An Artificial-Compressibility Physics-Informed Neural Network for the Unsteady Incompressible Navier--Stokes Equations

Aytekin \c{C}ibik

REVIEW 3 major objections 4 minor 25 references

An Artificial-Compressibility Physics-Informed Neural Network for the Unsteady Incompressible Navier--Stokes Equations

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read This paper claims that replacing the divergence-free constraint with an artificial-compressibility relaxation makes unsteady incompressible Navier–Stokes equations trainable for PINNs, and that the method's reach is bounded by…

desk verdict A genuinely useful and honest map of the AC-PINN trade-off, with clean Taylor–Green scaling; the cylinder validation is circular but disclosed, and the paper deserves serious review. read the letter →

arxiv 2608.04191 v1 pith:JKCGPZHB submitted 2026-08-04 physics.flu-dyn cs.NAmath.NA

Aytekin Çibik This is my paper · ORCID

classification physics.flu-dyncs.NAmath.NA MSC 35Q3065M6068T0776D05

keywords artificialcompressibilityphysics-informedneuralnetworksincompressibleNavier-StokesequationsTaylor-Greenvortexlid-drivencavitycylinderwakesheddingparameteridentification

topics Navier-Stokes Regularity

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

Artificial compressibility, the replacement of the divergence-free condition by $\varepsilon\,\partial_t p + \nabla\cdot u = 0$, converts the stiff differential-algebraic constraint of incompressible flow into a purely local evolution residual that a physics-informed neural network can minimize directly. The paper establishes this on three benchmarks of increasing difficulty. On the Taylor–Green vortex, the residual divergence scales as $\varepsilon|\partial_t p|$, and fixed-$\varepsilon$ training is stable across five decades of $\varepsilon$ with sub-0.5% velocity error, so no elaborate training aids are required on smooth problems. On the lid-driven cavity, a long enough time window recovers reference centerline profiles to 7.7%. The boundary of the method is the cylinder wake: the forward solve collapses to the steady symmetric branch, while assimilating a few hundred sparse velocity sensors recovers the vortex street to 7% and its frequency to within 3% of the reference, a bound set by the reference solver's own fidelity. The paper's larger point is that this relaxation maps where a lightweight toolbox suffices and where data assimilation is unavoidable.

What carries the argument

The load-bearing object is the relaxed continuity equation $\varepsilon\,\partial_t p + \partial_x u + \partial_y v = 0$, with $\varepsilon$ playing the role of an inverse squared artificial sound speed. It removes the differential-algebraic constraint and makes every governing equation carry a time derivative, so the residual system becomes purely local; the divergence error then scales as $\varepsilon|\partial_t p|$. This single parameter carries the argument by trading modelling error against optimisation stiffness, and the paper maps that trade-off on a closed-form benchmark and then applies it to flows where the steady branch competes with an unsteady one.

What would settle it

Train the plain forward AC-PINN on the Re=100 cylinder with no data-misfit term and a strong, sustained symmetry-breaking inflow, then measure the wake probe's transverse velocity RMS and spectral peak; the paper predicts collapse to the steady branch with RMS around 3e-3 and no Strouhal peak near 0.17. If the network sheds with Strouhal approximately 0.17 without data, the claim that data-free forward solves cannot reach the instability is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the artificial-compressibility relaxation with a single scalar $\varepsilon$ turns the global divergence constraint into a local residual, at a modelling-error price controlled by $\varepsilon$ and quantified by the scaling $\partial_x u + \partial_y v = -\varepsilon\,\partial_t p$. The paper shows that both divergence and velocity error decrease monotonically and saturate as $\varepsilon\to 0$, with no stiffness-induced degradation down to $\varepsilon=10^{-4}$ on a smooth benchmark. On the $Re=100$ cylinder wake the data-free forward AC-PINN relaxes to the steady symmetric wake, a valid low-residual solution; this failure is not a training artifact but a structural property, so sparse-sensor data assimilation is required to recover shedding. The paper also claims that assimilation fidelity is bounded by the reference fidelity, and that viscosity is identifiable only when sensors are placed in the boundary layer where the viscous term is significant.

