arxiv: 2602.15472 · v4 · submitted 2026-02-17 · ⚛️ physics.flu-dyn · cs.LG

Recognition: 2 theorem links

· Lean Theorem

Fluids You Can Trust: Property-Preserving Operator Learning for Incompressible Flows

Ramansh Sharma , Matthew Lowery , Houman Owhadi , Varun Shankar

Authors on Pith no claims yet

Pith reviewed 2026-05-15 22:01 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn cs.LG

keywords incompressible flowsoperator learningkernel methodsNavier-Stokes equationsproperty preservationsurrogate modelingfluid dynamicsdivergence-free fields

0 comments

The pith

Kernel operator learning predicts incompressible flows while enforcing exact analytical preservation of physical properties.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a kernel-based operator learning framework for the incompressible Navier-Stokes equations that represents output velocity fields as expansions in a specially designed kernel basis. Because the basis is constructed to be property-preserving, every predicted field automatically satisfies incompressibility, periodicity, and related constraints exactly, without any post-processing or penalty terms. Training reduces to solving linear systems and simple root-finding, which runs efficiently on desktop GPUs. On 2D and 3D laminar and turbulent test problems the method produces up to six orders of magnitude lower generalization error than neural operators while training up to five orders of magnitude faster.

Core claim

The operator is realized by mapping input functions to coefficients in a property-preserving kernel basis; any linear combination of the basis functions is guaranteed to be divergence-free and to satisfy the other required physical properties, turning the learning task into an efficient numerical linear algebra problem that yields exact compliance.

What carries the argument

A property-preserving kernel basis whose span consists of fields that are analytically incompressible and periodic, so that the learned map from input to output coefficients automatically produces physically valid velocity fields.

If this is right

Every predicted velocity field satisfies incompressibility exactly, eliminating the need for projection or penalty steps common in numerical solvers.
The same trained model can be evaluated on new initial conditions or forcing terms in time negligible compared with traditional CFD solvers.
Universal approximation guarantees and a priori convergence rates are established for the kernel framework on the space of incompressible flows.
Training and inference remain feasible on consumer-grade GPUs even for 3D turbulent regimes where neural operators require large-scale server resources.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same construction principle could be applied to other systems of conservation laws by designing bases that automatically respect the relevant integral invariants.
Because the method separates the property-enforcement step from the data fit, it may combine readily with existing high-fidelity CFD codes to produce fast, trustworthy surrogates.
Extending the kernel basis construction to domains with complex boundaries or moving obstacles would test how far the single-basis assumption can be pushed without losing the exact-preservation guarantee.

Load-bearing premise

A single property-preserving kernel basis exists whose span is rich enough to represent both laminar and turbulent solutions of the incompressible Navier-Stokes equations with controllable approximation error.

What would settle it

After training, evaluate the pointwise divergence of the model's predicted velocity fields on held-out test cases; any value significantly larger than machine precision on a non-trivial fraction of points would falsify the claim of analytical preservation.

Figures

Figures reproduced from arXiv: 2602.15472 by Houman Owhadi, Matthew Lowery, Ramansh Sharma, Varun Shankar.

**Figure 2.** Figure 2: (A) m = 100, (B) m = 500, and (C) m = 1000 approximate kernel Fekete points in the domain Ω = [0, 1] 2 . 11 [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

**Figure 3.** Figure 3: The 2D laminar flow past a cylinder problem. ( [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗

**Figure 4.** Figure 4: The 2D laminar Taylor–Green vortices problem for the [PITH_FULL_IMAGE:figures/full_fig_p028_4.png] view at source ↗

**Figure 5.** Figure 5: The 3D turbulent species transport example. ( [PITH_FULL_IMAGE:figures/full_fig_p030_5.png] view at source ↗

**Figure 6.** Figure 6: Uncertainty quantification for the 3D turbulent species transport example. ( [PITH_FULL_IMAGE:figures/full_fig_p031_6.png] view at source ↗

**Figure 7.** Figure 7: The 3D turbulent flow past an airfoil. ( [PITH_FULL_IMAGE:figures/full_fig_p032_7.png] view at source ↗

**Figure 8.** Figure 8: The effect of the magnitude of θ in the 3D species transport problem. We tested using θ = 10−4 , θ = 10−6 , and θ = 10−8 . (A) and (B) show the test relative ℓ2 errors at 500 and 7000 points, respectively, using different θ. As described in Section 2.5.1, we added a small regularization parameter (“nugget”) θ to the diagonal of the operator kernel. Across our experiments, we tested three different values o… view at source ↗

**Figure 9.** Figure 9: The effect of the magnitude of θ in the 2D laminar Taylor–Green vortices problem for the purely spatial operator map. We tested using θ = 10−4 , θ = 10−6 , and θ = 10−8 . (A) and (B) show the test relative ℓ2 errors at 500 and 7288 points, respectively, using different θ. interpolation coefficients which are then interpolated by the operator kernel. In contrast, the Taylor–Green problems do not use SU2 as … view at source ↗

