{"id":"20a56477-769b-4098-8083-df3c8bea30bd","arxiv_id":"2505.13463","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Fourier Neural Operator trained on volume-of-fluid simulations predicts 2D liquid-vapour interface evolution with validation R2=0.95 and 1-8 ms inference time.","lead":"This paper trains Fourier Neural Operators on volume-of-fluid simulation data to predict how liquid-vapour interfaces evolve over time. It reports fast, 1-8 millisecond predictions and good agreement with simulation on held-out cases, suggesting such surrogates could one day enable real-time multiphase process control.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"R2=0.95 is computed on a singular, unspecified SDF transform; it can be high even if the zero level set (the interface) is misplaced.","rationale":"The SDF metric gap is the most load-bearing because every downstream assertion—interface prediction, control readiness, generalization—depends on the FNO output actually tracking the interface. The reader's weakest assumption identifies the same gap; I agree. I also considered the Case 2 sample-level split (simulations may appear in both train and validation at different times) and the absence of error bars, but those affect the strength of the generalization evidence rather than the definition of the object being measured. The validation metrics in Section 2 are written for α, yet Table 2 reports ζ; this mismatch, together with the singularity in Eq. (31), means the reported R2=0.95 cannot be interpreted as an interface-accuracy number. The proposed contour-distance test would settle the issue directly and is feasible even without the full dataset, since it only requires the trained model's outputs and the VOF α fields. If the test passes, the conditional verdict can be upgraded; if it fails, the paper's central claim is unsupported. Therefore the reader's CONDITIONAL verdict should remain unchanged.","tokens_in":9019,"tokens_out":9615,"duration_ms":107705,"concrete_test":"On the held-out Case 2 set, extract the predicted interface as the zero level set of ζ_pred (e.g., marching squares at ζ=0) and the true interface from the VOF α field (α=0.5 contour, or the same Eq. 31 applied to α_true). Compute the mean and Hausdorff distance between the two contours, and recompute R2/MSE restricted to interface cells (0<α<1 or |ζ|<ε). If the contour error is sub-cell and the near-interface R2 is close to 0.95, the concern is resolved; if the interface is displaced by several cells despite a global R2 of 0.95, the headline claim fails. Report the result for at least two values of ε/clipping so the metric's sensitivity is known.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—very high accuracy in predicting interface evolution—rests on Table 2's R2=0.95, but that number is computed on the reconstructed field ζ from Eq. (31), not on the physical volume fraction α or on the extracted interface. Two properties make this metric unable to carry the claim. First, Eq. (31), ζ = ε·atanh(1−2α), is singular at α=0 and α=1, so as written it is undefined on most VOF cells; the actual clipping/regularization and the value of ε are never reported, so the metric is not reproducible and its scale is arbitrary. Second, even with a fixed regularization, a global R2 on ζ is dominated by bulk cells, where |ζ| is large and easy to predict, while the interface is the zero level set {ζ=0}, a small fraction of the domain. A model can systematically misplace the contour and still achieve R2≈0.95 if bulk regions are correct. The paper reports no near-interface error (|ζ|≈0 or 0<α<1), no contour/Hausdorff distance between predicted and true interfaces, and no quantitative error on α reconstructed from ζ, despite Figure 3 showing only visual comparisons. Thus the specific quantity the paper claims to predict—the liquid-vapour interface—is never directly evaluated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes using Fourier Neural Operators (FNOs) as a surrogate for two-phase volume-of-fluid simulations. The authors train an FNO on signed-distance fields derived from the volume fraction α via a regularized inverse Heaviside transformation (Eq. 31), taking as input the initial SDF plus a scalar time channel and outputting the predicted SDF at a future time. Two benchmark cases are presented: a dam-break forecasting scenario and a dataset of random liquid-blob initial conditions. The authors report validation R²=0.95 on the