Load-bearing premise

The load-bearing premise is that minimizing the mean-squared PDE residual at sampled collocation points is a faithful proxy for solving the equations, so the optimizer finds the physically correct solution rather than a low-residual alternative such as the steady symmetric cylinder wake.

Editorial extensions

If this is right

On smooth problems with known boundary and initial data, the AC-PINN needs only variable scaling and a plain fixed-$\varepsilon$ residual to reach sub-0.5% velocity error; adaptive weighting and $\varepsilon$-continuation improve accuracy but are not necessary.
For steady-limit flows in the lid-driven cavity, the physical time horizon, not collocation density or near-wall sampling, determines accuracy; wall-clustered sampling actively degrades the result.
A data-free forward PINN cannot be expected to produce self-sustained instabilities such as vortex shedding, because the steady branch is a low-residual solution; sparse velocity observations are sufficient to recover the unsteady branch.
Any validation of a data-driven wake reconstruction is only as good as the reference solver's boundary-layer resolution; a reference biased high in Strouhal passes that bias on to the learned solution.
Inverse parameter recovery depends on sensor placement, not sensor count: viscosity estimation fails from wake sensors alone and becomes identifiable once boundary-layer sensors are included.

Reading between the lines

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

The same relaxation may transfer to other DAE-constrained PDEs where a global constraint couples all collocation points, giving a local residual at the cost of a tunable modelling error; this is testable by comparing convergence on a problem with a known constraint.
The collapse to the steady branch suggests a testable fix for forward instability: adding a term that penalizes low-residual steady solutions, or explicitly breaking symmetry in the loss landscape, may let a data-free solve reach shedding without sensors.
The sensor-placement result implies an active-learning strategy for inverse problems: place observations where the term to be inferred is largest, such as the boundary layer for viscosity, and use a prior or denser near-wall data to close residual bias.
The paper's negative result on boundary-layer-adaptive sampling for the cavity suggests that local residual balancing via gradient-norm weighting can be destabilized by adding near-wall terms, pointing toward curriculum or hard-constraint treatments of thin layers.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

You should know about this one: it is a methods paper that maps the artificial-compressibility (AC) relaxation for PINNs on the unsteady incompressible Navier–Stokes equations. The new empirical content is a quantitative epsilon-scaling study on Taylor–Green (five decades, multiple seeds), a documented forward-solve collapse to the steady symmetric cylinder wake, and a sensor-placement identifiability study for viscosity inference. Those are genuinely useful and I do not think they appear in the prior literature.

What it does well: the Taylor–Green tests are clean. The residual divergence scales as epsilon*|partial_t p|, which is an algebraic consequence of equation (4), but they verify it and show that fixed-epsilon training is stable down to 1e-4 with sub-0.5% velocity error. Seed-averaged numbers are given. The cavity result is honest: 7.7% centerline error, with the physical time horizon identified as the controlling factor, plus a reported negative result that wall clustering makes it worse. The cylinder forward collapse is also honestly reported and consistent with prior findings.

The soft spots are real but disclosed. The cylinder data-assimilation validation is circular: the sensors and the ground truth come from the same FEM reference, so the 7% velocity error and 3% frequency match measure how well the network fits that reference, not how close to the true flow. The authors say exactly this in the abstract and Section 7.2, so it is not a hidden flaw. Still, the reference Strouhal is 0.176, above the accepted 0.164–0.172 band, so the 3% claim could correspond to a ~7% error against the true shedding frequency if the literature band is right. A converged-mesh Strouhal check inside the band would strengthen the claim. Also, there is no code or data, and no same-budget comparison against a hard-constrained PINN baseline, so the claimed advantage of AC over divergence-penalty or stream-function formulations is not established. That latter point is minor because the paper mostly claims to map the trade-off, not to beat baselines.

