Pith. sign in

REVIEW 2 major objections 5 minor 56 references

Physics-informed, boundary-constrained Gaussian process regression for the reconstruction of fluid flow fields

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Flow fields can be reconstructed with no boundary observations by building the no-penetration condition into a Gaussian process covariance kernel.

desk verdict Solid, honest methods paper with real contributions; the main soft spot is the unquantified quadrature error between the continuous claim and the finite-Nyström implementation. read the letter →

arxiv 2507.17582 v4 pith:U2B2NMLE submitted 2025-07-23 physics.flu-dyn stat.ML

Adrian Padilla-Segarra , Pascal Noble , Olivier Roustant , Éric Savin This is my paper
classification physics.flu-dynstat.ML MSC 60G1560G6062M3065C2076D0576M25
keywords Gaussianprocessregressionboundary-constrainedprocessesKarhunen-Loèveexpansiondivergence-freerandomfieldsslipboundaryconditionphysics-informedkernelsflowreconstructionuncertaintyquantification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper's central claim is that the no-penetration condition on an aerodynamic profile can be moved out of the observation set and into the Gaussian-process prior itself. Starting from any GP with a continuous covariance kernel, the profile boundary is treated as a compact set, and the process is modified by subtracting its Karhunen-Loève projection onto that set. The resulting process vanishes on the whole boundary in the mean-square sense, and when used as a stream function, the velocity field is automatically divergence-free and tangential to the profile at every boundary point. The numerical implementation truncates the spectral expansion and computes the constrained kernel and its derivatives offline, so velocity, vorticity, and uncertainty fields can be reconstructed from interior and outer-boundary data alone. A sympathetic reader would take this as a shift of emphasis: physical boundary conditions become part of the prior's geometry rather than data to be collected.

What carries the argument

The load-bearing object is the boundary-constrained Gaussian process (BCGP) kernel $G_0$, built from the Karhunen-Loève expansion of the base kernel localized on the profile $\Gamma$. The construction hinges on the lifted eigenfunctions $\tilde{\phi}_n$: they carry the boundary geometry into the whole domain through integrals of the base kernel against boundary eigenfunctions, and the infinite sum $\sum_n \lambda_n \tilde{\phi}_n(x)\tilde{\phi}_n(x')$ removes exactly the boundary variance from $G$. In the numerical method, the profile is parameterized as a curve, the integral eigenvalue problem is discretized by a Nyström quadrature into the generalized problem $GHE=E\Lambda$, and the spectral factor $S=H^{1/2}\tilde{E}\Lambda^{-1/2}$ lets every BCGP evaluation be written as $G_0(x,x')=G(x,x')-(G(x,X_\Gamma)S)(G(x',X_\Gamma)S)^T$. The same factor applies to derivatives $\partial_x^\alpha\otimes\partial_{x'}^\beta G_0$, which is why velocity estimates, vorticity estimates, and posterior covariances all inherit the boundary constraint with no boundary observations.

What would settle it

Fix a profile and a continuous base kernel with slowly decaying boundary spectrum, for example a Matérn kernel with smoothness $\nu=3/2$ on the NACA 0412 leading edge, run the paper's spectral algorithm with increasing $J$, and track both the spectral accuracy $\epsilon(I,J)$ and the velocity normal-fit indicator $\epsilon_n$; if reaching $\epsilon_n<10^{-4}$ demands $J$ close to the discretization size $I$, the continuous-enforcement claim and its computational saving fail for that kernel.

Watch

Extended reading notes

Core claim

Proposition 3.1 is the analytical core. For a zero-mean GP $Z$ with continuous kernel $G$ on a domain $\Omega$ and a compact set $\Gamma$ carrying a fully supported Borel measure $\nu$, let $(\lambda_n, \phi_n)$ be the eigenpairs of the restricted covariance operator $T_\Gamma[h](x)=\int_\Gamma G(x,y)h(y)\,d\nu(y)$, and let $\tilde{\phi}_n$ be their extension to $\Omega$ defined by $\tilde{\phi}_n(x)=\lambda_n^{-1}\int_\Gamma G(x,y)\phi_n(y)\,d\nu(y)$. The paper defines $Z_0(x)=Z(x)-\sum_n \sqrt{\lambda_n}\,\tilde{\phi}_n(x)\,\xi_n$, which is a centered GP with kernel $G_0(x,x')=G(x,x')-\sum_n \lambda_n\,\tilde{\phi}_n(x)\tilde{\phi}_n(x')$, and shows $Z_0=0$ on $\Gamma$ in the mean-square sense, with $G_0(\cdot,x')$ vanishing uniformly on $\Gamma$ for every $x'$. In the fluid setting $\psi=Z_0$ gives $u=\nabla^\perp\psi=\operatorname{curl} Z_0$, so $\nabla\cdot u=0$ by construction and $u\cdot n=0$ on $\Gamma$ because $\psi$ is constant there; the paper verifies the construction numerically on a cylinder at $\mathrm{Re}=3\cdot 10^3$ and a NACA 0412 leading edge at $\mathrm{Re}=2.7\cdot 10^3$, achieving a three-orders-of-magnitude improvement in the boundary normal-fit indicator compared with using the base kernel plus discrete boundary collocation.