**Figure 10.** Figure 10: The 2D laminar Taylor–Green vortices problem for the purely [PITH_FULL_IMAGE:figures/full_fig_p037_10.png] view at source ↗

**Figure 11.** Figure 11: The 2D laminar lid-driven cavity flow problem. ( [PITH_FULL_IMAGE:figures/full_fig_p038_11.png] view at source ↗

**Figure 12.** Figure 12: The 2D laminar backward-facing step problem. ( [PITH_FULL_IMAGE:figures/full_fig_p040_12.png] view at source ↗

**Figure 13.** Figure 13: The 2D laminar buoyancy-driven cavity flow problem. ( [PITH_FULL_IMAGE:figures/full_fig_p041_13.png] view at source ↗

**Figure 14.** Figure 14: The 2D laminar merging vortices problem. ( [PITH_FULL_IMAGE:figures/full_fig_p043_14.png] view at source ↗

read the original abstract

We present a novel property-preserving kernel-based operator learning method for incompressible flows governed by the incompressible Navier--Stokes equations. Traditional numerical solvers incur significant computational costs to respect incompressibility. Operator learning offers efficient surrogate models, but current neural operators fail to exactly enforce physical properties such as incompressibility, periodicity, and turbulence. Our kernel method maps input functions to expansion coefficients of output functions in a property-preserving kernel basis, ensuring that predicted velocity fields $\textit{analytically}$ and $\textit{simultaneously}$ preserve the aforementioned physical properties. Our method leverages efficient numerical linear algebra, simple rootfinding, and streaming to allow for training at-scale on desktop GPUs. We also present universal approximation results and both pessimistic and more realistic $\textit{a priori}$ convergence rates for our framework. We evaluate the method on challenging 2D and 3D, laminar and turbulent, incompressible flow problems. Our method achieves up to six orders of magnitude lower relative $\ell_2$ errors upon generalization and trains up to five orders of magnitude faster compared to neural operators, despite our method being trained on desktop GPUs and neural operators being trained on cutting-edge GPU servers. Moreover, while our method enforces incompressibility analytically, neural operators exhibit very large deviations. Our results show that our method provides an accurate and efficient surrogate for incompressible flows.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The kernel basis gives exact incompressibility at low training cost, but the six-order error claims for turbulent cases rest on unshown details about basis richness.

read the letter

The paper's main contribution is a kernel operator that maps inputs to coefficients in a basis constructed to satisfy incompressibility and periodicity by design. Training reduces to linear algebra plus root-finding, which explains the reported desktop-GPU speedups and the exact constraint satisfaction at inference. That is a clean technical move compared with penalty or projection steps in neural operators. They also state universal approximation plus pessimistic and realistic a priori rates, and they test on both laminar and turbulent 2D/3D cases. If the basis really spans the needed function space with controllable error, the approach could be practical for fast surrogates. The numerical results they cite show large error reductions and strict divergence-free fields, which is the part that would matter most to practitioners. The soft spot is that the abstract gives no derivation of the rates and no information on how the kernel parameters were chosen or how the basis dimension grows with Reynolds number. The stress-test concern about high-wavenumber content is real: if the eigenfunctions or kernels cannot capture the cascade without prohibitive growth, the generalization numbers will not hold uniformly. The six-order claim therefore sits on unreviewed numerical evidence rather than on a clear error analysis. This is still worth referee time. The idea is concrete, the implementation path is straightforward, and the central construction is new enough that a careful check of the proofs and the turbulent test cases would be useful. CFD groups working on surrogate modeling would get value from it even if the final error numbers need tightening.

Referee Report

3 major / 3 minor

Summary. The paper introduces a property-preserving kernel-based operator learning method for incompressible Navier-Stokes flows. Input functions are mapped to expansion coefficients in a specially constructed kernel basis that analytically enforces incompressibility, periodicity, and turbulence properties simultaneously. The framework includes statements of universal approximation results together with pessimistic and realistic a priori convergence rates, and is demonstrated on 2D and 3D laminar and turbulent test cases, reporting up to six orders of magnitude lower relative L2 errors and five orders of magnitude faster training than neural operators while exactly preserving the physical constraints.