Overall: a sincere and useful methods paper. The central scaling story on Taylor–Green holds up, and the honest documentation of failures is valuable. The inverse sensor-placement study is a nice practical lesson. I would send it to peer review and ask for code/data and ideally an independent Strouhal check before accepting. I would bring it to our reading group when we discuss PINN training pathologies.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes an artificial-compressibility physics-informed neural network (AC-PINN) for unsteady incompressible Navier–Stokes equations. The divergence-free constraint is replaced by the relaxed continuity equation ε ∂_t p + ∂_x u + ∂_y v = 0 (Eq. 4), so that every governing equation carries a time derivative and all PINN residuals become local. The paper reports three benchmark studies: Taylor–Green vortex, where a seeded study over five decades of ε shows sub-0.5% velocity error and divergence scaling; lid-driven cavity at Re = 100, where a smoothed-lid unsteady solve reaches 7.7% relative-L2 agreement with Ghia et al. centerline profiles after a sufficiently long time window; and cylinder wake at Re = 100, where the plain forward solve collapses to the steady symmetric branch, data assimilation from a finite-element reference recovers the wake to 7% velocity error and Strouhal number to within 3% of that reference, and viscosity inference requires boundary-layer sensors. The paper is explicitly honest about its negative results and about the reference-fidelity limitations of the cylinder validation.

Significance. The paper's main value is its systematic and unusually transparent characterization of a simple formulation modification, including clear negative results. The Taylor–Green experiments are strengthened by three-seed statistics and by the observation that plain fixed-ε training is stable across five decades, which is a useful practical fact for practitioners. The explicit identification of the forward cylinder failure and the sensor-placement dependence of viscosity identification are genuinely informative. The authors also state plainly that the cylinder validation is circular and that the FEM reference itself sits above the literature Strouhal band. However, the central methodological claim—that artificial compressibility makes incompressible-flow PINN training easier—is not supported by any same-budget comparison against a standard PINN or a hard-constrained alternative, and the cavity benchmark is modified by a smoothed lid, so the reported agreements with classical references are not as clean as they appear. These gaps need to be addressed before the paper's conclusions can be fully accepted.

major comments (3)

[Section 6] The cavity benchmark is altered by the smoothed lid profile u_lid(x) = tanh(50x) tanh(50(1−x)), so the problem actually solved is not the sharp-lid Ghia benchmark used for comparison. The reported 7.7% centerline error therefore conflates the effect of the smoothing (a free parameter that is not studied) with the method's accuracy. The authors should either solve the sharp-lid problem, possibly with an appropriate corner treatment, or reformulate the claim as agreement with a smoothed-lid reference and provide a sensitivity study of the smoothing width.
[Sections 7.1–7.2] The cylinder data-assimilation validation is circular: the sparse sensors are drawn from the same finite-element reference against which the 7% velocity error and 3% Strouhal error are computed, and that reference has St = 0.176, above the accepted 0.164–0.172 literature band. Although the text repeatedly acknowledges this, the abstract's 'within 3%' statement is an agreement with a biased reference, not with the true flow. The unsteady-recovery claim would be substantially strengthened by comparing against an independent reference solution or by explicitly reporting the error relative to the literature Strouhal band.
[Section 5] No comparison is made against a standard PINN with a divergence penalty, a streamfunction parameterization, or another hard-constrained architecture under the same network size, collocation count, and training budget. Since the paper's motivation is that the algebraic divergence constraint is stiff and that artificial compressibility makes the residuals local, the empirical claims in Section 5.5 ('none of the advanced fixes were strictly necessary') are absolute statements about the AC-PINN and do not establish an advantage over existing formulations. A same-budget baseline comparison is needed to support the central methodological contribution.

minor comments (4)

[Section 7.2] The exact number of sensors, their placement distribution, and whether the sensor data are noiseless are not specified; the abstract says 'a few hundred' but the text gives no precise count, which hinders reproducibility.
[Table 4 and Section 7.3] The inverse-problem results are reported as single runs without seed variability or uncertainty estimates; given the paper's otherwise careful use of seeds in the Taylor–Green study, the viscosity-identifiability conclusion would be more robust with multiple seeds.
[Table 2 and Section 5.4] The final row of Table 2 compares the ε-continuation run (16,000 Adam epochs plus L-BFGS) with fixed-ε runs (4,000 Adam epochs plus L-BFGS); the text notes the budget difference, but the factor-of-three improvement should be explicitly described as partly a training-budget effect, not purely an effect of continuation.
[Section 7.1, Table 3] The Strouhal numbers are reported to three decimal places, but no uncertainty or statistical variability from the finite-element snapshots is given; a short statement on the temporal resolution of the Strouhal estimate would help the reader judge the 3% figure.

