{"id":"b57f8768-c0f0-40b8-ad3c-18544220758e","arxiv_id":"2507.05218","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 3D Volume of Fluid scheme is proposed in which a neural network directly predicts the numerical flux from cell volume fractions, removing explicit interface reconstruction and showing faster empirical convergence than upwind and limited downwind schemes on advection tests.","lead":"This paper trains a neural network to replace the interface-reconstruction step in 3D Volume of Fluid simulations, predicting fluxes directly from local volume fractions. If the reported convergence rates hold beyond the tested shapes, it could simplify multiphase flow codes and handle complex interfaces more easily.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing issue is distribution shift: the network is validated only on interfaces made from the same half-space/ellipsoid family used for training, so out-of-family interfaces could invalidate the claimed convergence advantage; an explicit out-of-distribution flux and advection test would…","rationale":"The reader's weakest assumption is exactly the load-bearing point: the paper validates on interfaces drawn from the same geometric family used to generate training data, so out-of-distribution failure would invalidate the claimed generality. I examined other possible objections. The projection bound in Eq. (3.7)/(3.8) is internally inconsistent because the admissibility interval is written in terms of alpha^x_{i,j,k}, the quantity being computed, so as written it cannot be applied before the update; this is a real defect in the stability argument, but it is secondary to the central numerical claim. The synthetic data generation is sound, the symmetry averaging and projection are coherent, and the provided tests do support convergence on those problems. The absence of released code and weights is a reproducibility problem, not a mathematical one. The decisive issue is that the empirical demonstration never leaves the training family, while the paper's central claim is about the method's general efficiency and accuracy in 3D. A concrete out-of-distribution flux and advection test would settle whether the network approximates the true flux operator or only the synthetic family. Therefore I keep the verdict CONDITIONAL: accept the mechanism and the reported numerical evidence, but require the OOD test before accepting the broader claim of a general 3D VOF-ML advantage.","tokens_in":22248,"tokens_out":5657,"duration_ms":75839,"concrete_test":"Build a fresh test set of 10^4 local 3x3x3 stencils whose exact interface is drawn from out-of-family geometries: thin ellipsoidal shells, two intersecting ellipsoids, oblique cylinders, and wavy interfaces with wavelength comparable to h. For each, compute the exact x-flux by the same convex-hull procedure and compare with the published network (or a retrained copy) at Courant numbers in [0,0.6]. If the flux MAE stays close to the in-distribution value 4.66e-3, the generalization premise holds; if it degrades by severalfold, rerun Test 1 with a thin-shell or cylinder initial condition and check whether the reported convergence rate of about 1.01 survives. This single experiment separates 'trained on the test distribution' from 'approximates the VOF flux operator'.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that a trained network maps the 3x3x3 volume-fraction stencil plus a scalar Courant number to the exact directional flux accurately enough to give 3D VOF-ML a convergence rate of about 1.01 versus 0.58 for limited downwind. For this to hold beyond the displayed test problems, the flux must be effectively representable as a single-valued function of these inputs for arbitrary interfaces. The training set contains only intersections of a cube with one to three half-spaces and ellipsoid interiors, with the central cell forced to contain both fluids, and Table 2 measures error on a held-out split of the same generator. The numerical tests are assembled from the same primitives: Test 1 is a sphere minus a rectangular prism, Test 2 adds three rectangular bars to a sphere, and Test 3 is a sphere. Locally these are close to the training family, but nothing tests genuinely new local topology such as two disconnected interface fragments in one stencil, a thin shell with thickness below the mesh scale, or a high-curvature interface crossing the stencil multiple times. Because no theorem shows that exact flux is uniquely determined by stencil averages, the network may be memorizing the parametric training family rather than approximating the universal VOF flux operator. The convergence advantage over LD therefore rests on an unverified generalization assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a three-dimensional Volume-of-Fluid (VOF) scheme in which the numerical interface flux is computed directly by a feed-forward neural network from a 3x3x3 stencil of volume fractions and the local Courant number, without any explicit interface reconstruction. The network is