ρs = 1002 kg/m³, this gives ρ(1) ≈ 998.2038 kg/m³, not ρs = 1002 kg/m³. The density contrast is reduced by roughly 1/ρ0 ≈ 10⁻³, making the buoyancy term in Eq. (3) negligible. Consequently, the stated sampling of 'pressure-driven to density-driven flow' in the abstract would not actually be realized: every simulation would be effectively a passive-scalar transport problem in a heterogeneous permeability field. STONet would still predict that scalar well, but the central claim of being a surrogate for density-driven solute transport in micro-cracked reservoirs would be unsupported. The paper provides no direct evidence—such as measured plume shapes, density contrasts, or reproduction of a published density-driven benchmark—that the FEM solver correctly captures variable-density effects. Moreover, the code and data are promised but not yet released, so an independent reader cannot determine whether Eq. (4) is a typographical error or the exact formula implemented in the dataset generation. If it is the latter, the training labels are non-physical, and any agreement with FEM is agreement with a defective reference solution. This concern is the single most load-bearing issue because it attacks the validity of the ground truth itself, not merely the breadth or architecture of the surrogate. Other limitations (single geometry, restricted boundary conditions, no baseline comparisons) limit generalizability but do not undermine the internal accuracy claim in the same way.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces STONet, an enriched DeepONet that incorporates transformer-style multi-head attention blocks with residual connections, for surrogate modeling of solute transport in micro-cracked porous media. The model takes the equivalent permeability tensor and pressure boundary conditions as inputs, predicts the concentration rate, and reconstructs the concentration field autoregressively via forward Euler time stepping. Training and test data come from 500 and 25 finite-element simulations, respectively, with random micro-fracture statistics and pressure conditions. The authors claim that STONet achieves relative errors typically below 1% compared with FEM simulations and reduces runtime by approximately two orders of magnitude, with code and data to be published at a promised GitHub repository.","tokens_in":13087,"tokens_out":4282,"duration_ms":39185,"significance":"If the reported accuracy and speed hold, STONet would be a practically useful surrogate for the studied parameter ranges, and the effort to encode the permeability tensor and to output the concentration rate are sensible design choices. The architecture itself is an incremental but reasonable modification of En-DeepONet, and the paper includes an honest limitations section. The central claims are currently weakened by the absence of quantitative error tables and repeated-seed statistics, and by an apparent normalization error in the density state equation that, if present in the data-generation pipeline, would eliminate density-driven flow from the training set. Because no code or data are yet available, independent verification of the implementation is not possible.","major_comments":[{"comment":"As printed, Eq. (4) gives rho(c) = rho0 + (rho_s - rho0)/rho0 * c, which with the Table 1 values yields rho(1) = 998.2 + 3.8/998.2 ≈ 998.2038 kg/m^3 rather than rho_s = 1002 kg/m^3. The density contrast is therefore reduced by three orders of magnitude, making the buoyancy term in Eq. (3) negligible and eliminating the 'balance of pressure-driven to density-driven flow' claimed in the Abstract. The authors must either correct the equation to rho(c) = rho0 + (rho_s - rho0)*c or demonstrate that the FEM label generation used a different, physically correct relation; if the printed form was actually implemented, the trained surrogate is a passive-scalar model and the central physical claim is unsupported.","section":"Section 2, Eq. (4), and Table 1"},{"comment":"The headline claim that relative errors are 'typically below 1%' is supported only by qualitative error distributions and mean-curve plots. The paper provides no numeric table of per-sample or aggregate relative errors, no maximum errors, no quantiles, and no repeated-seed statistics or confidence intervals. Without such quantitative reporting, the central accuracy claim is not verifiable as stated. Please add a table reporting mean, median, maximum, and quantiles of relative error on the 25 test samples, ideally across multiple training seeds.","section":"Section 4.4, Figures 10 and 11"},{"comment":"All test labels come from the same FEM code and equivalent-continuum formulation used