{"id":"86fb54a4-6a7d-48fc-ae12-e9c3074f0219","arxiv_id":"1908.05582","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A nonlinear space-time upscaling framework, using constraint-based test functions and machine-learned local maps, achieved 0.3-3.5% coarse-grid errors on two porous-media flow tests.","lead":"This paper builds a faster coarse-grained model for fluid flow in complex, fractured porous media by summarizing fine-scale details with a handful of macroscopic variables. It uses machine learning to learn the local relationships among those variables, and reports smaller errors than standard upscaling on two test problems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 0.3–3.5% accuracy figures are in-sample: the CNN is trained and evaluated on local patches from the same global fine-grid solution used as the reference, and Theorem 1 does not cover the learned transmissibility scheme.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing issue: the numerical results in Section 4 are in-sample evaluations with no unseen-media test and no error bound linking the learned transmissibilities to the final coarse-grid error. This is the correct focus because the headline contribution is the ML-accelerated nonlinear NLMC framework, and the 0.3–3.5% relative L2 errors are the primary quantitative evidence that the simplified ML coarse model actually works. The theoretical Theorem 1 is for a monotone elliptic problem and explicitly does not cover the time-dependent two-phase flow application, so it cannot compensate for the missing out-of-sample validation. This is a significant evidence gap, but it is not evidence of internal inconsistency or a demonstrated failure of the method; a held-out-realization experiment or a transfer argument would settle it. I therefore agree with the reader's CONDITIONAL verdict and recommend no change.","tokens_in":17547,"tokens_out":6812,"duration_ms":68984,"concrete_test":"Hold out one or more entire permeability-and-fracture realizations sampled from the same statistical model (not 20% of local patches from the same realization), retrain the CNNs on the remaining realizations, and recompute e(uNL), e(pNL), and e(sNL) on the held-out realization(s). Also report the per-realization spread over at least 10 training seeds. If the held-out errors are comparable to the Section 4 values, the generalization concern is settled; if they degrade substantially (e.g., to the classical upscaling level), the reported accuracy is an in-sample artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest numerical claims (e(uNL)=2.155% for Test 1; e(pNL)=0.281%, e(sNL)=3.512% for Test 2, Section 4) depend on a condition that is not established: that transmissibilities trained on local patches extracted from fine-grid calculations on one global domain Ω remain accurate on the coarse-grid states encountered in the global solve, and on other realizations. Section 4 uses an 80:20 train/validation split of local samples drawn from the same Ω, and the reported errors are computed against the same reference fine-grid solution. A held-out 20% of patches is not a held-out medium; the split does not test generalization to unseen permeability fields, which is the intended use of an upscaling method. In addition, the theoretical analysis in §3.4 bounds the full NLMC downscaling map Fms for a monotone elliptic problem; it does not apply to the simplified ML transmissibility model actually used in Section 4, and no estimate connects the Table 1 RMSE/MAE values to the final relative L2 errors. Without such a link or an out-of-sample test, the 0.3–3.5% accuracy cannot be separated from the training data used to produce it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a space-time nonlinear nonlocal multi-continuum upscaling framework for problems with multiscale coefficients. In the framework, test functions on each coarse space-time element define continua (macroscopic variables); a downscaling map is built by solving nonlinear local problems on oversampled regions subject to constraints set by the test functions; and coarse macroscopic variables are found from a variational formulation using the same test functions. Section 3.4 develops a convergence analysis for a monotone elliptic model (Eq. (11)) and states Theorem 1, an error bound of the form ||Fms_1(Ums)-u||_a <= CH + C1(M) + C2(M), with O(H) accuracy when M ~ O(log(H^-1) + log(C_kappa)). Section 4 replaces the expensive local solves with CNN-learned upscaled transmissibilities and reports relative L2 errors for an unsaturated flow problem (Test 1: e(uNL)=2.155% versus e(uUP)=11.773%) and a two-phase