Circularity Check

2 steps flagged · score 5.0 of 10

Partially circular validation: the ε-divergence scaling restates Eq. (4), and the cylinder-wake accuracy is measured against the same FEM reference that supplied the sensors, not against the true flow.

self definitional [Section 5.4, Eq. (4)]
"The relaxed continuity equation (4) forces a residual divergence ∂xu +∂yv =−ε∂tp, so the divergence error scales as ε|∂tp|; for the Taylor–Green vortex the pressure decays slowly (∝e−4νt with ν = 0.05), so even the loosest relaxation ε = 10−1 yields a modest RMS divergence."

Equation (4) is ε∂tp + ∂xu + ∂yv = 0, so rearranging gives ∂xu + ∂yv = −ε∂tp. The claimed 'divergence scales as ε|∂tp|' is therefore a restatement of the model's defining equation, not an empirical result produced by the trained network. Reporting it as the ε-trade-off is a consistency check of the residual definition rather than an independent physical prediction. It is not load-bearing for the accuracy claims, but it is a genuine self-definitional element of the paper's framing.
other [Sections 7.1–7.2, Table 3]
"To provide assimilation data and a ground truth we solve the same problem with a standard finite-element method ... 100 velocity snapshots are stored on a regular grid ... the AC-PINN reproduces a clean, sustained vortex street ... with ... a shedding frequency St = 0.181, matching the FEM’s 0.176 to within 3% ... the 3% figure should therefore be read as a bound on assimilation fidelity set by the reference solver, not as an independent validation against the true shedding frequency."

The same FEM solution is used twice: its snapshots provide the sparse sensor observations, and the same FEM field is the 'ground truth' for the reported 7% velocity error and 3% frequency match. A network trained to fit those snapshots is then measured against the source of those snapshots, so the accuracy numbers quantify self-consistency with the reference, not fidelity to the true flow. The paper explicitly concedes this, but the central unsteady validation remains a self-referential loop. Moreover, because the reference Strouhal 0.176 sits above the 0.164–0.172 literature band, the recovered St = 0.181 could be roughly 7% away from the true shedding frequency while still being 'within 3%' of the reference.

full rationale

The paper is unusually transparent and includes independent benchmarks: the Taylor–Green vortex has a closed-form exact solution, and the lid-driven cavity is compared to the external Ghia et al. data. Those parts are not circular. There is no load-bearing self-citation, no imported uniqueness theorem, and no ansatz smuggled in by citation. However, two self-referential elements prevent a clean bill. First, the central ε-scaling result is an algebraic identity obtained by rearranging the defining equation (4); it is a consistency check rather than a derived prediction. Second, the strongest unsteady-flow claim—data-assimilated recovery of the cylinder wake to 7% and within 3% frequency—is validated against the same FEM reference that generated the sensors, so the match measures agreement with the reference, not with the true flow. The paper itself flags this limitation, which reduces the severity, but the reported accuracy numbers are still not externally validated. On balance, the core method has independent content, but the headline unsteady validation and the ε-scaling 'finding' are partially circular, warranting a score of 5 rather than 0 or 2.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central method rests on one hand-chosen scalar epsilon, several domain assumptions about optimizer behavior and reference fidelity, and an applied-analysis convergence theorem for AC as epsilon goes to zero. No new physical entities are introduced. The Taylor-Green benchmark has an independent closed-form solution and therefore provides the cleanest non-circular evidence; the cylinder and inverse experiments inherit their ground truth from the authors' own FEM reference.

