A physics-constrained neural solver for incompressible flows without labeled data

arXiv:2511.18820 · physics.flu-dyn, cs.LG · Submitted 2025-11-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "A physics-constrained neural solver for incompressible flows without labeled data".

Paper summary: Kai: So, we're looking at this paper called "A physics-constrained neural solver for incompressible flows without labeled data". This research is tackling a problem where physics informed neural networks have struggled with incompressible flows, especially when things get really fast, like high Reynolds numbers.

Mira: Exactly. The core idea here is that existing methods need some kind of labeled data or reference solutions to work well, and this paper claims they've found a way to do it without that help.

Lev: It’s about overcoming the lack of a solid mechanism for enforcing the divergence-free constraint and boundary conditions with strict tolerances, which is what conventional solvers absolutely need for incompressible flow accuracy.

Kai: So, this method uses something called the physics and equality constrained artificial neural network framework, PECANN, paired up with a conditionally adaptive augmented Lagrangian method, CA-ALM.

Mira: That CA-ALM part is key because it recasts the problem into an unconstrained form by using Lagrange multipliers and penalty parameters to handle those constraints.

Lev: From where I look at real hardware, the overhead of these constraint enforcement mechanisms needs to be managed carefully when you try to run them on something that isn't perfectly structured.

Kai: The objective function they are minimizing is defined by minimizing the mean square residual of the pressure Poisson equation, but this minimization is subject to residuals from momentum equations in each direction, the divergence-free condition, and boundary conditions.

Mira: They set up a constrained optimization problem where they're trying to minimize that pressure Poisson residual while keeping those other equations satisfied with certain squared norms being zero over the entire domain.

Lev: If you look at how they set up the scaling factors for those PDE constraints, they have to make sure those are smaller than the scaling factors for the boundary and initial conditions so that everything locks in early.

Kai: To stabilize things, they propose two specific techniques: first, computing an entropy residual directly from the predicted velocity and pressure field by calculating (velocity minus a reference velocity) dotted with some vector F.

Mira: And then they use an exponential moving average of the squared entropy norm to update an adaptive vanishing entropy viscosity based on its unbiased root at each epoch.

Lev: That adaptive vanishing entropy viscosity is specifically for high Reynolds number cases, which sounds like it's a way to stabilize early training without messing up the final prediction fidelity.

Kai: On the benchmark results, they tested this method on several problems without any labeled data or supervised pretraining.

Mira: They showed that this approach captures the spontaneous onset of periodic vortex shedding in unsteady cylinder flow even when starting from a randomly initialized neural network.

Lev: And for the 2D lid-driven cavity flow benchmark, they found that the velocity-pressure formulation achieved good results up to Reynolds numbers of seven thousand five hundred.

Kai: So, the method seems robust across different tests, including three dee Beltrami flow and both steady and unsteady two-dimensional cylinder flows <ref:2511.18820#pg2>.

Mira: It's important to remember this is a meshless flow solver where all the flow variables are represented entirely by neural networks trained on randomly distributed collocation points.

Lev: That means the training cost for this method ends up being roughly two orders of magnitude higher than that of conventional solvers, which is a significant overhead for practical use.

Kai: So, what does this mean for us? It shows that even with all those constraint enforcement machinery, the baseline naïve formulation just doesn't work on its own.

Mira: The paper argues that this pressure-Poisson-based formulation is a new way to enforce the divergence-free condition that’s fundamental to solving these incompressible flow equations.

Lev: For someone building real hardware, it implies that you need this kind of physics constraint baked into your loss function if you want a neural network approach to actually work for fluid dynamics.

Kai: The authors say they are providing a robust foundation for future extensions toward things like complex moving boundaries and inverse problems in fluid flow.

Mira: It suggests that we might be able to get better results in these complex areas without relying on pre-existing data, which is a big deal.

Conclusion: Kai: So, this paper is about taking those physics informed neural networks, which are usually good for some stuff, and trying to make them actually work for incompressible flows without needing any pre-labeled data or reference solutions.

Mira: Right, it's titled "A physics-constrained neural solver for incompressible flows without labeled data." The authors are tackling the fundamental problem that these models struggle when things get really complex, like high Reynolds numbers.

Lev: They’re using a pressure Poisson-based objective function to force the divergence-free constraint and boundary conditions to be very tight tolerances. That sounds like a lot of work just to set up the math right.

Kai: They are using this physics and equality constrained artificial neural network framework, PECANN, combined with something called a conditionally adaptive augmented Lagrangian method, CA-ALM. It’s essentially minimizing the pressure Poisson equation while making sure all those momentum and continuity equations stay zero over the whole domain.

Mira: That means they're setting up an unconstrained problem by using Lagrange multipliers and penalty parameters to handle all those physical constraints simultaneously. It’s a clever way to structure the loss function.