flow-transport problem (Test 2: e(pNL)=0.281% and e(sNL)=3.512% versus e(pUP)=14.063% and e(sUP)=13.354%).","tokens_in":17797,"tokens_out":9107,"duration_ms":84555,"significance":"If demonstrated as stated, the framework would be a useful extension of nonlocal multi-continuum methods to nonlinear and space-time heterogeneous problems. The paper's strengths include a conceptually clean construction based on test-function-defined continua, a self-contained error analysis for the model monotone elliptic problem, and a favorable numerical comparison against classical transmissibility upscaling on the two test problems. The numerical study is presented against a fine-grid reference, which is an appropriate benchmark. However, the analysis covers only a stationary monotone elliptic model, while the numerical claims are for time-dependent models solved with a machine-learned surrogate; the numerical protocol is in-sample, so the reported accuracy figures are not yet evidence of generalization to unseen media.","major_comments":[{"comment":"Theorem 1 analyzes only the monotone elliptic equation (11), but Section 4 tests the method on a time-dependent unsaturated flow problem and a two-phase flow problem using simplified local problems and CNN-fitted transmissibilities. No estimate in Section 3.4 connects the learned transmissibilities (Table 1) to the downscaling map Fms used in Theorem 1, so the reported relative L2 errors (e(uNL)=2.155% in Test 1; e(pNL)=0.281% and e(sNL)=3.512% in Test 2) are not consequences of the convergence analysis. Please add an error-propagation statement for the ML surrogate, or explicitly present the numerical results as heuristic evidence.","section":"Section 3.4 vs Section 4"},{"comment":"The neural networks are trained on local patches extracted from fine-grid calculations on the same global domain Omega used to compute the reference solution, with an 80:20 train/validation split of those patches. Because the validation set is a random subset of patches from the same medium rather than an unseen medium, the claimed accuracy is in-sample: it does not test whether the learned transmissibilities generalize to new permeability or fracture configurations, which is the intended use of an upscaling method. Please provide an out-of-sample test (e.g., train on one realization or region and test on another) or provide a theoretical link between the Table 1 errors and the final coarse-grid errors.","section":"Section 4, training and validation protocol"},{"comment":"Theorem 1 and Lemmas 3-5 are conditional on three assumptions that are neither proved nor referenced for the continua used in this paper. Assumption 1, in particular, is a nontrivial simultaneous approximation bound comparing the s-norm and a-norm after projection, and it directly produces the O(H) term in Lemma 4. The paper should prove these assumptions for the chosen test functions or provide precise references before Theorem 1 can be considered established.","section":"Section 3.4, Assumptions 1-3"},{"comment":"The decay bound in Lemma 5 uses the factor (1 - C^(-1) C_1^(-1) C_2), but the proof does not state that this factor lies in (0,1), which is needed for the powers (1 - ...)^M to be meaningful and decaying. From (12)-(13) one can expect C_2 <= C_1, so the point is probably fixable, but it should be written explicitly and the range of the constants should be verified.","section":"Section 3.4, Lemma 5"}],"minor_comments":[{"comment":"The expression 'C2||u-F(Pi u)||^2_a' in the proof of Lemma 4 is inconsistent with the statement; it should refer to F1(Uglo).","section":"Section 3.4, Lemma 4 proof"},{"comment":"The proof places the term (1/2)||Fms_1(Ums)-u||_a on the right-hand side and stops; the standard absorption argument should be stated so the final O(H) conclusion is explicit.","section":"Section 3.4, Theorem 1 proof"},{"comment":"The quantity called MSE is defined as a sum of squared errors, not a mean squared error; either rename it or change the definition.","section":"Section 4, error definitions"},{"comment":"Notation in Section 3 is inconsistent: the downscaling map appears as Fms, Fms_1, Floc, Floc,K, and F1 across Section 3.4; standardize the notation.","section":"Section 3, notation"},{"comment":"There are several typos, including 'tis lemma' in the proof of Lemma 5 and 'RELU' for ReLU in Section 4; the error definition for u=(p,s) should also be written more carefully.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the main risk is the in-sample numerical validation. If the authors can supply a genuine out-of-sample test or an error-propagation bound, the paper would be substantially strengthened. The paper is within scope and the analysis is coherent, but the current presentation overclaims what the numerical results establish."