Load-bearing premise

The continuous boundary constraint in practice is only as accurate as the truncated Karhunen-Loève spectrum: if the base kernel and profile geometry do not make the eigenvalues decay quickly, a finite number of modes cannot represent the vanishing condition, and the method degenerates into a discrete boundary approximation.

Editorial extensions

If this is right

  • The profile-boundary condition no longer consumes any of the $N$ observations: the GP prior enforces it continuously, so the same observation budget can be spent on interior and outer-boundary data.
  • Reconstructions of velocity and vorticity automatically satisfy $\nabla\cdot u=0$ and $u\cdot n=0$ on $\Gamma$ in the prior and posterior, avoiding spurious particle penetration in Lagrangian simulations.
  • Because the BCGP kernel and its derivatives are computed offline once, at cost $O(I^2(J+1))$ for the spectrum, online regression remains $O(N^3)$ in the number of observations, with no boundary collocation points inflating the Gram matrix.
  • Vorticity and stream-function estimates are obtained directly from the same kernel derivatives, without solving a pressure equation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same subtraction could be applied to derivative processes to impose no-slip (zero tangential velocity) or flux conditions, since the paper's derivative extension already provides the needed kernel derivatives; the paper itself lists no-slip as future work.
  • The practical advantage over discrete boundary collocation depends on how quickly the base kernel's restricted spectrum decays; for kernels or geometries with slow decay, the mode count $J$ will approach the discretization size $I$, and the offline computation will no longer be negligible. Testing with low-regularity Matérn kernels would map this boundary.
  • Because the constrained prior already prevents penetration, embedding it in a sequential or particle-filter data-assimilation loop could act as a soft wall constraint and reduce spurious particle crossings in Lagrangian estimates; the paper notes the Lagrangian connection but does not run such a filter.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes a Gaussian process (GP) framework for reconstructing two-dimensional incompressible flow fields around aerodynamic profiles. The main mathematical tool is a spectral procedure, attributed to Gauthier and Bay [12], that modifies an arbitrary GP with a continuous kernel so that the resulting process vanishes on a compact set Γ in the mean-square sense. The authors apply this construction to the stream function ψ, obtaining a velocity prior u = curl ψ that is divergence-free in the domain and satisfies a slip boundary condition on Γ. The boundary condition is imposed through the prior rather than through discrete boundary observations, which is the central advertised advantage. The paper also introduces a multi-scale RBF kernel whose energy-decay exponent γ=6 is derived from a scaling argument, and it provides numerical demonstrations for flow around a cylinder and around the leading edge of a NACA 0412 airfoil. The numerical method approximates the infinite spectral expansion by a Nyström discretization of the boundary, with a spectral accuracy bound δ controlling the number of retained modes. The paper includes comparisons against discrete boundary enforcement and reports reconstruction errors, uncertainty coverage, and computational timings, with code publicly available.

Significance. If the results hold, the paper offers a useful and fairly general recipe for building physics-informed GP priors with boundary constraints on arbitrary compact sets, and it demonstrates the recipe on nontrivial geometries. The method's ability to avoid spending part of the observation budget on boundary collocation points is attractive for scarce-data aerodynamic reconstructions, and the paper gives concrete evidence that this can lead to more accurate boundary-condition satisfaction than discrete constraint enforcement at equal Gram-matrix size. Strengths include the explicitly derived γ=6 scaling law, the clear comparison of two choices of spectral measure, the offline/online cost separation, and the availability of reproducible code with convergence studies of the spectral truncation. The main caveat is that the numerical kernel is built on a finite Nyström discretization, and the paper does not quantify the resulting between-node boundary error; this gap affects the strength of the 'continuous enforcement' and 'no boundary observations needed' claims.