trained offline on a synthetic dataset generated from intersections of half-spaces and ellipsoids, with no retraining during the simulations. The authors also propose symmetry-averaging and flux-projection post-processings to enforce physical constraints. They test the scheme on three 3D advection problems (constant, directionally conservative, and general divergence-free velocity fields) and report empirical convergence rates of about 1.01 for VOF-ML, compared with 0.58 for the limited downwind scheme and 0.16 for upwind.","tokens_in":22602,"tokens_out":10760,"duration_ms":104888,"significance":"The conceptual simplification of bypassing interface reconstruction is attractive and could genuinely simplify three-dimensional VOF implementations. The symmetry-averaging construction (Section 3.1.1) and the flux-bounding strategy (Section 3.2) are cleanly motivated and, apart from the index issue discussed below, the algebraic steps are correct. The synthetic dataset generation is reproducible, and the network is used in simulations without any fitting to the advection results, so the scheme is not circularly validated. If the convergence advantage is robust on broader test families, this would be a valuable contribution to the VOF literature. The main caveat is that the numerical evidence so far is confined to interfaces drawn from the same geometric family as the training data, which limits the generality of the central claim.","major_comments":[{"comment":"The upper bounds in (3.7) are written as α_{i+1/2,j,k} ≤ (1/β^x) α^x_{i,j,k} and 1 − α_{i+1/2,j,k} ≤ (1/β^x)(1 − α^x_{i,j,k}), with α^x_{i,j,k} denoting the intermediate x-sweep value. However, the proof of Lemma 3.3 and the compatibility check in Remark 3.4 treat the bounds as involving the pre-update value α_{i,j,k}: the proof requires α_{i,j,k} − β^x α_{i+1/2,j,k} ≥ 0, and Remark 3.4 claims that the choice α_{i+1/2,j,k} = α_{i,j,k} satisfies the inequalities, which is not true for the α^x version in general. Since Definition 3.5 uses these bounds to project the network output, the maximum-principle argument as written is invalid. The authors should correct the index (likely replacing α^x_{i,j,k} with α_{i,j,k} in (3.7)) or revise the projection to be based on an explicitly computable quantity.","section":"Section 3.2, Eq. (3.7) and Remark 3.4"},{"comment":"The initial conditions are all composed of spheres, boxes, and bars, which are the same geometric primitives used to generate the synthetic training data (half-spaces and ellipsoids). No test exercises a genuinely different local interface topology, such as two disconnected interface fragments in one stencil, a thin shell of sub-cell thickness, or an interface crossing a stencil multiple times. The claimed convergence advantage (rates 1.01 vs. 0.58) is therefore demonstrated only within the training distribution. To support the stronger claim that the network approximates the universal stencil-to-flux operator, the authors should include at least one out-of-distribution advection test, e.g., an initial interface generated with a different random process (Fourier modes, random radial functions) or a known challenging benchmark, and report the resulting L1 errors.","section":"Section 4, Tests 1-3"},{"comment":"Each error curve in these figures is a single deterministic simulation per mesh, with no error bars or repeated trials. Since the network weights are initialized randomly, the reported rates (especially 1.01 for Test 1) could be sensitive to the specific training run. The authors should either fix and clearly state the random seed, or better, repeat the training several times (or evaluate with multiple network checkpoints) and show the spread of the L1 errors and of the fitted convergence rates.","section":"Figures 8, 10, and 12"},{"comment":"The mapping from stencil volume fractions to the exact flux is not single-valued: many different sub-cell interfaces can produce the same 3x3x3 vector of volume fractions but different fluxes. The synthetic dataset contains only one interface realization per input vector, so the network is trained to mimic a particular drawing from this conditional distribution rather than the true conditional expectation or worst-case flux. This could explain why accuracy degrades on interface topologies not represented in the training family. The authors should explicitly discuss this non-uniqueness and, ideally, train with multiple interface samples per stencil input (or add a variance penalty) to make the learned flux more robust.","section":"Section 2.2 and Remark 2.1"}],"minor_comments":[{"comment":"The equation is written as ∂tα + ∇·(αu) = ∂tα + u·∇αu = 0; the notation ∇αu is nonstandard and appears to be a typo for u·∇α.","section":"Abstract and Section 1, Eq. (1.1)"},{"comment":"'octohedral' should be 'octahedral'.","section":"Section 3.1"},{"comment":"There is a typo: 'a anti-clockwise rotations' should be 'an anti-clockwise rotation'.","section":"Section 4.1"},{"comment":"The ratio R_{mix}^{T/0} is reported without explaining how a 'mixed cell' is defined; consider adding a sentence defining the threshold used (ε_mark is defined in Appendix B, but a brief reminder in Section 4 would improve readability).","section":"Tables 3-5"},{"comment":"The claim that the normalization procedure 'preserves the mass conservation' is not immediately obvious for a general directional splitting, since normalizing per cell could alter the total mass of each phase; the authors should clarify why the additivity condition (Lemma 3.2) makes the normalization neutral in this respect.