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a serious look. It extends the authors' NLMC framework to space-time nonlinear problems, gives a convergence analysis for a monotone elliptic model, and adds a CNN-based surrogate for the local maps. That combination is genuinely new relative to their own prior work. The numerical comparison is also striking: on two test problems, the ML-based coarse model lands at 0.3–3.5% relative L2 error while simple upscaling sits around 11–14%. Those gains are the reason to read the paper.\n\nWhat it does well: the framework itself is coherent. Defining continua via test functions, building local downscaling maps in oversampled regions, and then imposing the coarse variational form is a clean way to generalize both CEM-GMsFEM and NLMC to nonlinear problems. The proof in Section 3.4 gives a rigorous O(H) error bound for a monotone elliptic problem once the oversampling grows like log(H^{-1}), assuming the three technical Assumptions hold. That is a real result, even if limited.\n\nThe soft spots are real, but not all equally soft. First, the analysis rests on Assumptions 1–3 that are stated but not verified or discussed for the examples. Lemma 5's decay factor requires the product C^{-1}C1^{-1}C2 to lie strictly below 1; the paper never checks or even names that condition. For linear problems it holds, but for the nonlinear cases here it is not obvious. This is a gap a referee should push on, not a fatal flaw.\n\nSecond, the theory and the numerics are decoupled. The theorem covers the monotone elliptic model; the tests are unsaturated and two-phase flow with simplified local solves and learned transmissibilities. The ML surrogate is not the downscaling map Fms analyzed in Section 3.4, so Theorem 1 does not justify the reported accuracies. That disconnect is the main structural weakness.\n\nThird, and most important for the practical claim: the 0.3–3.5% numbers come from training and testing on the same global fine-grid solution. The 80/20 split is over local patches, not over media realizations. An upscaling method has to work on unseen permeability fields; in-sample accuracy does not show that. This is a fixable issue but it changes the strength of the conclusion.\n\nWho is this for? Researchers in multiscale modeling and computational geoscience who care about nonlinear upscaling. It is a solid methodology paper with an honest (if partial) analysis. It deserves a serious referee, but the referee should demand an out-of-sample test and a bridge between the theory and the learned surrogate. I would recommend conditional acceptance after major revision, not desk rejection.","headline":"A real NLMC extension with a partial convergence proof, but the headline accuracy numbers are in-sample and the analysis does not cover the ML-based numerical scheme.","tokens_in":18327,"tokens_out":2830,"would_cite":true,"duration_ms":29037,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M60","65N30","76S05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes a space-time nonlinear upscaling framework in which test functions define macroscopic continua and oversampled local solves build a downscaling map, with $O(H)$ convergence proved for a monotone elliptic model.","keywords":["nonlinear upscaling","nonlocal multi-continuum","space-time heterogeneity","oversampling","machine learning","porous media","two-phase flow","high contrast"],"falsifier":"Take the same two test problems but evaluate the trained networks on a permeability/fracture realization not present in the training data, ideally one with a different channel or fracture geometry; if the relative $L^2$ errors of the ML-NLMC solution rise above the classical upscaling errors or fail to shrink as the coarse mesh is refined, the central claim that the learned local maps support accurate coarse models is falsified.","tokens_in":17339,"feed_emoji":"💧","tokens_out":6503,"duration_ms":57604,"temperature":0.7,"pith_summary":"This paper tries to establish that nonlinear multiscale problems in porous media can be upscaled in both space and time without assuming scale separation, by replacing each coarse block with several macroscopic 'continua' defined through test functions. Local nonlinear solves in oversampled regions, constrained by averages of these test functions, define a downscaling map from macroscopic variables to fine-grid fields; the coarse model then enforces the original equation weakly through the same test functions. For