major comments (2)
  1. [§3.3, Eqs. (20)–(27), Algorithm 1; Figures 8 and 16] The paper's central claim that the slip condition is enforced continuously, so that no boundary observations are needed, rests on the numerical kernel G0^(I,J) vanishing on Γ. What the Nyström discretization (20)–(22) actually enforces is that the kernel vanishes at the I node points X_Γ: in the full-rank limit J=I−1, G0^(I,J)(γ(s),x') = G(γ(s),x') − G(γ(s),X_Γ)G(X_Γ,X_Γ)^{-1}G(X_Γ,x'), which is the kriging residual and is generally nonzero for γ(s) between nodes. The spectral accuracy ϵ(I,J) in Eq. (23) measures the fraction of the discretized operator's trace captured by the retained modes; it does not bound this interpolation residual. The convergence studies in Figure 8 (I=400) and Figure 16 (I=300) vary only δ/J with I held fixed, so no evidence is given that the boundary error decays as I increases. The statement in §3.3 that the procedure 'does not represent a discrete enforcement of the boundary constraint at collocation points' is therefore inaccurate for finite I: the constraint is enforced exactly at the collocation points and only approximately between them. Please either prove a bound on the between-node residual in terms of I, J, and the kernel regularity, or add I-convergence studies with decreasing boundary error; in either case, the 'continuous enforcement' and 'no boundary observations needed' claims should be made conditional on this residual.
  2. [§5.3.2, Table 2] The comparison between the BCGP approach and discrete boundary enforcement is described as being at constant observation budget, but the two designs differ in more than the enforcement mechanism: the discrete approach uses N=344 velocity observations plus NΓ=142 boundary constraints, while the BCGP approach uses N=415 velocity observations. The three-orders-of-magnitude improvement in ϵn (1.611×10^{-2} versus 1.823×10^{-5}) may therefore reflect the extra 71 interior velocity measurements as much as the continuous constraint. This does not undermine the method, but the claim that the BCGP approach 'allows to satisfy the slip boundary condition at a better order of accuracy than that of the discrete method for the same observation budget' should be reworded, or supplemented by a comparison in which both methods use the same interior and boundary data and differ only in how the boundary condition is imposed. The paper should also clarify in Table 2 that the budget is fixed in terms of Gram-matrix size, not in terms of the number of physical observations.
minor comments (5)
  1. [§4.3] The derivation of γ=6 uses an asymptotic range '3<q≤M'; with the numerical choice M=3 this range is empty. The paper should state that the scaling law is justified in the limit M→∞ and that M=3 is a finite-scale approximation, and it should discuss how many scales are needed for the O(∥h∥²) behavior to be visible.
  2. [Corollary 3.4] The text says H^{d−1}(Γ) is the '(p−1)-dimensional measure of the manifold,' which appears to be a typo: it should be the (d−1)-dimensional measure. In the case p=1 the statement about arc length is correct, but the indexing should be cleaned up.
  3. [§3.3, Eq. (23) and Figure 16] The horizontal axis of Figure 16 is described as the negative logarithm of ϵ, while the experiment varies the spectral threshold δ; please clarify that the axis is −log₁₀(δ) and that ϵ(I,J)≤δ by construction.
  4. [§3.4, Table 1 and §5.4] The complexity comparison says the spectral factor cost is negligible because 'J≪N', but in the NACA experiment J=112–200 modes are used with only N=184 training observations, so J is comparable to N. The asymptotic claim should be qualified with the actual working regime.
  5. [§4.2, Corollary 4.1] Corollary 4.1 asserts uniform convergence of the differentiated series (36)–(37) and Remark 4.2 gives a sufficient condition, but no proof is included. Since the numerical method relies on differentiated BCGP kernels, a short justification or a precise reference would help the reader verify that the interchange of differentiation and summation is legitimate for the anisotropic Gaussian and multi-scale RBF kernels used here.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the boundary-constrained GP is an explicit conditional-projection construction, the physics parameters are set from stated scaling laws and cross-validated data, and the cited spectral result is external to the authors.

full rationale

The central construction in Proposition 3.1 is not a disguised prediction of its own inputs: Z0(x)=Z(x)-Σ sqrt(λn) φ̃n(x) ξn is defined as the residual after removing the projection of Z onto the span of its boundary values on Γ, and the paper proves (and sketches) that this residual has zero covariance on Γ. The zero boundary condition is an imposed constraint, not an empirical quantity being inferred; the kernel G0 is then the covariance of that residual by construction. The result is attributed to [12], an external source, and the proof sketch is self-contained enough to show the mechanism. Similarly, the energy-decay parameter γ=6 is obtained algebraically from the stated reference structure ℓm=ℓ0/2^{3m}, σm=σ0/2^{γm}, and the known 2D spectrum E(k)∝k^{-3}, so it is not fitted to the reconstruction targets; σ0 and ℓ0 are tuned by UQ-based cross-validation on training folds and evaluated on a separate test set. The only self-citation involving an author ([14], which includes Éric Savin) is a pointer in the related-work and future-work discussion, not a load-bearing premise. The numerical implementation replaces the integral operator on Γ by an I-point Nyström quadrature (Eqs. 20-22, 25-27), so for finite I the continuous zero condition holds exactly at the quadrature nodes in the full-rank limit and only approximately between them; this is a numerical accuracy and convergence question about the I-discretization, not a circularity in the derivation. No equation in the paper reduces to a fitted quantity or to a self-citation chain, so no circular step is present.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central construction rests on standard spectral theory (Mercer/KLE), a C^{2,2} regularity assumption for derivative kernels, a chosen multi-scale energy-decay model, and a numerical Nystrom eigenproblem. The only numbers fitted to data are the global kernel hyperparameters sigma0 and ell0, plus algorithmic choices (M, I, J, delta, eta, alpha). No new physical entities are introduced; the constrained kernel is a mathematical construct. The energy exponent gamma=6 is derived rather than fitted, but the scale ratios feeding that derivation are chosen by the authors.