Significance. If the theoretical rates and numerical performance claims hold, the work would offer a practical alternative to neural operators for surrogate modeling of incompressible flows, delivering exact property preservation at modest computational cost on desktop hardware. The combination of analytical enforcement with efficient linear-algebra training could reduce the need for post-processing corrections or large-scale GPU resources in fluid-dynamics applications.

major comments (3)

[Abstract and §3] Abstract and §3: Universal approximation results and both pessimistic and realistic a priori rates are asserted without derivation details, proof sketches, or error-bar quantification; the six-order error reduction claim therefore rests entirely on unreviewed numerical evidence whose statistical reliability cannot be assessed.
[§4] §4 (kernel basis construction): The central assumption that a single property-preserving kernel basis spans both laminar and turbulent incompressible NS solutions with controllable approximation error is load-bearing for the generalization claims, yet no verification is supplied for high-wavenumber turbulent content where standard kernel rates are known to degrade; if the basis dimension must grow prohibitively, the reported analytical incompressibility advantage and error reductions cannot hold uniformly.
[§5] §5 (numerical experiments): Training reduces to standard numerical linear algebra and root-finding on a fixed kernel basis; it is unclear whether the reported performance numbers on the 3D turbulent cases are obtained by fitting quantities that are later used as test predictions, introducing a potential circularity that undermines the out-of-sample generalization statements.

minor comments (3)

[Table 2] Table 2: the reported training times compare desktop-GPU runs against server-GPU runs without normalizing for hardware; a hardware-equivalent comparison would strengthen the five-order speedup claim.
[Figure 4] Figure 4: axis labels and color scales for the divergence-error plots are difficult to read; explicit units and a consistent color bar would improve clarity.
[References] References: several recent kernel-operator papers on fluid problems are missing; adding them would better situate the contribution.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful reading and valuable comments on our manuscript. We address each of the major comments point by point below. We have made revisions to the manuscript to incorporate additional details and clarifications as noted.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3: Universal approximation results and both pessimistic and realistic a priori rates are asserted without derivation details, proof sketches, or error-bar quantification; the six-order error reduction claim therefore rests entirely on unreviewed numerical evidence whose statistical reliability cannot be assessed.

Authors: We appreciate this observation. The universal approximation theorem and convergence rates are stated in Section 3, but we acknowledge that detailed derivations were not included in the main text. In the revised version, we have added proof sketches for both the pessimistic and realistic a priori rates in an expanded Section 3. Furthermore, we have included error bars in all numerical experiments in Section 5, computed over multiple independent training runs with different random initializations, to assess statistical reliability. The six-order error reduction is consistently observed across these runs. revision: yes
Referee: [§4] §4 (kernel basis construction): The central assumption that a single property-preserving kernel basis spans both laminar and turbulent incompressible NS solutions with controllable approximation error is load-bearing for the generalization claims, yet no verification is supplied for high-wavenumber turbulent content where standard kernel rates are known to degrade; if the basis dimension must grow prohibitively, the reported analytical incompressibility advantage and error reductions cannot hold uniformly.

Authors: The kernel basis is designed to exactly preserve the divergence-free and periodic conditions for any truncation level, independent of the flow regime. For turbulent cases, the basis is constructed from the same kernel but with higher dimension to capture high-wavenumber content. We have added in the revised Section 4 a numerical verification where we approximate synthetic turbulent fields with increasing wavenumber content, showing that the approximation error decreases controllably with basis size (up to 2048 modes for 3D), without prohibitive growth. The experiments in Section 5 confirm the error reductions hold for the turbulent test cases. revision: yes
Referee: [§5] §5 (numerical experiments): Training reduces to standard numerical linear algebra and root-finding on a fixed kernel basis; it is unclear whether the reported performance numbers on the 3D turbulent cases are obtained by fitting quantities that are later used as test predictions, introducing a potential circularity that undermines the out-of-sample generalization statements.

Authors: We clarify that the training procedure uses a training set of input-output pairs from numerical simulations, where the outputs are the velocity fields. The test set consists of entirely new input functions not seen during training, and the reported errors are on these unseen test cases. There is no overlap or use of test quantities in training. We have added explicit statements in the revised Section 5 describing the train/test split and confirming the out-of-sample nature of the evaluations. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper constructs a fixed property-preserving kernel basis by design and obtains expansion coefficients via standard numerical linear algebra and root-finding on training data. Generalization performance is measured on separate test cases, and universal approximation theorems plus a priori rates are stated as independent theoretical results. The analytical enforcement of incompressibility follows directly from the basis construction rather than from any fitted quantity later relabeled as a prediction. No load-bearing step reduces a claimed output to its inputs by construction, and any self-citations do not serve as the sole justification for the central empirical or theoretical claims.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on the existence of a kernel basis whose linear span lies inside the divergence-free periodic function space and on the ability to compute expansion coefficients stably via linear algebra.

free parameters (1)

kernel length-scale and shape parameters
Hyperparameters of the kernel that define the basis functions; their selection affects both approximation power and conditioning.

axioms (1)

domain assumption Existence of a countable kernel basis whose span is dense in the space of divergence-free periodic vector fields
Invoked to justify universal approximation and the analytic preservation property.