free parameters (3)

epsilon (AC relaxation parameter) = swept 1e-4 to 1e-1; continuation ladder 0.1, 2.15e-2, 4.64e-3, 1e-3
Chosen by hand; controls the relaxation of the divergence-free constraint and is the central free scalar of the method. Not fitted to data, but the reported scaling law is a function of it.
lid smoothing width = 50
The smoothed lid u_lid(x) = tanh(50x)tanh(50(1-x)) is an ad hoc choice to soften corner singularities; the Ghia benchmark uses a sharp lid, so this may shift the reference profiles.
FEM boundary-layer resolution = about 9 cells across the O(Re^-1/2) layer
The reference Strouhal number depends on this resolution choice; it was adjusted until St=0.176, close to but above the cited literature band.

assumptions (4)

standard math Artificial-compressibility solutions converge to the incompressible Leray weak solution as epsilon approaches zero.
Invoked in Section 3 to justify the epsilon-to-zero limit; taken from the applied-analysis literature [17, 18] and not reproved.
domain assumption Low mean-squared collocation residual implies an accurate PDE solution.
The method relies on minimizing loss (5) to solve the PDE; no generalization guarantee is given, and Section 7 shows a low-residual steady branch exists for the cylinder.
ad hoc to paper The smoothed-lid cavity solution is comparable to the sharp-lid Ghia et al. benchmark.
The lid is smoothed with tanh(50x)tanh(50(1-x)), but the results are compared against the sharp-lid Ghia profiles; the paper assumes the smoothing does not compromise the comparison.
domain assumption The boundary-layer-resolved FEM solution is a sufficiently accurate ground truth for the cylinder wake.
This FEM solution is both the sensor source and the validation reference in Sections 7.1 and 7.2; its Strouhal number 0.176 is above the literature band, so the reference has known bias.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An Artificial-Compressibility Physics-Informed Neural Network for the Unsteady Incompressible Navier--Stokes Equations." pith.science (2026). https://pith.science/paper/JKCGPZHB

@misc{pith2026260804191,
  author       = {Pith},
  title        = {Pith review of: An Artificial-Compressibility Physics-Informed Neural Network for the Unsteady Incompressible Navier--Stokes Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JKCGPZHB}},
  note         = {Machine review of arXiv:2608.04191}
}

abstract

We study a physics-informed neural network (PINN) for the unsteady, two-dimensional incompressible Navier--Stokes equations in which the stiff divergence-free constraint is replaced by an artificial-compressibility (AC) relaxation governed by a single scalar parameter $\eps$. The relaxation reintroduces a pressure time derivative, converting a differential-algebraic constraint into an ordinary residual that a PINN can minimise directly. On the Taylor--Green vortex, which admits a closed-form unsteady solution, we quantify the effect of $\eps$: the residual divergence scales as $\eps\,|\partial_t p|$, so larger $\eps$ raises both the divergence and the velocity error, and both decrease monotonically and saturate as $\eps$ is reduced. On the $Re=100$ cylinder wake the plain forward AC-PINN collapses to the steady symmetric branch and does not reproduce von K\'arm\'an shedding; assimilating a few hundred sparse velocity sensors from a boundary-layer-resolved finite-element reference (whose Strouhal number, $0.176$, we bring close to the $0.164$--$0.172$ literature band by resolving the separating shear layer, though it remains just above it) recovers the unsteady vortex street to $7\%$ over the wake and its shedding frequency to within $3\%$ of that same reference --- a bound set by the reference's own fidelity rather than an independent validation against the true flow.

Figures

Figures reproduced from arXiv: 2608.04191 by the authors.