free parameters (8)
  • sigma_0 (base kernel amplitude) = RBF 0.20; DF-RBF 0.01; M-RBF 0.05; PI-RBF 0.10
    Global variance scale, tuned by UQ-based cross-validation in Sec. 5.4.1; changes reconstruction amplitude and uncertainty.
  • ell_0 (coarse-scale correlation length) = RBF/DF-RBF 0.1; M-RBF/PI-RBF 1.0
    Correlation length at the coarsest scale, tuned by cross-validation; controls smoothness of the reconstructed flow.
  • eta (nugget) = 1e-6 cylinder; 1e-8 or 1e-10 NACA
    Regularization added to the Gram matrices for invertibility; chosen by hand at the minimal stable order.
  • M (number of scales) = 3 for multiple-scale and PI kernels
    Hand-chosen number of additive RBF components for the energy-decay construction in Sec. 4.3.
  • alpha (anisotropy) in Sec 5.3 = 1.5
    Fixed by hand for the cylinder visualizations; in the multi-scale kernel alpha_m is set by the formula max(1, 2^(3-m)).
  • I, J (boundary discretization and spectral truncation) = I=400, J=18 (cylinder); I=300, J=23-200 (NACA for delta=1e-9 to 1e-14)
    Algorithmic parameters controlling the accuracy of the approximate boundary constraint; chosen to reach spectral accuracy delta=1e-12.
  • delta (spectral accuracy threshold) = 1e-12
    Threshold used in Eq. (24) to select the number of modes J; directly controls the numerical error of the slip condition.
  • gamma (energy decay exponent) = 6
    Derived analytically in Sec. 4.3 from a k^-3 2D energy spectrum with imposed scale ratios; not fitted to data, but dependent on the chosen construction.
assumptions (6)
  • standard math Uniform convergence of Mercer and Karhunen-Loeve expansions for continuous kernels on compact sets.
    Used in Proposition 3.1 and Remark 3.2 to justify the infinite sums defining G0 and their uniform convergence; standard spectral theorem for compact self-adjoint operators.
  • standard math Convergence of the Nystrom quadrature approximation of the integral eigenvalue problem (Eqs. 19-22).
    The continuous boundary constraint is replaced by the discrete generalized eigenvalue problem GHE=E Lambda; the paper assumes the resulting spectral factor S in Eq. (28) approximates the operator spectrum well enough for the chosen I and delta.
  • domain assumption The base kernel G is at least C^{2,2}(Omega x Omega).
    Needed for the GP derivatives defining velocity and vorticity (Remark 4.2); satisfied by the anisotropic RBF kernels used here.
  • domain assumption Two-dimensional incompressible flow is represented by a stream function with u=curl psi, and the slip condition is equivalent to psi being constant on the profile boundary.
    Standard 2D fluid mechanics; the paper imposes the homogeneous Dirichlet condition on psi to enforce u dot n = 0 on Gamma (Sec. 4.1-4.2).
  • ad hoc to paper Small-scale energy decay follows the constructed multi-scale RBF model with ell_m=ell_0/2^(3m), alpha_m=max(1,2^(3-m)), sigma_m=sigma_0/2^(gamma m), and gamma=6.
    This spectral-shape assumption in Sec. 4.3 is justified by a k^-3 power-law argument, but the exact scale ratios are author choices; a different energy spectrum would require a different construction.
  • domain assumption The ground-truth flows used for evaluation satisfy the same homogeneous slip condition that the prior enforces on the profile boundary.
    The cylinder setup declares slip on the profile and the AirfRANS data are adapted to slip in Sec. 5.4; if the target flow were no-slip, the prior would be misspecified near the boundary.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-informed, boundary-constrained Gaussian process regression for the reconstruction of fluid flow fields." pith.science (2026). https://pith.science/paper/U2B2NMLE

@misc{pith2026250717582,
  author       = {Pith},
  title        = {Pith review of: Physics-informed, boundary-constrained Gaussian process regression for the reconstruction of fluid flow fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U2B2NMLE}},
  note         = {Machine review of arXiv:2507.17582}
}
read the original abstract

Gaussian process regression techniques have been used in fluid mechanics for the reconstruction of flow fields from a reduction-of-dimension perspective. A main ingredient in this setting is the construction of adapted covariance functions, or kernels, to obtain such estimates. In this paper, we present a general method for constraining a prescribed Gaussian process on an arbitrary compact set. The kernel of the pre-defined process must be at least continuous and may include other information about the studied phenomenon. This general boundary-constraining framework can be implemented with high flexibility for a broad range of engineering applications. From this, we derive physics-informed kernels for simulating two-dimensional velocity fields of an incompressible (divergence-free) flow around aerodynamic profiles. These kernels allow to define Gaussian process priors satisfying the incompressibility condition and the prescribed boundary conditions along the profile in a continuous manner. We describe an adapted numerical method for the boundary-constraining procedure parameterized by a measure on the compact set. The relevance of the methodology and performances are illustrated by numerical simulations of flows around a cylinder and a NACA 0412 airfoil profile, for which no observation at the boundary is needed at all.