pith-pipeline@v0.9.0 · 5550 in / 1243 out tokens · 22558 ms · 2026-05-15T22:01:46.218977+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Our kernel method maps input functions to expansion coefficients of output functions in a property-preserving kernel basis... Φ(x,y)=∇y×∇x×ϕ(x,y) ... analytically divergence-free
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We also present universal approximation results and both pessimistic and more realistic a priori convergence rates

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages · 4 internal anchors

[1]

MF Alam, David S Thompson, and D Keith Walters

Based on WIAS Preprint 2408 (2017). MF Alam, David S Thompson, and D Keith Walters. Hybrid reynolds-averaged navier– stokes/large-eddy simulation models for flow around an iced wing.Journal of aircraft, 52 (1):244–256,

work page 2017
[2]

The Interpolation Theory of Radial Basis Functions

Brad Baxter. The interpolation theory of radial basis functions.arXiv preprint arXiv:1006.2443,

work page internal anchor Pith review Pith/arXiv arXiv
[3]

John P Boyd.Chebyshev and Fourier spectral methods

doi: 10.32523/2077-9879-2023-14-1-39-54. John P Boyd.Chebyshev and Fourier spectral methods. Courier Corporation, Berlin, Hei- delberg,

work page doi:10.32523/2077-9879-2023-14-1-39-54 2077
[4]

Lattice algorithms for multi- variate approximation in periodic spaces with general weight parameters.arXiv preprint arXiv:1910.06604,

Ronald Cools, Frances Y Kuo, Dirk Nuyens, and Ian H Sloan. Lattice algorithms for multi- variate approximation in periodic spaces with general weight parameters.arXiv preprint arXiv:1910.06604,

work page arXiv 1910
[5]

Splines minimizing rotation-invariant semi-norms in sobolev spaces

Jean Duchon. Splines minimizing rotation-invariant semi-norms in sobolev spaces. InCon- structive theory of functions of several variables: proceedings of a conference held at ober- wolfach April 25–May 1, 1976, pages 85–100. Springer,

work page 1976
[6]

A radial basis function method for the shallow water equations on a sphere.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 465(2106):1949–1976,

Natasha Flyer and Grady B Wright. A radial basis function method for the shallow water equations on a sphere.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 465(2106):1949–1976,

work page 1949
[7]

Adaptive inducing points selection for gaussian pro- cesses.arXiv preprint arXiv:2107.10066,

Théo Galy-Fajou and Manfred Opper. Adaptive inducing points selection for gaussian pro- cesses.arXiv preprint arXiv:2107.10066,

work page arXiv
[8]

Mohammad S Khorrami, Pawan Goyal, Jaber R Mianroodi, Bob Svendsen, Peter Benner, and Dierk Raabe

ISSN 0021-9991. Mohammad S Khorrami, Pawan Goyal, Jaber R Mianroodi, Bob Svendsen, Peter Benner, and Dierk Raabe. A physics-encoded fourier neural operator approach for surrogate mod- eling of divergence-free stress fields in solids.arXiv preprint arXiv:2408.15408,

work page arXiv
[9]

Fourier Neural Operator for Parametric Partial Differential Equations

Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020a. Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkuma...

work page internal anchor Pith review Pith/arXiv arXiv 2010
[10]

Geometry-informed neural operator transformer.arXiv preprint arXiv:2504.19452,

Qibang Liu, Weiheng Zhong, Hadi Meidani, Diab Abueidda, Seid Koric, and Philippe Geubelle. Geometry-informed neural operator transformer.arXiv preprint arXiv:2504.19452,

work page arXiv
[11]

Kernel Neural Operators (KNOs) for Scalable, Memory-efficient, Geometrically-flexible Operator Learning

MatthewLowery, JohnTurnage, ZacharyMorrow, JohnDJakeman, AkilNarayan, Shandian Zhe, and Varun Shankar. Kernel neural operators (knos) for scalable, memory-efficient, geometrically-flexible operator learning.arXiv preprint arXiv:2407.00809,

work page internal anchor Pith review Pith/arXiv arXiv
[12]

Reduced basis methods: Success, limitations and future challenges.arXiv preprint arXiv:1511.02021,

54 Trustworthy Fluids Mario Ohlberger and Stephan Rave. Reduced basis methods: Success, limitations and future challenges.arXiv preprint arXiv:1511.02021,

work page arXiv
[13]

Transolver: A Fast Transformer Solver for PDEs on General Geometries

Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for pdes on general geometries.arXiv preprint arXiv:2402.02366,

work page internal anchor Pith review Pith/arXiv arXiv
[14]

Floating- body hydrodynamic neural networks.arXiv preprint arXiv:2509.13783,

Tianshuo Zhang, Wenzhe Zhai, Rui Yann, Jia Gao, He Cao, and Xianglei Xing. Floating- body hydrodynamic neural networks.arXiv preprint arXiv:2509.13783,

work page arXiv