to generate the training data, so the reported accuracy measures fit quality within one simulation family rather than physical fidelity of the density-driven transport model. To support the density-driven claim independently of the Eq. (4) concern, the authors should validate the FEM pipeline against a published variable-density benchmark (e.g., the Henry or Schincariol problem) and report the simulated density contrast or a dimensionless buoyancy measure for the training/test distributions.","section":"Section 4.4 and Discussion"}],"minor_comments":[{"comment":"The claim that STONet improves performance 'without incurring additional computational overhead' is based only on parameter counts in Figure 6; please report measured training and inference times or memory usage to support the computational-cost claim.","section":"Section 3.2, Figure 6"},{"comment":"The y-axis label 'average loss' should specify the exact loss function (e.g., mean squared error of the concentration rate) and state whether the reported values are on training or validation data.","section":"Section 4.3, Figures 6 and 7"},{"comment":"The symbol Sigma is used for both the REV volume in Eq. (5) and the Poisson-distributed fracture density in Algorithm 1; please use distinct notation to avoid ambiguity.","section":"Algorithm 1 and Eq. (5)"},{"comment":"The boundary-condition notation pright = U(4976, 4996) + 9792.34y should be described explicitly as a uniform perturbation added to a hydrostatic pressure gradient, with units stated for the coefficient 9792.34.","section":"Section 4.1"},{"comment":"For each displayed test sample, please list the corresponding input parameters (mean fracture orientation, Poisson rate, pressure perturbation) so that the reader can assess the diversity of the shown cases.","section":"Figures 8 and 9"},{"comment":"The Data availability statement promises a repository that is not yet provided; please make the code and data available with a working URL or DOI so that the implementation of Eq. (4) and the reported results can be independently checked.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The Eq. (4) issue is potentially decisive: if the printed normalization was used in the FEM label generation, the physical problem being solved is not density-driven flow, and the paper's central claim collapses. Because the manuscript is otherwise a reasonable surrogate-modeling study, I recommend major revision rather than rejection, contingent on the authors correcting the equation, releasing the code/data, and adding the missing quantitative error reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: STONet is a reasonable engineering contribution—a DeepONet variant with attention-style fusion and residual connections, trained on FEM simulations of transport in random micro-fracture fields. The authors show it beats their own En-DeepONet at equal parameter counts and report plausible sub-1% error on 25 held-out cases with a two-order speedup. The limitations are stated honestly: single 2D geometry, fixed injection, no varying boundary conditions. If the only question is 'can a neural operator learn this FEM family,' the answer is yes.\n\nThe soft spots are real, and one is potentially fatal. Eq. (4) is dimensionally wrong as printed: it divides the density contrast by ρ0, so ρ(1) ≈ 998.2 instead of 1002. If that formula was used in the FEM code, every simulation is effectively a passive-scalar problem with negligible buoyancy. The abstract's claim that the data covers 'pressure-driven to density-driven flow' would then be false, and the <1% accuracy would be against a surrogate of the wrong physics. The paper provides no independent check—no comparison with a published density-driven benchmark, no measured plumes—so a reader cannot tell whether this is a typo or the implemented equation. This is the first thing I'd ask for in revision: correct the equation, confirm the implementation, and show one validation case where density contrast actually matters.\n\nOther issues are less severe but worth fixing: no numeric error table (we get only distributions), no repeated-seed statistics, and no comparison with other attention-based operator models. The comparison is only against En-DeepONet, which makes the 'novelty' claim weaker than it could be. The test set is from the same FEM family, so accuracy measures fit quality, not physical correctness—normal for surrogates, but worth stating.