{"id":"f71b9ac1-dcb5-4b58-a78f-08a0be7c0d5d","arxiv_id":"2502.04562","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"POU-MOR-Physics uses spatially gated mixtures of Fourier neural operators to impose boundary conditions, select expert models, and learn an LES closure for Re=1000 channel flow with uncertainty.","lead":"This paper introduces POU-MOR-Physics, a neural operator that combines a partition-of-unity gating network with several Fourier-based expert operators to learn PDEs with non-periodic boundary conditions. It demonstrates the method on nonlinear operators in a disk and quarter disk, and on large eddy simulation models for turbulent channel flow with uncertainty estimates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quarter-disk validation is internally inconsistent: the manufactured data from Eq. 19 do not satisfy the Dirichlet boundary condition advertised in Eq. 20, so the demonstration of learning mixed boundary conditions is not established.","rationale":"The reader's weakest assumption focuses on the fidelity of the H1-minimizing extension. That is a plausible concern, but it is not the sharpest vulnerability: the extension map is linear and injective, so a sufficiently expressive operator could in principle learn the composition P∘E, and the disk example provides some evidence the strategy works. The more concrete and checkable flaw is that the quarter-disk data do not realize the advertised boundary value problem. On the unit circle the construction gives v=(1/2)Σcos(...), not v=1, so the learned operator is not demonstrably the solution operator for the mixed-boundary Poisson problem described in Eq. 20. The LES claims have separate reproducibility issues, but those are secondary to the missing logical support for the central boundary-condition contribution. The method is still a useful and potentially correct proposal, so a conditional verdict is appropriate with the specific condition that the quarter-disk experiment be corrected or the claim revised.","tokens_in":11985,"tokens_out":15090,"duration_ms":164748,"concrete_test":"Regenerate the quarter-disk dataset exactly as specified in Eq. 19 (correcting the typo so \\u is defined via \\v) and evaluate the residual v−1 on the circular boundary and ∇v·n on x_1=0 and x_2=0. If the circular residual is not at machine precision, the stated BVP in Eq. 20 is not the one being learned; re-run the experiment with data that genuinely solves the mixed-boundary problem, for example by solving the PDE numerically with those BCs, and recompute the relative RMSE.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 3.2's central demonstration is invalid as stated. The data are manufactured by setting ψ(x)=1−||x||²/2 on the unit disk, \\v = ψ Σ cos(f_{m,1}x_1)cos(f_{m,2}x_2), and \\u = ∇·tanh(∇\\v), then restricting to the quarter disk. The paper claims this data solves ∇·tanh(∇v)=u with v=1 on the circular arc and ∇v·n=0 on the straight edges (Eq. 20). The Neumann conditions on the axes do hold, because the cosine sum has zero normal derivative there. However, the Dirichlet condition does not: on the unit circle, ψ=1/2, so v=(1/2)Σcos(...)≠1 on the arc. The reported ~1% relative RMSE therefore validates approximation of a manufactured mapping, not solution of the advertised mixed-boundary BVP. Since the paper's core claim includes recovering operators with non-trivial boundary conditions, the clean two-dimensional test of general boundary conditions does not support it. This is correctable, but the demonstration must be regenerated with data satisfying the stated BCs, or the claim must be revised.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces POU-MOR-Physics, a mixture-of-experts neural operator that combines Fourier-based MOR-Physics experts with a spatially conditioned partition-of-unity gating network. The method is designed to learn nonlinear operators on non-periodic domains by first smoothly extending input functions to the periodic torus and then letting spatially localized experts, including a zero expert, enforce boundary behavior and select among submodels. The authors demonstrate the approach on a synthetic nonlinear operator on the unit disk, a nonlinear Poisson problem with mixed boundary conditions on the quarter disk, and a large eddy simulation (LES) closure for Re=1000 turbulent channel flow using the JHTDB database, the last with mean-field variational inference to produce uncertainty estimates and out-of-distribution detection. The central claims are that