Figures

Figures reproduced from arXiv: 2507.17582 by the authors.

Figure 1
Figure 1. Summary of the methodology to enforce physical constraints on GP priors for velocity and vorticity fields. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. GPR setting for approximating the velocity field of an incompressible fluid flow around a profile in a [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Spatial GPR interpolations for velocity (arrows) and vorticity (colormap) fields around different profiles [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Ground truth velocity field of an incompressible fluid flow around a cylinder profile in a wind-tunnel setting. [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Initial reconstruction by physics-informed GPR of the velocity field of an incompressible fluid flow ( [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Total standard deviation of the velocity field reconstructed by GPR with or without BCGP kernels. Design [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Reconstruction by physics-informed GPR of the velocity field of an incompressible fluid flow ( [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Evaluation of indicators of physical constraints around profile boundary with respect to spectral expansion [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 10
Figure 10. Figure 10: Sensitivity analysis of the normal fit indicator [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Stationary ground truth velocity field u † of an incompressible fluid flow around the NACA 0412 airfoil. The colormap represents the norm of the velocity field in m/s. Our objective is to compare different choices of GP priors used in the reconstruction formula (43) b…
Figure 12
Figure 12. Figure 12: Estimation of hyperparameters σ0 and ℓ0 through cross-validation using UQ coverage indicators. The base 10 logarithm of the mean loss function from Eq. (63) obtained over the grid {0.005, 0.01., 0.05, 0.1, 0.15, 0.2} × {0.01, 0.1, 0.2, 0.4, 0.6, 0.8, 1.0} is depicted …
Figure 13
Figure 13. Figure 13: Reconstruction of the velocity field of an incompressible flow at the leading edge of a NACA 0412 airfoil [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 14
Figure 14. Figure 14: UQ of the reconstructed velocity field of an incompressible flow at the leading edge of a NACA 0412 airfoil [PITH_FULL_IMAGE:figures/full_fig_p025_14.png]
Figure 15
Figure 15. Figure 15: Estimates and confidence intervals of the horizontal (from (a) to (d)) and vertical (from (e) to (h)) [PITH_FULL_IMAGE:figures/full_fig_p025_15.png]
Figure 16
Figure 16. Figure 16: Evaluation of indicators of physical constraints around profile boundary with respect to spectral expansion [PITH_FULL_IMAGE:figures/full_fig_p026_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 44 canonical work pages

  1. [12]

    Spectral approach for kernel-based interpolation.Annales de la Facult´ e des Sciences de Toulouse: Math´ ematiques, 21(3):439–479, 2012

    Bertrand Gauthier and Xavier Bay. Spectral approach for kernel-based interpolation.Annales de la Facult´ e des Sciences de Toulouse: Math´ ematiques, 21(3):439–479, 2012

    work page 2012

  2. [1]

    Numerical study of slightly viscous flow.Journal of Fluid Mechanics, 57(4):785–796, 1973

    Alexandre Joel Chorin. Numerical study of slightly viscous flow.Journal of Fluid Mechanics, 57(4):785–796, 1973

    work page 1973

  3. [2]

    A stochastic Lagrangian representation of the three-dimensional incompressible Navier-Stokes equations.Communications on Pure and Applied Mathematics, 61(3):330–345, 2008

    Peter Constantin and Gautam Iyer. A stochastic Lagrangian representation of the three-dimensional incompressible Navier-Stokes equations.Communications on Pure and Applied Mathematics, 61(3):330–345, 2008

    work page 2008

  4. [3]

    Forward–backward stochastic differential systems associated to Navier–Stokes equations in the whole space.Stochastic Processes and their Applications, 125(7):2516–2561, 2015

    Freddy Delbaen, Jinniao Qiu, and Shanjian Tang. Forward–backward stochastic differential systems associated to Navier–Stokes equations in the whole space.Stochastic Processes and their Applications, 125(7):2516–2561, 2015

    work page 2015

  5. [4]

    An introduction to 3D stochastic fluid dynamics

    Franco Flandoli. An introduction to 3D stochastic fluid dynamics. In Giuseppe Da Prato and Michael R¨ uckner, editors,SPDE in Hydrodynamic: Recent Progress and Prospects: Lectures given at the C.I.M.E. Summer School held in Cetraro, Italy August 29–September 3, 2005, pages 51–150. Springer, Berlin, 2008

    work page 2005

  6. [5]