**Figure 1.** Taylor–Green vortex at t = 0.6: AC-PINN prediction (left), exact solution (6) (centre), and pointwise absolute error (right) for u, v, and p. 6 Test 2: Lid-Driven Cavity The unit-square lid-driven cavity at Re = 100 (ν = UL/Re = 10−2 , U = L = 1) is a classical benchmark with tabulated reference data [25]. The top lid moves with a smoothed horizontal profile ulid(x) = tanh(50x)tanh(50(1 − x)) that tapers to zero at … view at source ↗

**Figure 2.** Taylor–Green diagnostics. (a) Per-field error over the simulation window. (b) The [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Lid-driven cavity at Re = 100, statistically steady state (t = 20): AC-PINN centerline velocities against Ghia et al. (1982). Left: u along the vertical centerline x = 0.5. Right: v along the horizontal centerline y = 0.5. Centerline-u relative-L2 error 7.7%. attached shear layers are captured, but no coherent vortices are shed downstream, and the far field carries low-amplitude under-resolution artifacts rather tha… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Cylinder wake at Re = 100, forward AC-PINN, t = 12. The attached shear layers are resolved near the cylinder but the wake does not destabilise into a von K´arm´an street; downstream vorticity is negligible apart from far-field under-resolution artifacts. The plain forw…

**Figure 5.** Figure 5: Data-assimilating AC-PINN, wake-band vorticity at [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 23 canonical work pages

[1]

A. J. Chorin. A numerical method for solving incompressible viscous flow problems.J. Comput. Phys., 2(1):12–26, 1967. 11

work page 1967
[2]

A. J. Chorin. Numerical solution of the Navier–Stokes equations.Math. Comput., 22(104):745– 762, 1968

work page 1968
[3]

Raissi, P

M. Raissi, P. Perdikaris, G. E. Karniadakis. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.J. Comput. Phys., 378:686–707, 2019

work page 2019
[4]

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang. Physics-informed machine learning.Nat. Rev. Phys., 3:422–440, 2021

work page 2021
[5]

X. Jin, S. Cai, H. Li, G. E. Karniadakis. NSFnets (Navier–Stokes flow nets): physics-informed neural networks for the incompressible Navier–Stokes equations.J. Comput. Phys., 426:109951, 2021

work page 2021
[6]

S. Wang, Y. Teng, P. Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks.SIAM J. Sci. Comput., 43(5):A3055–A3081, 2021

work page 2021
[7]

Krishnapriyan, A

A. Krishnapriyan, A. Gholami, S. Zhe, R. Kirby, M. W. Mahoney. Characterizing possible failure modes in physics-informed neural networks. InAdvances in Neural Information Processing Systems (NeurIPS), 2021

work page 2021
[8]

Self-adaptive loss balanced Physics-informed neural networks for the incompressible Navier-Stokes equations

Z. Xiang, W. Peng, X. Zheng, X. Zhao, W. Yao. Self-adaptive loss balanced physics-informed neural networks for the incompressible Navier–Stokes equations.arXiv:2104.06217, 2021

work page Pith review arXiv 2021

Show all 25 references

[9]

S. Wang, S. Sankaran, P. Perdikaris. Respecting causality for training physics-informed neural networks.Comput. Methods Appl. Mech. Engrg., 421:116813, 2024

2024
[10]

S. Ko, S. H. Park. VS-PINN: a fast and efficient training of physics-informed neural networks us- ing variable-scaling methods for solving PDEs with stiff behavior.J. Comput. Phys., 529:113860, 2025

2025
[11]

S. Wang, X. Yu, P. Perdikaris. When and why PINNs fail to train: a neural tangent kernel perspective.J. Comput. Phys., 449:110768, 2022

2022
[12]

Tancik, P

M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. T. Barron, R. Ng. Fourier features let networks learn high frequency functions in low dimensional domains.arXiv:2006.10739, 2020

2006 arXiv
[13]

S. Wang, B. Li, Y. Chen, P. Perdikaris. PirateNets: physics-informed deep learning with residual adaptive networks.J. Mach. Learn. Res., 25(402):1–51, 2024

2024
[14]

S. Wang, S. Sankaran, H. Wang, P. Perdikaris. An expert’s guide to training physics-informed neural networks.arXiv:2308.08468, 2023

2023 arXiv
[15]