\n\nOverall: the paper is a competent empirical study, and I'd engage with it seriously. But the density equation issue must be resolved before the headline claim is credible. As it stands, I'd recommend conditional acceptance: insist on corrected equations, artifact release, and at least one external validation. If the authors can confirm the implementation is correct, this becomes a useful contribution for the subsurface-flow community.","headline":"Useful surrogate paper with a load-bearing typo in the density equation; the physics claim needs verification before the 1% accuracy claim can be trusted.","tokens_in":13642,"tokens_out":4715,"would_cite":false,"duration_ms":37466,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","76S05","65M60"],"pacs":[],"model":"deepseek-v4-flash","headline":"STONet, a neural operator for solute transport in micro-cracked reservoirs, predicts concentration fields from fracture statistics and pressure conditions, with relative errors typically below 1% against FEM simulations and roughly two…","keywords":["neural operator","solute transport","fractured porous media","DeepONet","transformer attention","density-driven flow","surrogate modeling","finite element method"],"falsifier":"Run STONet on a test case with a solute density contrast similar to real seawater intrusion (e.g., $\\rho_s/\\rho_0 \\approx 1.025$, giving a much larger buoyancy term) and check whether pointwise relative errors stay below 1%; if accuracy degrades sharply outside the near-uniform-density regime sampled in training, the 1%-error claim holds only for weak density contrasts.","tokens_in":12525,"feed_emoji":"💧","tokens_out":4425,"duration_ms":38039,"temperature":0.7,"pith_summary":"The paper introduces STONet, a neural operator designed to model density-driven solute transport in micro-cracked porous media. Its central claim is that, once trained on 500 finite-element simulations spanning randomized micro-fracture densities, orientations, apertures, and pressure boundary conditions, STONet predicts the spatio-temporal concentration field on unseen fracture networks with relative errors typically below 1% while cutting runtime by approximately two orders of magnitude. This matters because repeated forward simulations for parameter identification, risk assessment, and remediation design are computationally expensive, and a fast, accurate surrogate would make such tasks tractable. The architecture predicts the time rate of change of concentration and integrates it auto-regressively, yielding a nearly constant relative error distribution over time.","feed_headline":"STONet predicts contaminant flow in fractured rock at under 1% error","feed_subtitle":"Once trained on 500 finite-element runs, the neural operator answers new fracture and pressure scenarios about 100x faster.","key_machinery":"The central object is the STONet architecture itself: an enriched DeepONet in which the branch and trunk encodings are combined through elementwise multiplication, addition, and subtraction, then passed through L attention blocks with residual connections, and finally decoded by a root network. The attention block with residual connections is the mechanism that improves accuracy over the base En-DeepONet at the same computational cost. The output is trained on the concentration rate, so predictions are rolled out in time with a forward-Euler update $c_{t+\\Delta t} = c_t + \\dot{c}\\,\\Delta t$, which allows flexible time stepping while keeping the relative error distribution stable over the forecast horizon.","core_discovery":"STONet revises the Enriched DeepONet architecture by replacing the single fusion step with a stack of attention-like blocks with residual connections that operate on the elementwise product, sum, and difference of branch and trunk encodings. The network takes as input the equivalent permeability tensor (which encodes fracture statistics and intrinsic permeability) and pressure boundary conditions, and outputs the concentration rate. Trained on 500 FEM simulations covering a wide range of fracture densities (Poisson λ ~ 30–70), orientations (normal distribution with uniform mean), lengths, apertures, and right-hand pressure values, the model achieves typical pointwise relative errors below 1% on 25 unseen test cases and runs about 100 times faster than FEM inference, with the same architecture outperforming the base En-DeepONet at equal parameter count.","pith_inferences":["Because Eq. (4) gives $\\rho(1) \\approx \\rho_0$, the actual density contrast in the training data is tiny; whether STONet handles strongly density-driven flows (e.g., seawater intrusion with ~2.5% density contrast) is untested and is the natural next experiment.","The reported 1% accuracy is measured against the same FEM solver that generated the labels, so the surrogate is calibrated to that solver's physical and numerical assumptions; if