the method recovers non-trivial boundary conditions, performs interpretable model selection, and yields LES predictions with calibrated-looking uncertainties far beyond the DNS training horizon.","tokens_in":12313,"tokens_out":8198,"duration_ms":79394,"significance":"If the results hold, the paper makes a useful conceptual contribution: the volume-penalization / mixture-of-experts analogy gives an interpretable way to impose boundary conditions in Fourier-based neural operators, and the learned zero expert is a clean mechanism for no-slip and Dirichlet behavior. The 2D disk example is well executed with very strong quantitative validation (R2>99.999%), and the visual expert partitions are compelling. The extension to LES with Bayesian variational inference, including an OOD detection test, addresses an important practical need. However, the paper does not release code or detailed training configurations, and, as detailed below, the quarter-disk boundary-condition demonstration and the LES validation are not yet sufficient to support the strength of the claims.","major_comments":[{"comment":"The manufactured data for the quarter-disk Poisson problem do not satisfy the boundary conditions stated in Eq. (20). With ψ(x)=1−||x||²/2, on the unit circular arc ||x||=1 one has ψ=1/2, so v=(1/2)∑cos(f_{m,1}x1)cos(f_{m,2}x2)≠1, contradicting the advertised Dirichlet condition v=1 on that arc. In addition, Eq. (19) writes 'ˆu = ∇·tanh(∇u)', which mixes u and v; the intended relation is presumably u=∇·tanh(∇v), as in Eq. (20). As a result, the reported ~1% relative RMSE validates approximation of a manufactured mapping, not solution of the advertised mixed-boundary BVP. This is a load-bearing issue because the abstract and contributions explicitly cite recovery of non-trivial boundary conditions. The demonstration must be regenerated with data satisfying the stated BCs, or the claim about learning mixed boundary conditions must be revised.","section":"Section 3.2, Eqs. (19)-(20)"},{"comment":"The LES evidence is too thin to support the central claim of 'accurately modeling Re=1000 3d wall bounded turbulence via neural operators and quantifying uncertainties.' The model is trained and evaluated on a single DNS trajectory, the deterministic R2=98.8% is reported without error bars or a specification of which fields and time steps it covers, and the spectral and statistical comparisons in Figure 10 are qualitative. The OOD experiment in Figure 11 is a single illustrative case. I recommend reporting ensemble or rollout statistics over multiple initial conditions, quantitative metrics with confidence intervals, and a calibration check for the MFVI predictive variance before the LES claims can be accepted.","section":"Section 3.3, Table 1 and Figure 10"},{"comment":"As written, the LES update operator in Eq. (24) is not the Chorin projection it claims to be: the pressure-projection term has the opposite sign to the nonlinear term (so the updated field need not be divergence-free), the viscous term appears as -||κ||²/2 rather than νΔt||κ||², and no time step Δt or viscosity ν is defined. If these are notation or convention choices, please state them explicitly; otherwise the 'a priori known physics' component of the LES model is not correctly specified.","section":"Appendix C, Eq. (24)"}],"minor_comments":[{"comment":"The line 'ˆu = ∇·tanh(∇u)' uses u on both sides; it should presumably read u=∇·tanh(∇v), consistent with Eq. (20).","section":"Section 3.2, Eq. (19)"},{"comment":"The text refers to 'Kologmorov’s 3/5’s rule' and the caption to 'Kolmogorov’s 5/3 rule'; the scaling exponent should be 5/3 consistently, and the typo 'Kologmorov' should be corrected.","section":"Section 3.3 and Figure 10"},{"comment":"The paper reports validation R2>99.999% but does not state the train/test split, the Gaussian process kernel, the number of experts (I), the number of MOR layers (L), or the hyperparameters; adding a table of architecture and training details would aid reproducibility.","section":"Section 3.1"},{"comment":"Table 1 lists 'Time Dimension t=4000' but the text says the full time resolution is kept; please clarify whether the 4000 frames are all DNS output times or a subsample, and define the box filter and spatial stride precisely.","section":"Section 3.3 and Table 1"},{"comment":"The gating network is described as taking only coordinates as inputs, but Figure 8 shows a 3D partition; the paragraph should state the input dimension and architecture of the gating network for the LES case.","section":"Section 2.2.2"},{"comment":"The manuscript does not include a data/code availability statement. Since the method is implemented with a complex distributed training setup, release of code or at least pseudocode and hyperparameters would be important for verification.