SDF, inference times of 1–8 ms, and compare FNO with U-Net and GCN, concluding that FNO-based surrogates are fast enough for real-time control of multiphase processes.","tokens_in":9281,"tokens_out":5934,"duration_ms":59376,"significance":"If the accuracy claim held at the level of the physical interface, the work would be a useful demonstration of neural-operator surrogates for multiphase-flow control. The manuscript has clear strengths: it uses a high-fidelity VOF solver, provides a detailed architectural description, includes a baseline comparison with U-Net and GCN, and reports inference times. However, the reported accuracy is measured on a smooth transformed field rather than on the volume fraction or the extracted interface geometry, and the transformation contains an unreported free parameter. As it stands, the central claim of highly accurate interface prediction is not established.","major_comments":[{"comment":"The central accuracy claim rests on metrics computed on the reconstructed distance function ζ, not on the physical volume fraction α or on the extracted interface. Because Eq. (31), ζ = ε·tanh⁻¹(1−2α), is singular at α=0 and α=1, the implementation must involve clipping or regularization, and the smoothing parameter ε is never reported. The reported MSE=9.72, MAE=1.69, and R²=0.95 are therefore not reproducible and are not interpretable as physical errors. Please report ε and the regularization strategy, and re-evaluate the predictions using α-based metrics (for example, errors restricted to interface cells with 0<α<1) or geometric interface measures such as the distance between predicted and true zero level sets (Hausdorff or contour distance).","section":"§3, Eq. (31) and Table 2"},{"comment":"The temporal-extrapolation claim for Case 1 is supported only by the qualitative comparison in Figure 1 and the loss curves in Figure 2; no quantitative error metric is reported for the held-out final 40% of the simulation. Since the case is explicitly presented as a test of extrapolation beyond the training horizon, the same numeric metrics used in Table 2 (MSE, MAE, R², and ideally interface-location error) should be reported for Case 1 separately.","section":"§3, Case 1"},{"comment":"The validation metrics defined in Eqs. (27)–(30) are written in terms of α, yet Table 2 reports them on SDF fields. This inconsistency matters because a global R² on a smooth field is dominated by bulk regions, so R²=0.95 can coexist with systematically misplaced interfaces. Please state explicitly which field each metric is evaluated on, and add interface-focused error statistics (e.g., errors in the band |ζ|<ε or in cells with 0<α<1) so that the reader can assess the actual interface accuracy.","section":"§2 vs §3, Table 2"},{"comment":"The bottom row of Figure 3 shows 'binary volume fractions reconstructed from each RDF', but the reconstruction procedure is not described (for example, is it a threshold at ζ=0? Is there a re-initialization step?). Without this information, the visual comparison cannot be reproduced, and no quantitative error on α is given. Please describe the reconstruction and report error statistics for α as well.","section":"§3, Figure 3"}],"minor_comments":[{"comment":"The 'Lp loss' label in Figures 2 and 4 is not defined; please specify the norm (presumably p=2) and use consistent terminology with Eq. (25).","section":"§2, Figures 2 and 4"},{"comment":"The sentence 'The dataset comprises T simulations performed using compressibleInterIsoFoam' leaves T undefined; please provide the actual number of training and validation simulations used in Case 1.","section":"§3, Case 1"},{"comment":"The statement 'Training on a N=900 base simulation dataset (yielding 1800 SDF samples)' appears in the Hardware section but actually describes the Case 2 dataset; this placement is confusing and should be moved to the relevant experimental section.","section":"§2, Hardware paragraph"},{"comment":"The caption says 'Predicted vs. ground-truth volume fraction' while the main text refers to the signed-distance field; please clarify which field is plotted.","section":"Figure 2, caption (b)"},{"comment":"The claim that the model 'successfully captures the underlying physics and extrapolates the evolution beyond the training window' is not substantiated by quantitative results for Case 1; please either soften this statement or add the missing metrics.","section":"§4, Conclusions"},{"comment":"The 1–8 ms inference time is