the equivalent-continuum upscaling is inaccurate, the surrogate inherits the error.","The attention fusion over branch and trunk features is architecture-agnostic and could be transferred to other DeepONet-family operators for problems with heterogeneous coefficient fields, not just fracture permeability.","Extending to 3D domains or multiple, varying contaminant sources would require substantially more training samples, as the authors note, and would be a direct test of the method's scalability."],"forward_implications":["STONet can replace repeated FEM solves in parameter identification and fracture-network characterization, reducing the simulation budget of inverse problems by about two orders of magnitude.","The architecture supports building digital twins for rapid contamination-risk assessment and optimization of remediation strategies in micro-cracked reservoirs.","Auto-regressive prediction of the concentration rate gives a nearly constant relative error over time, which is desirable for long-horizon forecasts.","The attention-with-residual design yields better training loss than En-DeepONet for the same parameter count, suggesting it scales to other operator-learning tasks.","The model generalizes to unseen fracture networks within the sampled range of fracture statistics and pressure conditions."],"supporting_citations":[{"why":"Defines the En-DeepONet architecture that STONet modifies and compares against.","marker":"[16]"},{"why":"Supplies the FEM model and equivalent continuum formulation used to generate the training dataset.","marker":"[26]"},{"why":"Provides the homogeneous benchmark problem that the heterogeneous setup extends.","marker":"[45]"},{"why":"Introduces DeepONet, the operator-learning framework underlying the branch/trunk design.","marker":"[33]"},{"why":"Origin of the multi-head attention mechanism that STONet's attention block resembles.","marker":"[47]"},{"why":"Introduces residual connections used in STONet's attention blocks for stable training.","marker":"[18]"}],"fun_headline_variants":["STONet: 1% error, 100x faster contaminant transport modeling","Neural operator simulates micro-crack flow 100x faster","STONet predicts solute movement in cracked rock with 1% error","AI models contaminant transport in micro-cracked reservoirs","STONet: transformer-powered neural operator for subsurface flow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The trained model inherits the physics of the finite-element solver that produced its labels; if the equivalent-continuum density relation (Eq. 4) in the FEM pipeline misrepresents the driving density contrast, the surrogate will reproduce that error rather than true physics.","fun_headline_variants_meta":{"raw":{"variants":["STONet: 1% error, 100x faster contaminant transport modeling","Neural operator simulates micro-crack flow 100x faster","STONet predicts solute movement in cracked rock with 1% error","AI models contaminant transport in micro-cracked reservoirs","STONet: transformer-powered neural operator for subsurface flow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000155,"raw_usage":{"total_tokens":1210,"prompt_tokens":936,"completion_tokens":274,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":182}},"tokens_in":552,"tokens_out":274,"duration_ms":2700,"temperature":1.0,"reasoning_tokens":182,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:34:27.271315+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run STONet on a test case with a solute density contrast similar to real seawater intrusion (e.g., $\\rho_s/\\rho_0 \\approx 1.025$, giving a much larger buoyancy term) and check whether pointwise relative errors stay below 1%; if accuracy degrades sharply outside the near-uniform-density regime sampled in training, the 1%-error claim holds only for weak density contrasts.","supporting_citations":[{"cited_title":"Fuks and H","cited_arxiv_id":null,"evidence_quote":"Defines the En-DeepONet architecture that STONet modifies and compares against."},{"cited_title":"Jiménez-Martínez, A","cited_arxiv_id":null,"evidence_quote":"Supplies the FEM model and equivalent continuum formulation used to generate the training dataset."},{"cited_title":"Raissi, P","cited_arxiv_id":null,"evidence_quote":"Provides the homogeneous benchmark problem that the heterogeneous setup extends."},{"cited_title":"Saeibehrouzi, R","cited_arxiv_id":null,"evidence_quote":"Origin of the multi-head attention mechanism that STONet's attention block resembles."},{"cited_title":"Goswami, A","cited_arxiv_id":null,"evidence_quote":"Introduces residual connections used in STONet's attention blocks for stable training."}],"review_version":1}