a monotone elliptic model, the paper proves that with oversampling of $M$ coarse layers the error is bounded by $CH + C_1(M) + C_2(M)$, and when $M\\sim O(\\log(H^{-1})+\\log(C_\\kappa))$ the error is $O(H)$. It also reports that replacing expensive local solves with neural-network-learned transmissibilities yields relative $L^2$ errors of $2.155\\%$ for an unsaturated flow problem and $0.281\\%$ (pressure) / $3.512\\%$ (saturation) for two-phase flow, against $11.8$–$14.1\\%$ for classical upscaling. A sympathetic reader would care because the framework offers a route to coarse models for highly heterogeneous, nonlinear, time-dependent problems where standard upscaling is known to be process-dependent.","feed_headline":"Nonlinear upscaling cuts coarse-grid error to 2.2%","feed_subtitle":"Multi-continua test functions plus neural-network transmissibilities beat classical upscaling on two-phase flow.","key_machinery":"The carrying mechanism is the nonlinear nonlocal multi-continuum (NLMC) construction with three ingredients: test functions $\\{\\psi^{(j)}_i\\}$ selecting continua, local downscaling maps obtained by solving constrained nonlinear problems on oversampled regions $K^+_i$ (with a Lagrange multiplier enforcing the continuum constraints), and a coarse-scale variational formulation using the same test functions. The convergence argument uses the monotone operator $A_\\omega(u,w)=\\int_\\omega \\kappa(x,\\nabla u)\\cdot \\nabla w$, three assumptions (approximation by the auxiliary projection, stability of the auxiliary space, and a norm equivalence), and a geometric decay lemma for the localization error $\\|F_i(U)-F^{loc}_i(U)\\|$ in oversampling layers. In the numerical part, the downscaling map is replaced by convolutional neural networks that learn upscaled nonlinear transmissibilities as functions of local permeability, fracture geometry, and coarse solution averages.","core_discovery":"The central claim is that a nonlinear version of nonlocal multi-continuum upscaling works for space-time heterogeneous problems: choose test functions that define continua (macroscopic variables) on each coarse block, solve local nonlinear problems in oversampled space-time regions with constraints locking the continua values, and use the resulting downscaling map in a global weak formulation. Under monotonicity, Lipschitz continuity, and a coercivity condition on the flux $\\kappa(x,\\nabla u)$, the paper proves Theorem 1: $\\|F^{ms}_1(U^{ms})-u\\|_a \\le CH + C_1(M)+C_2(M)$, with $C_1,C_2$ decaying geometrically in the number $M$ of oversampling layers, and $M\\sim O(\\log(H^{-1})+\\log(C_\\kappa))$ yields $\\|F^{ms}_1(U^{ms})-u\\|_a \\le CH$. The numerical section claims this framework, with convolutional neural networks predicting coarse-grid transmissibilities, outperforms classical upscaling in relative $L^2$ error on two porous-media test problems.","pith_inferences":["A natural test the paper does not run is cross-realization generalization: train the networks on local data from one permeability/fracture field and evaluate on a different field; if the reported accuracy degrades sharply, the method's practical value depends on training-data coverage rather than on the upscaling construction itself.","The analysis is for a monotone elliptic model with $p=2$; extending Theorem 1 to the degenerate parabolic two-phase flow case would require additional compactness or monotonicity structure, since the numerical tests do not carry an error bound.","One could combine the ML local solves with online corrections: when a coarse state falls outside the training distribution, fall back to a constrained local solve, which would make the approach more robust without abandoning the learned map."],"forward_implications":["For monotone elliptic problems, the coarse solution converges linearly in coarse-mesh size $H$ once oversampling grows logarithmically with $H^{-1}$ and the contrast $C_\\kappa$.","The local downscaling maps are stable in the energy norm, so the same framework can be reused as a component in larger coarse systems without blowing up.","The two numerical tests show relative $L^2$ errors of about $0.3$–$3.5\\%$ for the ML-based nonlinear NLMC solution, compared with roughly $12$–$14\\%$ for classical upscaling at final time.","Because the local problems mirror the original global problem, the framework applies to time-dependent nonlinear models such as two-phase flow and transport, not just elliptic problems.","The machine-learned transmissibilities can be trained once from local fine-grid data and then used in the coarse solve, avoiding repeated expensive local solves