reported without specifying the hardware, the batch size, or the number of samples averaged, and there is no comparison with the per-step CFD solve time; please provide these details to support the real-time control claim.","section":"§3, Inference time"}],"recommendation":"major_revision","confidential_remarks":"This is a promising proof-of-concept, but the evaluation protocol does not yet support the headline claim. The missing ε and the absence of any interface-specific error metric make the reported R² essentially uninterpretable. The fix is straightforward: evaluate on α and on the extracted interface, and report the missing parameters. I also note the manuscript structure interleaves Case 1 and Case 2, and several experimental details (dataset size, SDF reconstruction procedure) are incomplete, which made the review more difficult than necessary."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a plausible first application of Fourier neural operators to liquid-vapour interface evolution, with genuinely fast inference (1-8 ms) and a straightforward comparison against U-Net and GCN. But the headline accuracy number (R2=0.95) is computed on a smoothed signed-distance field, not on the interface itself, and the paper never directly measures how well the predicted zero level set matches the true interface. That gap is central, not cosmetic.\n\nWhat is actually new is narrow: the standard FNO architecture applied to volume-of-fluid data preprocessed through Eq. (31), a regularized inverse-Heaviside transform into a reconstructed distance function. The transform is a sensible preprocessing step, and the two dam-break benchmarks show visually plausible interface evolution. The inference-time claim is credible for this architecture and is the real selling point.\n\nThe soft spots, in order of severity:\n\n1. The validation metrics in Section 2 are defined on α, but Table 2 reports them on the SDF field, and no metric targets the interface location. Global R2 on an SDF is dominated by bulk cells where the field is large and easy to predict; a model can systematically misplace the zero level set and still score 0.95. There is no near-interface error, no contour/Hausdorff distance, and no quantitative error on α reconstructed from the predicted SDF. Figure 3 is visual only.\n\n2. Eq. (31) is singular at α=0 and α=1, so clipping or regularization is required, and the paper never reports ε. Without ε the reported MSE (9.72) and MAE (1.69) have no interpretable scale, and the whole metric is not reproducible as written.\n\n3. Case 1 is a single trajectory with the last 40% predicted from the first 60%; it is shown as images, with no error numbers for the extrapolation. Case 2 uses a more typical 90/10 split over 1000 random initial conditions, but still without error bars over seeds.\n\n4. No code or data is shipped, and the conclusion overreaches when it says the results 'prove' the method can control processes. The speed result is promising; the control claim does not follow from 2D, two-time-step surrogates.\n\nThe citation pattern is fine; the self-citation [29] is minor and relevant.\n\nThis paper is for researchers who want a baseline for neural-operator surrogates in multiphase flows. It is an application paper, not a methodological breakthrough.\n\nRecommendation: send it to peer review, but request major revision. The authors need to report ε, add direct interface-error metrics (near-interface error or level-set distance), and quantify Case 1 extrapolation error. If they do that, the paper is a solid contribution; as it stands, the central accuracy claim is not yet supported.","headline":"Useful first FNO-for-interface-evolution demo, but the headline R2=0.95 is on a smoothed field, not the interface, so the accuracy claim outruns the evidence.","tokens_in":9787,"tokens_out":3179,"would_cite":false,"duration_ms":31169,"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 Fourier Neural Operator trained on volume-of-fluid simulations predicts liquid-vapour interface evolution with $R^2=0.95$ and 1–8 ms inference, making real-time multiphase control plausible.","keywords":["Fourier Neural Operators","operator learning","multiphase flow","volume of fluid","interface tracking","signed distance function","surrogate modeling","real-time control"],"falsifier":"Reconstruct the volume fraction from a predicted SDF via the inverse of Eq. (31), threshold it at $\\alpha = 0.5$, and compare the resulting