    Butterworth-Heinemann, Oxford, second edition, 2007

    Charles Hirsch.Numerical Computation of Internal and External Flows. Butterworth-Heinemann, Oxford, second edition, 2007

    work page 2007

  7. [6]

    Pope.Turbulent Flows

    Stephen B. Pope.Turbulent Flows. Cambridge University Press, Cambridge, first edition, 2000

    work page 2000

  8. [7]

    Cambridge University Press, Cambridge, second edition, 2024

    Stavros Tavoularis and Jovan Nedi´ c.Measurement in Fluid Mechanics. Cambridge University Press, Cambridge, second edition, 2024

    work page 2024

Show all 56 references
  1. [8]

    H. G. Maas, A. Gr¨ un, and D. A. Papantoniou. Particle tracking velocimetry in three-dimensional flows. Part I. Photogrammetric determination of particle coordinates.Experiments in Fluids, 15:133–146, 1993

    1993

  2. [9]

    N. A. Malik, Th. Dracos, and D. A. Papantoniou. Particle tracking velocimetry in three-dimensional flows. Part II. Particle tracking.Experiments in Fluids, 15:279–294, 1993

    1993

  3. [10]

    State observer data assimilation for RANS with time-averaged 3D-PIV data.Computers & Fluids, 218:104827, 2021

    Edoardo Saredi, Nikhilesh Tumuluru Ramesh, Andrea Sciacchitano, and Fulvio Scarano. State observer data assimilation for RANS with time-averaged 3D-PIV data.Computers & Fluids, 218:104827, 2021

    2021

  4. [11]

    McKeon, Denis Sipp, and Peter J

    Sean Symon, Nicolas Dovetta, Beverley J. McKeon, Denis Sipp, and Peter J. Schmid. Data assimilation of mean velocity from 2D PIV measurements of flow over an idealized airfoil.Experiments in Fluids, 58:61, 2017

    2017

  5. [13]

    Gaussian process hydrodynamics.Applied Mathematics and Mechanics, 44(7):1175–1198, 2023

    Houman Owhadi. Gaussian process hydrodynamics.Applied Mathematics and Mechanics, 44(7):1175–1198, 2023

    2023

  6. [14]

    Learning “best” kernels from data in Gaussian process regression

    Jean-Luc Akian, Luc Bonnet, Houman Owhadi, and ´Eric Savin. Learning “best” kernels from data in Gaussian process regression. With application to aerodynamics.Journal of Computational Physics, 470:111595, 2022

    2022

  7. [15]

    Kernel Flows: From learning kernels from data into the abyss.Journal of Computational Physics, 389:22–47, 2019

    Houman Owhadi and Gene Ryan Yoo. Kernel Flows: From learning kernels from data into the abyss.Journal of Computational Physics, 389:22–47, 2019

    2019

  8. [16]

    Learning dynamical systems from data: a simple cross-validation perspective, part I: parametric kernel flows.Physica D: Nonlinear Phenomena, 421:132817, 2021

    Boumediene Hamzi and Houman Owhadi. Learning dynamical systems from data: a simple cross-validation perspective, part I: parametric kernel flows.Physica D: Nonlinear Phenomena, 421:132817, 2021

    2021

  9. [17]

    Covariance Models for Divergence-Free and Curl-Free Random Vector Fields.Stochastic Models, 28(3):433–451, 2012

    Michael Scheuerer and Martin Schlather. Covariance Models for Divergence-Free and Curl-Free Random Vector Fields.Stochastic Models, 28(3):433–451, 2012

    2012

  10. [18]

    Boundary constrained Gaussian processes for robust physics-informed machine learning of linear partial differential equations.Journal of Machine Learning Research, 25:1–61, 2024

    David Dalton, Alan Lazarus, Hao Gao, and Dirk Husmeier. Boundary constrained Gaussian processes for robust physics-informed machine learning of linear partial differential equations.Journal of Machine Learning Research, 25:1–61, 2024

    2024

  11. [19]

    Gaussian process regression constrained by boundary value problems.Computer Methods in Applied Mechanics and Engineering, 388:114117, 2022

    Mamikon Gulian, Ari Frankel, and Laura Swiler. Gaussian process regression constrained by boundary value problems.Computer Methods in Applied Mechanics and Engineering, 388:114117, 2022

    2022

  12. [20]

    Know your boundaries: Constraining gaussian processes by variational harmonic features

    Arno Solin and Manon Kok. Know your boundaries: Constraining gaussian processes by variational harmonic features. In Kamalika Chaudhuri and Masashi Sugiyama, editors,Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, volume 89 ...