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript leans heavily on the authors' own earlier MOR-Physics and POU-Net work; this is acceptable as building blocks, but the novelty boundary with respect to those works should be drawn more sharply. Given the LES section's dependence on a single public dataset and the absence of code release, I would ask the authors to provide the training configuration and, if possible, code for verification. The journal fit is reasonable for a machine-learning venue, but the current validation depth is below the bar for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper combines partition-of-unity gating with MOR-Physics experts and a zero expert to handle non-periodic boundary conditions in Fourier-based neural operators. That combination is genuinely new, and the framing of volume penalization as a mixture-of-experts mechanism is a useful conceptual link. The 2D disk example is clean: R2 > 99.999% on held-out data, and the learned zero-expert partition is exactly what you want near the boundary. The LES section is also ambitious: extracting a closure from JHTDB channel flow at Re=1000 with variational UQ and showing domain-decomposed expert weights is ahead of the prior FNO-based wall-modeling work. The OOD detection via uncertainty is a nice bonus.\n\nBut there is a real problem in the quarter-disk section. The manufactured data come from multiplying a cosine sum by psi = 1 - ||x||^2/2. On the circular arc, ||x||^2/2 = 1/2, so v = 0.5 * sum(cos(...)), not 1. The data therefore do not satisfy the advertised Dirichlet condition v=1 on the arc — they only satisfy the Neumann conditions on the axes. The ~1% RMSE validates approximation of a manufactured mapping, not solution of the mixed-boundary BVP claimed in Eq. 20. This is load-bearing for the general-BC claim. It is fixable by regenerating the data with the correct BC, but the current demonstration does not support the headline claim about learning general boundary conditions.\n\nThe other soft spots are minor by comparison: one LES training run, no released code, no error bars on the deterministic R2=98.8%, qualitative spectral comparison, and a Kolmogorov 5/3 / 3/5 slip in the text. The self-citation to MOR-Physics and POU-Net is heavy but those are building blocks, not the target result. The smooth H1 extension is the real load-bearing assumption; the paper should isolate its contribution with an ablation. None of these are disqualifying on their own.\n\nBottom line: this deserves peer review, not a desk reject. The architecture is novel, the disk example is solid, and the LES experiment is worth referee time. The quarter-disk error must be corrected before acceptance, and the regenerated data should be checked against the stated BCs. I would bring it to reading group and, once the BC demonstration is straightened out, I would cite it.\n\nFair but firm: major revision.","headline":"The POU-MOR-Physics combination is real and the disk result is clean, but the quarter-disk demonstration does not satisfy the advertised boundary condition; it deserves peer review with major revision.","tokens_in":12799,"tokens_out":2046,"would_cite":false,"duration_ms":22628,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A mixture of Fourier operator experts with partition-of-unity gating can learn non-periodic boundary conditions, select submodels, and produce an LES closure for turbulent channel flow.","keywords":["neural operators","mixture of experts","partition of unity","boundary conditions","Fourier methods","large eddy simulation","turbulent channel flow","variational inference"],"falsifier":"Train POU-MOR-Physics on a domain with a reentrant corner, such as an L-shaped domain, and measure pointwise validation error against distance from the corner; if the error concentrates at the corner while the interior stays accurate, the $H^1$-extension premise is the limiting factor and the boundary-condition claim is only partially true.","tokens_in":11785,"feed_emoji":"🌊","tokens_out":11034,"duration_ms":90999,"temperature":0.7,"pith_summary":"Fourier-based neural operators are fast and simple but assume periodic domains, which rules out most real engineering geometries. This paper