interface contours to the ground-truth VOF interface at $t = 0.25$ s and $t = 0.50$ s; if contour-position errors are large while SDF $R^2$ is 0.95, the central claim of accurate interface prediction fails.","tokens_in":8810,"feed_emoji":"💧","tokens_out":6317,"duration_ms":60862,"temperature":0.7,"pith_summary":"This paper sets out to show that a Fourier Neural Operator (FNO) can act as a fast, accurate surrogate for volume-of-fluid simulations of liquid-vapour flows, predicting how the phase interface evolves in time. The authors train the operator on simulation data for a dam-break flow and for freely falling liquid blobs, representing the interface by a smoothed signed-distance field. They report validation $R^2 = 0.95$ and inference times of 1–8 ms, a speed that would make the surrogate fast enough to sit inside a real-time control loop, where conventional CFD is too slow. If the claim holds, neural operators become a practical forecasting layer for multiphase industrial processes and digital twins.","feed_headline":"Neural operator predicts liquid-vapour interfaces in milliseconds","feed_subtitle":"It scores R² = 0.95 on the signed-distance field and runs in 1–8 ms, opening real-time multiphase control.","key_machinery":"The load-bearing mechanism is the Fourier Neural Operator's spectral convolution layer, which learns a kernel directly in Fourier space, truncates to a fixed number of modes, and thereby captures global spatial dependencies at a cost of $O(N\\log N + k_{\\max}N)$ per layer. To make the discontinuous volume-fraction field learnable, the paper maps $\\alpha$ to a reconstructed distance function through $\\zeta = \\varepsilon\\tanh^{-1}(1-2\\alpha)$, a regularized inverse Heaviside transform that converts the sharp interface into a smooth field. The network takes the initial RDF plus a scalar time channel as input and outputs the predicted future RDF. The smoothing transform is what lets a continuous operator-learning architecture represent a discontinuous two-phase interface, while the spectral kernel is what lets the model propagate interface information across the whole domain in each layer.","core_discovery":"The central claim is that an FNO trained on compressed volume-of-fluid data learns the operator that maps an initial interface configuration to its future evolution, and that the learned operator generalizes both in time and across initial conditions. In the dam-break case, training on the first 60% of the simulated evolution lets the model forecast the remaining 40%, including the collapse and impact dynamics. In the second case, the model trained on two snapshots from 900 random initial conditions predicts the interface at later times for unseen configurations. Accuracy is measured on reconstructed distance fields (RDFs), where the model reaches MSE 9.72, MAE 1.69, and $R^2 = 0.95$, with 1–8 ms inference per prediction. The paper concludes that this makes neural-operator-based prediction fast enough to control multiphase applications.","pith_inferences":["The paper validates on smooth RDFs; a natural next check is to threshold the predicted RDF at zero and compare the extracted interface contour against the ground-truth VOF interface, since control actions depend on where the interface is, not just on field correlation.","The transform $\\alpha \\leftrightarrow \\zeta$ is invertible for $\\alpha \\in (0,1)$, so every SDF prediction can be mapped back to a volume-fraction field; the paper does not report errors in that physical space, and doing so on the published dataset would settle how much of the $R^2$ score is inherited from smooth regions away from the interface.","The unspecified smoothing parameter $\\varepsilon$ controls the sharpness of the reconstructed distance field; reporting its value and testing sensitivity to it would indicate whether the reported accuracy is robust or tuned to a favourable smoothness.","If the method extends to 3D and to phase change, as the paper names as next steps, the same operator-learning pipeline would apply to spray atomization and evaporating films, where interface evolution currently limits control bandwidth."],"forward_implications":["Predictions at 1–8 ms make the surrogate a candidate for real-time monitoring and feedback control in multiphase systems, a regime where full CFD is too slow.","A model trained only on the early dam-break evolution can forecast subsequent impact and nonlinear deformation, suggesting the learned operator