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core idea is interesting. The main risk is that the numerical validation is too narrow to support the advertised convergence advantage, and the index inconsistency in the maximum-principle proof is a load-bearing technical error. I recommend the paper be reconsidered after the authors address the generalization issue and correct the flux-bound derivation. No concerns about citation practices or novelty disclosure were identified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid extension of the 2D VOF-ML idea to 3D, with clean symmetry handling and honest numerical evidence, but the generalization claim is only supported on the same geometric family used for training, and there's a small notational slip in the positivity lemma.\n\nThe genuinely new piece is the 3D flux network: input 3x3x3 volume-fraction stencil plus one Courant number, output the directional flux, no interface reconstruction. The octahedral-symmetry averaging (8 permutations that preserve the flux) and the dataset enrichment by the 24 flux-changing symmetries are well thought out and correctly formalized. The synthetic dataset generation (exact polytope volumes for half-space intersections, Fibonacci-discretized ellipsoids) is reproducible in principle, and the numerical tests on three advection problems show the VOF-ML scheme converging at roughly 1.0 while limited downwind sits around 0.5-0.6. That's a meaningful result for the VOF community.\n\nSoft spots, in order of importance. First, the out-of-distribution question: the tests are all built from spheres and planar cuts, the same primitives as the training data. Nothing tests a genuinely different local topology — two disjoint interface pieces in one stencil, a subgrid thin shell, or repeated crossings. The paper itself claims the training family is 'sufficient' (Remark 2.1) but gives no evidence beyond the held-out split of the same generator. The central advantage could shrink on less friendly interfaces. Second, the error reporting: single deterministic runs, no seeds, no error bars. For a method that starts from a random initialization, that's a real gap. Third, there is a notational inconsistency in Lemma 3.3 and Remark 3.4: the flux bounds are written with the updated value αx_{i,j,k} where the proof needs the old value α_{i,j,k}. It's a typo, not a conceptual flaw, but it should be fixed. Fourth, no code or trained weights are released; given that the whole method is a black-box network, this hampers independent verification.\n\nOverall, the paper is coherent and the math of the symmetry projection is correct. The reader's conditional verdict is about right. I'd engage with it, but I'd want the authors to either narrow the claims or add a genuine out-of-distribution test, and to release the network weights.","headline":"Solid 3D extension of the ML-VOF idea with honest numerics, but the accuracy claim is only demonstrated on training-family interfaces and there is a small indexing typo in the positivity lemma.","tokens_in":23086,"tokens_out":2355,"would_cite":false,"duration_ms":26849,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35Q35","68T07","76-10","76M12"],"pacs":[],"model":"deepseek-v4-flash","headline":"A trained neural network can replace explicit interface reconstruction in a 3D volume-of-fluid scheme and still converge faster than standard schemes.","keywords":["Volume of Fluid","Machine Learning","CFD","3D","interface reconstruction","neural network","finite volume scheme","advection equation"],"falsifier":"Advect an initial condition whose interface is outside the training family, such as a thin cylindrical shell, a sharp wedge, or a three-material junction, on meshes with $N_h = 27, 54, 105$ and compare the final L1 error with limited downwind; if the VOF-ML error plateaus or its convergence rate drops below that of limited downwind, the transferability claim fails. A cheaper check is to evaluate the network on synthetically generated stencils from such shapes and measure whether the flux error remains competitive with the test-set error.","tokens_in":22068,"feed_emoji":"🧠","tokens_out":7809,"duration_ms":87182,"temperature":0.7,"pith_summary":"This paper claims that the hardest step of 3D volume-of-fluid simulation, reconstructing the interface shape to compute fluxes, can be dropped: a fully connected neural network, trained once on synthetic interfaces, maps the local 3×3×3 stencil of volume fractions and a Courant number