claims that the limitation can be overcome by mixing several MOR-Physics expert operators with a partition-of-unity gating network that is conditioned only on spatial location, in the spirit of volume penalization. The authors show the resulting POU-MOR-Physics model recovers a nonlinear operator on a disk and a quarter disk, and use it to extract a large eddy simulation closure for Re=1000 turbulent channel flow with Bayesian variational inference that produces posterior predictive samples far past the DNS training horizon. The payoff, if the method holds, is that Fourier operators gain access to domains with walls and mixed boundary conditions while keeping their efficiency and gaining an interpretable spatial decomposition.","feed_headline":"Mixture of Fourier experts learns boundaries and picks submodels","feed_subtitle":"Spatial gating lets Fourier neural operators handle walls, pick submodels, and quantify turbulence uncertainty.","key_machinery":"The central object is the partition-of-unity mixture $P(u)(x)=\\sum_{i} G_i(x)\\, N_i(u)(x)$, where the $N_i$ are MOR-Physics operators acting in Fourier space and the $G_i$ are softmax-normalized gates that form a partition of unity and depend only on the spatial coordinate $x$. The load-bearing companion is the smooth extension of input functions to the periodic torus via the constrained problem $\\min_{u_e} \\int \\nabla u_e \\cdot \\nabla u_e \\, dx$ subject to $R u_e = u$, solved with a matrix-free conjugate gradient method; this extension is what prevents Gibbs phenomena and carries the boundary information into the Fourier representation.","core_discovery":"POU-MOR-Physics is a composite operator in which each expert is a MOR-Physics operator, a Fourier-domain composition of learned activations and modal weightings, and the experts are combined by softmax gates that depend only on the spatial coordinate. Inputs defined on a nontrivial domain are first extended to the periodic torus by minimizing the $H^1$ semi-norm subject to matching the data on the original domain, which suppresses Gibbs oscillations; the output is then restricted back. The mixture lets the model learn boundary behavior as a volumetric forcing, and one expert can be fixed to zero to encode no-slip walls. On the unit disk the validation $R^2$ exceeds 99.999%; on the quarter disk the learned solution operator achieves about 1% relative RMSE; and for Re=1000 channel flow a deterministic version reproduces filtered DNS fields with $R^2=98.8\\%$, while the Bayesian version matches the energy spectrum and mean velocity statistics at ten flow-through times.","pith_inferences":["Because the gating depends only on position, the learned decomposition is static; conditioning the gates on the input field as well would let the model adapt its submodels to changing flow regimes, an extension the authors do not explore.","The $H^1$-extension step is the most fragile part of the pipeline for domains with reentrant corners, where the minimizer itself develops a singularity; a testable extension would replace the $H^1$ objective with a weighted or boundary-adapted norm and compare errors near corners.","The volume-penalization connection suggests a direct use beyond boundaries: the same mixture could discover coupled multi-physics systems, such as fluid-structure interaction, by learning separate experts for each physical regime without prescribing the interface.","For LES, the uncertainty estimates could be turned into an online criterion for when the learned closure should yield to a coarse DNS solve, since the model already detects out-of-distribution initial fields from a single time slice."],"forward_implications":["Non-periodic boundary conditions become a learnable feature instead of a hard constraint: the model discovers where to apply boundary forcing, as the zero expert on the disk walls demonstrates.","The gating network produces an interpretable decomposition of the domain, with channel flow experiments showing separate experts for the boundary layer, the bulk, and the wall (zero) region.","The same architecture can serve as a closure model for under-resolved simulations, since the learned LES operator is stable enough to be run autoregressively for ten flow-through times.","Variational inference turns the operator into a probabilistic model whose predictive variance grows with time and flags out-of-distribution initial conditions, useful for deciding when a learned surrogate can be trusted."],"supporting_citations":[{"why":"Supplies