J. Shen. Pseudo-compressibility methods for the unsteady incompressible Navier–Stokes equa- tions. InProc. 1994 Beijing Symp. on Nonlinear Evolution Equations and Infinite Dynamical Systems, ZhongShan University Press, 1994

1994
[16]

Cappanera, S

L. Cappanera, S. Giordano. Artificial compressibility method for the incompressible Navier– Stokes equations with variable density.Numer. Methods Partial Differ. Equ., 2025

2025
[17]

R. Temam. Sur l’approximation de la solution des ´ equations de Navier–Stokes par la m´ ethode des pas fractionnaires (II).Arch. Ration. Mech. Anal., 33:377–385, 1969. 12

1969
[18]

Donatelli

D. Donatelli. On the artificial compressibility method for the Navier–Stokes–Fourier system. arXiv:0807.3842, 2008

2008 arXiv
[19]

J. Song, W. Cao, W. Zhang. Solving compressible Navier–Stokes equations using the feature- enhanced neural network.arXiv:2602.18791, 2026

2026
[20]

Q. Hu, I. Senocak. Solution of incompressible flow equations with physics and equality con- strained artificial neural networks.arXiv:2511.18820, 2025

2025 arXiv
[21]

Franceschini, D

L. Franceschini, D. Sipp, O. Marquet. Mean-flow data assimilation based on minimal correction of turbulence models: application to a turbulent high-Reynolds-number backward-facing step. Phys. Rev. Fluids, 5(9):094603, 2020

2020
[22]

Mons, J.-C

V. Mons, J.-C. Chassaing, P. Sagaut. Optimal sensor placement for variational data assimilation of unsteady flows past a rotationally oscillating cylinder.J. Fluid Mech., 823:230–277, 2017

2017
[23]

Zhang, A

Z. Zhang, A. Alla, G. E. Karniadakis. Adjoint method versus physics-informed neural networks in PDE-constrained inverse problems.arXiv:2606.12337, 2026

2026 arXiv
[24]

H. Son, S. W. Cho, H. J. Hwang. Enhanced physics-informed neural networks with augmented Lagrangian relaxation method (AL-PINNs).Neurocomputing, 548:126424, 2023

2023
[25]

U. Ghia, K. N. Ghia, C. T. Shin. High-Re solutions for incompressible flow using the Navier– Stokes equations and a multigrid method.J. Comput. Phys., 48(3):387–411, 1982. 13

1982

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

A. J. Chorin. A numerical method for solving incompressible viscous flow problems.J. Comput. Phys., 2(1):12–26, 1967. 11

work page 1967

[2] [2]

A. J. Chorin. Numerical solution of the Navier–Stokes equations.Math. Comput., 22(104):745– 762, 1968

work page 1968

[3] [3]

Raissi, P

M. Raissi, P. Perdikaris, G. E. Karniadakis. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.J. Comput. Phys., 378:686–707, 2019

work page 2019

[4] [4]

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang. Physics-informed machine learning.Nat. Rev. Phys., 3:422–440, 2021

work page 2021

[5] [5]

X. Jin, S. Cai, H. Li, G. E. Karniadakis. NSFnets (Navier–Stokes flow nets): physics-informed neural networks for the incompressible Navier–Stokes equations.J. Comput. Phys., 426:109951, 2021

work page 2021

[6] [6]

S. Wang, Y. Teng, P. Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks.SIAM J. Sci. Comput., 43(5):A3055–A3081, 2021

work page 2021

[7] [7]

Krishnapriyan, A

A. Krishnapriyan, A. Gholami, S. Zhe, R. Kirby, M. W. Mahoney. Characterizing possible failure modes in physics-informed neural networks. InAdvances in Neural Information Processing Systems (NeurIPS), 2021

work page 2021

[8] [8]

Self-adaptive loss balanced Physics-informed neural networks for the incompressible Navier-Stokes equations

Z. Xiang, W. Peng, X. Zheng, X. Zhao, W. Yao. Self-adaptive loss balanced physics-informed neural networks for the incompressible Navier–Stokes equations.arXiv:2104.06217, 2021

work page Pith review arXiv 2021