captures physics rather than memorizing the training window.","Two-snapshot training on random initial conditions generalizes to unseen configurations, which would let users build surrogates from sparse simulation or experimental data.","Because the method learns from volume-fraction fields produced by a standard VOF solver, the same pipeline can be re-trained for other two-phase geometries and fluids without changing the architecture.","The reported time scale matches the response time of many multiphase processes, so process controllers could embed the FNO as a predictive model."],"supporting_citations":[{"why":"Defines the Fourier Neural Operator architecture and the spectral convolution that the paper uses as its surrogate model.","marker":"[14]"},{"why":"Introduces the volume-of-fluid method that produces the $\\alpha$-field training data for the interface.","marker":"[5]"},{"why":"Provides the isoAdvector geometric VOF reconstruction used by the solver to advance the interface.","marker":"[10]"},{"why":"Validates the compressibleInterIsoFoam solver used to generate the two-phase simulation dataset.","marker":"[24]"},{"why":"Supplies the open-source CFD framework in which the training simulations were run.","marker":"[25]"},{"why":"Introduces the reconstructed distance function concept from volume fractions, the basis for the smoothing transform in Eq. (31).","marker":"[28]"},{"why":"Provides the continuum surface tension force model used in the simulated two-phase flows.","marker":"[27]"}],"fun_headline_variants":["FNO predicts liquid-vapour interfaces in 1–8 ms","Neural operator maps interface evolution in 1–8 ms","From 60% training, FNO foretells collapse dynamics","Real-time multiphase control via neural operator forecasts","Interface forecaster hits 95% accuracy in milliseconds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy score is computed on the smoothed reconstructed distance field and never on the physical volume fraction or on the interface location, and the smoothing parameter $\\varepsilon$ is unspecified.","fun_headline_variants_meta":{"raw":{"variants":["FNO predicts liquid-vapour interfaces in 1–8 ms","Neural operator maps interface evolution in 1–8 ms","From 60% training, FNO foretells collapse dynamics","Real-time multiphase control via neural operator forecasts","Interface forecaster hits 95% accuracy in milliseconds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001134,"raw_usage":{"total_tokens":4689,"prompt_tokens":903,"completion_tokens":3786,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":3701}},"tokens_in":519,"tokens_out":3786,"duration_ms":28344,"temperature":1.0,"reasoning_tokens":3701,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:37:07.742890+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reconstruct the volume fraction from a predicted SDF via the inverse of Eq. (31), threshold it at $\\alpha = 0.5$, and compare the resulting interface contours to the ground-truth VOF interface at $t = 0.25$ s and $t = 0.50$ s; if contour-position errors are large while SDF $R^2$ is 0.95, the central claim of accurate interface prediction fails.","supporting_citations":[{"cited_title":"V olume of fluid (vof) method for the dynamics of free boundaries.J","cited_arxiv_id":null,"evidence_quote":"Introduces the volume-of-fluid method that produces the $\\alpha$-field training data for the interface."},{"cited_title":"A new volume-of-fluid method in openfoam","cited_arxiv_id":null,"evidence_quote":"Provides the isoAdvector geometric VOF reconstruction used by the solver to advance the interface."},{"cited_title":"Validation of volume- of-fluid openfoam® isoadvector solvers using single bubble benchmarks","cited_arxiv_id":null,"evidence_quote":"Validates the compressibleInterIsoFoam solver used to generate the two-phase simulation dataset."},{"cited_title":"Openfoam: A c++ library for complex physics simulations","cited_arxiv_id":null,"evidence_quote":"Supplies the open-source CFD framework in which the training simulations were run."},{"cited_title":"Cummins, Marianne M","cited_arxiv_id":null,"evidence_quote":"Introduces the reconstructed distance function concept from volume fractions, the basis for the smoothing transform in Eq. (31)."},{"cited_title":"A continuum method for modeling surface tension","cited_arxiv_id":null,"evidence_quote":"Provides the continuum surface tension force model used in the simulated two-phase flows."}],"review_version":1}