the MOR-Physics operator, the Fourier modal regression base that each expert is built from.","marker":"Patel and Desjardins [2018]"},{"why":"Extends the MOR-Physics framework to continuum model discovery and provides the operator architecture used in the LES experiments.","marker":"Patel et al. [2021]"},{"why":"Introduces partition-of-unity networks, the gating mechanism that spatially combines the experts.","marker":"Lee et al. [2021]"},{"why":"Provides the mixture-of-experts formulation that motivates competing expert models and model selection.","marker":"Shazeer et al. [2017]"},{"why":"Supplies the volume penalization viewpoint that motivates embedding non-periodic boundaries in a periodic domain with forcing.","marker":"Brown-Dymkoski et al. [2014]"},{"why":"Shows volume penalization applied to Fourier pseudo-spectral solvers, the discretization family the authors retain.","marker":"Kolomenskiy and Schneider [2009]"},{"why":"Supplies the turbulent channel-flow DNS data used to learn and evaluate the LES closure.","marker":"Graham et al. [2016]"},{"why":"Is the prior operator-learned LES closure for wall-bounded turbulence that this paper compares against and extends with uncertainty quantification.","marker":"Wang et al. [2024]"},{"why":"Provides the weight-uncertainty variational inference method that yields the posterior predictive samples and out-of-distribution detection.","marker":"Blundell et al. [2015]"}],"fun_headline_variants":["Fourier experts gated in space learn boundary conditions and submodels","Spatially gated Fourier experts learn boundaries and pick submodels","Neural operator experts master boundaries and pick submodels","Expert gating lets Fourier operators handle walls and choose models","Mixture-of-experts neural operators adapt to boundaries and select models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's success rests on the smooth $H^1$-minimizing extension of the input to the periodic torus retaining enough boundary and interior information that the Fourier experts can learn the true restricted operator; if that extension destroys information, every learned operator built on it is compromised.","fun_headline_variants_meta":{"raw":{"variants":["Fourier experts gated in space learn boundary conditions and submodels","Spatially gated Fourier experts learn boundaries and pick submodels","Neural operator experts master boundaries and pick submodels","Expert gating lets Fourier operators handle walls and choose models","Mixture-of-experts neural operators adapt to boundaries and select models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001187,"raw_usage":{"total_tokens":4880,"prompt_tokens":904,"completion_tokens":3976,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":3889}},"tokens_in":520,"tokens_out":3976,"duration_ms":24802,"temperature":1.0,"reasoning_tokens":3889,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T22:17:27.848285+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train POU-MOR-Physics on a domain with a reentrant corner, such as an L-shaped domain, and measure pointwise validation error against distance from the corner; if the error concentrates at the corner while the interior stays accurate, the $H^1$-extension premise is the limiting factor and the boundary-condition claim is only partially true.","supporting_citations":[{"cited_title":"A characteristic based volume penalization method for general evolution problems applied to compressible viscous flows","cited_arxiv_id":null,"evidence_quote":"Supplies the volume penalization viewpoint that motivates embedding non-periodic boundaries in a periodic domain with forcing."},{"cited_title":"A fourier spectral method for the navier--stokes equations with volume penalization for moving solid obstacles","cited_arxiv_id":null,"evidence_quote":"Shows volume penalization applied to Fourier pseudo-spectral solvers, the discretization family the authors retain."},{"cited_title":"A web services accessible database of turbulent channel flow and its use for testing a new integral wall model for les","cited_arxiv_id":null,"evidence_quote":"Supplies the turbulent channel-flow DNS data used to learn and evaluate the LES closure."},{"cited_title":"Prediction of turbulent channel flow using fourier neural operator-based machine-learning strategy","cited_arxiv_id":null,"evidence_quote":"Is the prior operator-learned LES closure for wall-bounded turbulence that this paper compares against and extends with uncertainty quantification."}],"review_version":1}