during simulation."],"supporting_citations":[{"why":"defines the nonlocal multi-continua upscaling concept and its three ingredients of continua, local basis, and global coupling.","marker":"[16]"},{"why":"is the preceding nonlinear NLMC framework that this paper extends to space-time problems.","marker":"[17]"},{"why":"supplies the energy-minimizing basis construction and the spatial-decay theory that motivates oversampled local problems.","marker":"[18]"},{"why":"gives the fast online variant of the energy-minimizing construction used for localization estimates.","marker":"[19]"},{"why":"provides the mixed-formulation version of the energy-minimizing multiscale method referenced for the decay of basis functions.","marker":"[14]"},{"why":"analyzes the NLMC upscaling method for high-contrast problems, supporting use of piecewise-constant continuum definitions.","marker":"[47]"},{"why":"documents the process dependence of pseudo-relative permeability upscaling that motivates a general nonlinear framework.","marker":"[6]"}],"fun_headline_variants":["Nonlinear upscaling beats classical on two-phase flow","Proven error bound for space-time nonlinear upscaling","ML-powered nonlinear upscaling improves coarse-grid accuracy","Nonlocal multi-continuum upscaling: thin oversampling layers suffice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned transmissibilities that produce the reported 0.3–3.5% errors are trained on local fine-grid data extracted from the same global domain used as the reference solution, with no test on unseen media and no error bound connecting learning error to the final $L^2$ error; if those learned functions fail on coarse states outside their training set, the numerical accuracy claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Nonlinear upscaling beats classical on two-phase flow","Proven error bound for space-time nonlinear upscaling","ML-powered nonlinear upscaling improves coarse-grid accuracy","Nonlocal multi-continuum upscaling: thin oversampling layers suffice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001421,"raw_usage":{"total_tokens":5789,"prompt_tokens":1054,"completion_tokens":4735,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":4665}},"tokens_in":670,"tokens_out":4735,"duration_ms":32146,"temperature":1.0,"reasoning_tokens":4665,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:08:49.881303+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same two test problems but evaluate the trained networks on a permeability/fracture realization not present in the training data, ideally one with a different channel or fracture geometry; if the relative $L^2$ errors of the ML-NLMC solution rise above the classical upscaling errors or fail to shrink as the coarse mesh is refined, the central claim that the learned local maps support accurate coarse models is falsified.","supporting_citations":[{"cited_title":"Non-local Multi-continua Upscaling for Flows in Heterogeneous Fractured Media","cited_arxiv_id":"1708.08379","evidence_quote":"defines the nonlocal multi-continua upscaling concept and its three ingredients of continua, local basis, and global coupling."},{"cited_title":"Nonlinear nonlocal multicon- tinua upscaling framework and its applications","cited_arxiv_id":null,"evidence_quote":"is the preceding nonlinear NLMC framework that this paper extends to space-time problems."},{"cited_title":"Constraint energy minimizing generalized multiscale finite element method","cited_arxiv_id":null,"evidence_quote":"supplies the energy-minimizing basis construction and the spatial-decay theory that motivates oversampled local problems."},{"cited_title":"Fast online generalized multiscale finite element method using constraint energy minimization","cited_arxiv_id":null,"evidence_quote":"gives the fast online variant of the energy-minimizing construction used for localization estimates."},{"cited_title":"Constraint energy minimizing generalized mul- tiscale finite element method in the mixed formulation","cited_arxiv_id":null,"evidence_quote":"provides the mixed-formulation version of the energy-minimizing multiscale method referenced for the decay of basis functions."},{"cited_title":"An analysis of the NLMC upscaling method for high contrast problems","cited_arxiv_id":"1904.11124","evidence_quote":"analyzes the NLMC upscaling method for high-contrast problems, supporting use of piecewise-constant continuum definitions."},{"cited_title":"Barker and S","cited_arxiv_id":null,"evidence_quote":"documents the process dependence of pseudo-relative permeability upscaling that motivates a general nonlinear framework."}],"review_version":1}