directly to the flux fraction across a face. The training data are generated purely geometrically from half-space intersections and ellipsoids, so no simulations or experiments feed the network. Embedded in a directional-splitting finite volume advection scheme, the network flux yields numerical convergence as the mesh is refined, with an empirical rate around 1.01 in a constant-velocity test versus 0.16 for upwind and 0.58 for limited downwind, and similar advantages in more general divergence-free velocity fields. This matters because 3D interface reconstruction, especially for irregular or multi-material interfaces, is the main implementation bottleneck of classical VOF schemes; the authors' scheme avoids it and needs no retraining when the equation, domain, or initial condition changes.","feed_headline":"Learned 3D VOF flux converges faster than two reference schemes","feed_subtitle":"One network trained on synthetic interfaces replaces the hardest step of 3D volume-of-fluid tracking.","key_machinery":"The object carrying the argument is the neural-network flux map $\\alpha_{NN}(x,\\beta)$: a fully connected feed-forward ReLU network with input dimension $m^3+1$ (here 28, from a 3×3×3 stencil plus the Courant number) and scalar output, trained by minimizing mean squared error between its output and exact geometrical fluxes. Exact fluxes are computed with a convex-hull volume algorithm on synthetic regions defined as cube intersections with one, two, or three half-spaces and with ellipsoids, sampled by Latin Hypercube sampling and enriched by non-flux-preserving permutations. Around this map the paper places three mechanisms: averaging over the eight flux-preserving octahedral symmetries, a complementary-fluid averaging so that $\\alpha_{NN}(x)+\\alpha_{NN}(Mx)=1$, and a projection onto an admissibility interval that keeps volume fractions between 0 and 1. These mechanisms are what let a purely geometric training set be used as a conservative flux in a finite volume scheme.","core_discovery":"The central claim is that the exact transported flux through a cell face can be treated as a regression problem: given the volume fractions in a local cubic stencil and one non-negative Courant number, the flux fraction is approximated by a ReLU network with 9151 trainable weights trained on 144000 synthetic input-output pairs. The paper shows, on advection problems in a periodic cube, that this network flux, when symmetrized over the eight flux-preserving cubic symmetries, made complementary under fluid exchange, and projected onto an admissible interval, produces a scheme whose L1 error decreases toward zero with mesh refinement at a higher empirical rate than the upwind and limited-downwind reference schemes. The same trained network is used across all numerical tests without retraining, and the scheme never reconstructs an interface during inference.","pith_inferences":["A testable extension the authors did not run is an initial condition whose interface lies outside the training family, such as a thin cylindrical shell or a sharp wedge; if the L1 error plateaus or the convergence rate drops below limited downwind, the claimed transferability would be falsified.","Because the synthetic dataset is parameterized and cheap to generate, an adaptive strategy could resample stencils that are poorly predicted and retrain, giving the scheme an accuracy guarantee concentrated on the interfaces that actually appear in a run.","The symmetry-averaging trick could plausibly be extended from the eight flux-preserving symmetries to the full 48-element octahedral group, potentially reducing the number of training samples needed and improving accuracy on rotated inputs.","Since the network learns the flux rather than the interface, the same flux map may be reusable inside an unsplit finite volume update or for more than two fluids by componentwise combination; whether such reuse preserves the observed convergence rates is a direct, testable consequence."],"forward_implications":["Interface reconstruction in 3D VOF can be replaced by one network evaluation, so the scheme avoids clipping and plane-positioning algorithms in its implementation.","A network trained once on synthetic planar and ellipsoidal interfaces transfers across equations, domains, and initial conditions without retraining.","On the constant-velocity test the scheme's empirical L1 convergence rate is about 1.01, versus 0.16 for upwind and 0.58 for limited downwind; on rotational and divergence-free shear tests it overtakes limited downwind once the mesh is fine enough.","Mixed-cell counts stay nearly constant when exact interface area should be constant, indicating that the learned flux controls numerical diffusion far better than upwind.","Flux reconstruction errors on the held-out test set are 43.7 times smaller in MSE than upwind and 13.2 times smaller than limited downwind, which matters because flux errors are injected at every time step."],"supporting_citations":[{"why":"Supplies the two-dimensional VOF-ML method