    2019

  13. [21]

    Brunton, Bernd R

    Steven L. Brunton, Bernd R. Noack, and Petros Koumoutsakos. Machine learning for fluid mechanics.Annual Review of Fluid Mechanics, 52:477–508, 2020

    2020

  14. [22]

    Data reconstruction for complex flows using AI: Recent progress, obstacles, and perspectives

    Michele Buzzicotti. Data reconstruction for complex flows using AI: Recent progress, obstacles, and perspectives. Europhysics Letters, 142(2):23001, 2023

    2023

  15. [23]

    Turbulence modeling in the age of data.Annual Review of Fluid Mechanics, 51:357–377, 2019

    Karthik Duraisamy, Gianluca Iaccarino, and Heng Xiao. Turbulence modeling in the age of data.Annual Review of Fluid Mechanics, 51:357–377, 2019

    2019

  16. [24]

    Bharath, and Chris D

    Mario Lino, Stathi Fotiadis, Anil A. Bharath, and Chris D. Cantwell. Current and emerging deep-learning methods for the simulation of fluid dynamics.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 479(2275):20230058, 2023

    2023

  17. [25]

    Maziar Raissi, Paris Perdikaris, and George Ern Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics, 378:686–707, 2019

    2019

  18. [26]

    The graph neural network model.IEEE Transactions on Neural Networks, 20(1):61–80, 2009

    Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model.IEEE Transactions on Neural Networks, 20(1):61–80, 2009

    2009

  19. [27]

    Fourier neural operator for parametric partial differential equations.arXiv:2010.08895, 2021

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv:2010.08895, 2021

    2010 arXiv

  20. [28]

    Karniadakis

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George E. Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3(3):218–229, 2021

    2021

  21. [29]

    Hou, and Max Tegmark

    Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljaˇ ci´ c, Thomas Y. Hou, and Max Tegmark. KAN: Kolmogorov-Arnold Networks.arXiv:2404.19756, 2025

    2025 arXiv

  22. [30]

    Poseidon: Efficient foundation models for PDEs.arXiv:2405.19101, 2024

    Maximilian Herde, Bogdan Raoni´ c, Tobias Rohner, Roger K¨ appeli, Roberto Molinaro, Emmanuel de B´ ezenac, and Siddhartha Mishra. Poseidon: Efficient foundation models for PDEs.arXiv:2405.19101, 2024

    2024 arXiv

  23. [31]

    Karniadakis

    Juan Diego Toscano, Theo K¨ aufer, Zhibo Wang, Martin Maxey, Christian Cierpka, and George E. Karniadakis. Inferring turbulent velocity and temperature fields and their statistics from Lagrangian velocity measurements using physics-informed Kolmogorov-Arnold networks.arXiv:240...

    2024 arXiv

  24. [32]

    Prediction of turbulent channel flow using Fourier neural operator-based machine-learning strategy.Physical Review Fluids, 9(8):084604, 2024

    Yunpeng Wang, Zhijie Li, Zelong Yuan, Wenhui Peng, Tianyuan Liu, and Jianchun Wang. Prediction of turbulent channel flow using Fourier neural operator-based machine-learning strategy.Physical Review Fluids, 9(8):084604, 2024

    2024

  25. [33]

    F. Bach. Information theory with kernel methods.IEEE Transactions on Information Theory, 69(2):752–775, 2023

    2023

  26. [34]

    Gaussian processes and kernel methods: A review on connections and equivalences.arXiv:1807.02582, 2018

    Motonobu Kanagawa, Philipp Hennig, Dino Sejdinovic, and Bharath K Sriperumbudur. Gaussian processes and kernel methods: A review on connections and equivalences.arXiv:1807.02582, 2018

    2018 arXiv

  27. [35]

    Smola.Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond

    Bernhard Sch¨ olkopf and Alexander J. Smola.Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond. The MIT Press, Cambridge MA, 2001

    2001

  28. [36]

    Carl Edward Rasmussen and Christopher K. I. Williams.Gaussian Processes for Machine Learning. The MIT Press, Cambridge MA, 2005

    2005

  29. [37]

    Solving and learning nonlinear PDEs with Gaussian processes.Journal of Computational Physics, 447:110668, 2021

    Yifan Chen, Bamdad Hosseini, Houman Owhadi, and Andrew M Stuart. Solving and learning nonlinear PDEs with Gaussian processes.Journal of Computational Physics, 447:110668, 2021

    2021

  30. [38]

    Characterization of the second order random fields subject to linear distributional PDE constraints.Bernoulli, 29(4):3396–3422, 2023

    Iain Henderson, Pascal Noble, and Olivier Roustant. Characterization of the second order random fields subject to linear distributional PDE constraints.Bernoulli, 29(4):3396–3422, 2023

    2023

  31. [39]

    On degeneracy and invariances of random fields paths with applications in Gaussian process modelling.Journal of Statistical Planning and Inference, 170:117–128, 2016

    David Ginsbourger, Olivier Roustant, and Nicolas Durrande. On degeneracy and invariances of random fields paths with applications in Gaussian process modelling.Journal of Statistical Planning and Inference, 170:117–128, 2016

    2016

  32. [40]

    Sobolev regularity of Gaussian random fields.Journal of Functional Analysis, 286(3):110241, 2024

    Iain Henderson. Sobolev regularity of Gaussian random fields.Journal of Functional Analysis, 286(3):110241, 2024

    2024

  33. [41]

    Cambridge Monographs on Applied and Computational Mathematics

    Houman Owhadi and Clint Scovel.Operator-Adapted Wavelets, Fast Solvers, and Numerical Homogenization: From a Game Theoretic Approach to Numerical Approximation and Algorithm Design. Cambridge Monographs on Applied and Computational Mathematics. Cambridge University Press, Camb...

    2019

  34. [42]

    V. I. Paulsen and M. Raghupathi.An Introduction to the Theory of Reproducing Kernel Hilbert Spaces. Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2016

    2016

  35. [43]

    Kernels for multi-task learning

    Charles Micchelli and Massimiliano Pontil. Kernels for multi-task learning. In L. Saul, Y. Weiss, and L. Bottou, editors,Advances in Neural Information Processing Systems, volume 17. MIT Press, Cambridge MA, 2004

    2004

  36. [44]

    Springer Science & Business Media, 2008

    Ingo Steinwart and Andreas Christmann.Support vector machines. Springer Science & Business Media, 2008

    2008

  37. [45]

    Springer, Berlin, 2007

    Vladimir Igorevich Bogachev.Measure Theory, volume 1. Springer, Berlin, 2007

    2007

  38. [46]

    CRC Press, 1992

    Lawrence Craig Evans and Ronald F Gariepy.Measure Theory and Fine Properties of Functions. CRC Press, 1992

    1992

  39. [47]

    Cambridge University Press, 1997

    Kendall E Atkinson.The Numerical Solution of Integral Equations of the Second Kind, volume 4. Cambridge University Press, 1997

    1997

  40. [48]

    Ladson, Jr

    Charles L. Ladson, Jr. Brooks, Cuyler W., Acquilla S. Hill, and Darrell W. Sproles. Computer program to obtain ordinates for NACA airfoils. Technical Memorandum 4741, NASA, 1996

    1996

  41. [49]

    Uriel Frisch.Turbulence: The Legacy of A. N. Kolmogorov. Cambridge University Press, 1995

    1995

  42. [50]

    Guido Boffetta and Robert E. Ecke. Two-dimensional turbulence.Annual Review of Fluid Mechanics, 44(1):427–451, 2012

    2012

  43. [51]

    A Dynamics-Informed Gaussian Process Framework for 2D Stochastic Navier-Stokes via Quasi-Gaussianity.arXiv:2511.21281, 2025

    Boumediene Hamzi and Houman Owhadi. A Dynamics-Informed Gaussian Process Framework for 2D Stochastic Navier-Stokes via Quasi-Gaussianity.arXiv:2511.21281, 2025

    2025

  44. [52]

    Smith, Mateusz Paprocki, Ondˇ rejˇCert ´ ık, Sergey B

    Aaron Meurer, Christopher P. Smith, Mateusz Paprocki, Ondˇ rejˇCert ´ ık, Sergey B. Kirpichev, Matthew Rocklin, AMiT Kumar, Sergiu Ivanov, Jason K. Moore, Sartaj Singh, Thilina Rathnayake, Sean Vig, Brian E. Granger, Richard P. Muller, Francesco Bonazzi, Harsh Gupta, Shivam Va...

    2017

  45. [53]

    H. G. Weller, G. Tabor, H. Jasak, and C. Fureby. A tensorial approach to computational continuum mechanics using object-oriented techniques.Computer in Physics, 12(6):620–631, 1998. See https://www.openfoam.com

    1998

  46. [54]

    AirfRANS: High fidelity computa- tional fluid dynamics dataset for approximating Reynolds-Averaged Navier-Stokes solutions

    Florent Bonnet, Ahmed Jocelyn Mazari, Paola Cinnella, and Patrick Gallinari. AirfRANS: High fidelity computa- tional fluid dynamics dataset for approximating Reynolds-Averaged Navier-Stokes solutions. In36th Conference on Neural Information Processing Systems Datasets and Benc...

    2023

  47. [55]

    Quasi-Gaussianity of the 2D stochastic Navier-Stokes equations

    James Coe, Martin Hairer, and Leonardo Tolomeo. Quasi-Gaussianity of the 2D stochastic Navier-Stokes equations. arXiv:2510.13460, 2025

    2025

  48. [56]

    Sparse Cholesky factorization for solving nonlinear PDEs via Gaussian processes.Mathematics of Computation, 94(353):1235–1280, 2025

    Yifan Chen, Houman Owhadi, and Florian Sch¨ afer. Sparse Cholesky factorization for solving nonlinear PDEs via Gaussian processes.Mathematics of Computation, 94(353):1235–1280, 2025. 29

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.