that this paper extends to 3D, including the idea of computing the flux directly from stencil averages without interface reconstruction.","marker":"[9]"},{"why":"Extends two-component VOF schemes to an arbitrary number of components and is the basis for the claim that a network trained on pairs of fluids also serves multi-material runs.","marker":"[2]"},{"why":"Provides the contact-discontinuity capturing flux framework whose bound constraints and projection are used to keep volume fractions in [0,1].","marker":"[10]"},{"why":"The Quickhull convex-hull algorithm is used to compute exact volume fractions for half-space intersections and approximated ellipsoid volumes in the training dataset.","marker":"[3]"},{"why":"The probabilistic analysis of the upwind scheme provides the theoretical 0.5 L1 convergence rate used to validate the reference implementation.","marker":"[8]"},{"why":"A directionally-split VOF technique that motivates the splitting strategy and one of the divergence-free velocity tests.","marker":"[13]"},{"why":"The Fibonacci spiral point distribution is used to discretize ellipsoids into polytopes for accurate volume computation in training data.","marker":"[19]"},{"why":"Supplies the octahedral group structure used to define the flux-preserving symmetries S= and the enrichment group S≠.","marker":"[23]"}],"fun_headline_variants":["Neural flux bypasses interface reconstruction in 3D VOF","Trained network accelerates 3D VOF convergence, no rebuild","ML flux from synthetic data lifts 3D VOF convergence rate","3D VOF without reconstruction: neural net predicts flux","Learned flux outperforms two schemes in 3D VOF advection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the exact flux is determined by a 3×3×3 stencil of volume fractions plus one Courant number, and that interfaces built from half-spaces and ellipsoids are representative enough that the network stays accurate on the curved and multi-scale interfaces that appear in real simulations.","fun_headline_variants_meta":{"raw":{"variants":["Neural flux bypasses interface reconstruction in 3D VOF","Trained network accelerates 3D VOF convergence, no rebuild","ML flux from synthetic data lifts 3D VOF convergence rate","3D VOF without reconstruction: neural net predicts flux","Learned flux outperforms two schemes in 3D VOF advection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000452,"raw_usage":{"total_tokens":2222,"prompt_tokens":840,"completion_tokens":1382,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":1291}},"tokens_in":456,"tokens_out":1382,"duration_ms":13377,"temperature":1.0,"reasoning_tokens":1291,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:31:30.558409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Advect an initial condition whose interface is outside the training family, such as a thin cylindrical shell, a sharp wedge, or a three-material junction, on meshes with $N_h = 27, 54, 105$ and compare the final L1 error with limited downwind; if the VOF-ML error plateaus or its convergence rate drops below that of limited downwind, the transferability claim fails. A cheaper check is to evaluate the network on synthetically generated stencils from such shapes and measure whether the flux error remains competitive with the test-set error.","supporting_citations":[{"cited_title":"Despr´es and H","cited_arxiv_id":null,"evidence_quote":"Supplies the two-dimensional VOF-ML method that this paper extends to 3D, including the idea of computing the flux directly from stencil averages without interface reconstruction."},{"cited_title":"Ancellin, B","cited_arxiv_id":null,"evidence_quote":"Extends two-component VOF schemes to an arbitrary number of components and is the basis for the claim that a network trained on pairs of fluids also serves multi-material runs."},{"cited_title":"Despr ´es and F","cited_arxiv_id":null,"evidence_quote":"Provides the contact-discontinuity capturing flux framework whose bound constraints and projection are used to keep volume fractions in [0,1]."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Quickhull convex-hull algorithm is used to compute exact volume fractions for half-space intersections and approximated ellipsoid volumes in the training dataset."},{"cited_title":"Delarue and F","cited_arxiv_id":null,"evidence_quote":"The probabilistic analysis of the upwind scheme provides the theoretical 0.5 L1 convergence rate used to validate the reference implementation."},{"cited_title":"F akhreddine, K","cited_arxiv_id":null,"evidence_quote":"A directionally-split VOF technique that motivates the splitting strategy and one of the divergence-free velocity tests."},{"cited_title":"Keinert, M","cited_arxiv_id":null,"evidence_quote":"The Fibonacci spiral point distribution is used to discretize ellipsoids into polytopes for accurate volume computation in training data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the octahedral group structure used to define the flux-preserving symmetries S= and the enrichment group S≠."}],"review_version":1}