Lev: From what I see, the scaling factors for those differential equation constraints have to be smaller than the ones for boundary conditions and initial conditions so that everything locks in early during training.

Kai: They also added some stabilization techniques, like computing an entropy residual directly from the predicted velocity and pressure field to keep things from blowing up. And they use an adaptive vanishing entropy viscosity specifically for high Reynolds number cases to stabilize early training without messing up the final prediction.

Mira: The results are interesting because they tested this on problems without any labeled data or supervised pretraining. They showed it can capture the spontaneous onset of periodic vortex shedding in unsteady cylinder flow even starting from a randomly initialized network.

Lev: And for the 2D lid-driven cavity flow, they got excellent agreement with reference data up to Reynolds numbers of seven thousand five hundred <ref:2511.18820#pg2>. That’s a big number for this kind of setup.

Kai: So, the authors are showing that even with all these constraint enforcement mechanisms in place, the baseline method isn't working on its own. This approach is a meshless flow solver where all the variables are just neural networks trained on randomly distributed points.

Mira: It shows that this formulation is robust across three dee Beltrami flow and both steady and unsteady two-dimensional cylinder flows. But Lev, how much real hardware overhead are we talking about for someone trying to actually run this thing?

Lev: The training cost is roughly two orders of magnitude higher than conventional solvers, which is a massive hurdle when you're thinking about putting this on real hardware.

Kai: So, the main point here is that they’ve created a robust foundation for future extensions toward complex moving boundaries and inverse problems in fluid flow. It shows that we might be able to get better results without relying on pre-existing data, which is a big deal for many applications.

Mira: It suggests that baking these physical constraints directly into the loss function is the way forward if we want neural networks to handle incompressible flows reliably. But Lev, what's next? Where does this lead us after they’ve proven it works on these specific benchmarks?

Qifeng Hu, Inanc Senocak

Department of Mechanical Engineering and Materials Science, University of Pittsburgh

physics.flu-dyn, cs.LG

Submitted: 2025-11-24

Updated: 2026-11-17

Comments: 36 pages, 24 figures

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: Physics-informed neural networks (PINNs) have shown considerable promise for solving partial differential equations…

Key concepts

PECANN
Physics-and equality-constrained artificial neural network framework used to solve problems where physical laws must be strictly satisfied. It minimizes the residual of the pressure Poisson equation while treating momentum and continuity equations as constraints on the network's output.
CA-ALM
Conditionally adaptive augmented Lagrangian method, a technique used to transform constrained optimization problems into unconstrained ones. It uses Lagrange multipliers and penalty parameters to enforce physical constraints like divergence-free flow in a manageable way during training.
Entropy Residual
A specific residual calculated by comparing the predicted velocity field against a reference velocity vector. This residual is used to adaptively update an entropy viscosity, which helps stabilize the training process, especially for high Reynolds number flows.
Meshless Flow Solver
A solver where the fluid flow variables are entirely represented by neural networks trained on randomly distributed points in space, rather than a traditional grid. This allows for flexible representation of complex geometries and boundary conditions.

Terminology

Summary

Physics-informed neural networks (PINNs) have shown considerable promise for solving partial differential equations, yet they have achieved limited success in simulating incompressible flows, particularly at high Reynolds numbers The proposed method introduces a pressure-Poisson-based objective function to enforce the divergence-free constraint and boundary conditions to strict tolerances

The Core Methodology

The method adopts the physics- and equality-constrained artificial neural network (PECANN) framework in conjunction with a conditionally adaptive augmented Lagrangian method (CA-ALM) This involves minimizing the residual of the pressure Poisson equation subject to momentum and continuity equations as equality constraints on primitive variables The objective function is defined by minimizing the mean square residual of the pressure Poisson equation This objective is constrained by residuals of momentum equations in each direction, the divergence-free condition, and boundary conditions The CA-ALM recasts this problem into an unconstrained form using Lagrange multipliers and penalty parameters

Constraint Enforcement Mechanisms

The paper details several mechanisms used to enforce the necessary physical constraints

[6] The constrained optimization problem is expressed as follows:

(6) min θ J (θ; x, t) = 1/NU X NU i=1 P(θ; xi, ti) squared 2 in U, subject to CF,k(θ; x, t) = 1 NU X NU i=1 Fk(θ; xi, ti) squared 2 = 0 in U, CM(θ; x, t) = 1 NU X NU i=1 M(θ; xi, ti) squared 2 = 0 in U...

The scaling factors for the PDE constraints should be chosen smaller than those for boundary and initial conditions to ensure boundary and initial conditions are satisfied early

Stabilization Techniques

To enhance robustness, the authors propose two specific stabilization techniques

(9) We compute the entropy residual directly from the predicted velocity and pressure field as r(x, t) = (u(x, t) − um) · F(x, t), where um represents a reference velocity vector.

(10) we maintain an exponential moving average of the squared entropy norm (r∞)2 and use its unbiased root to update the entropy viscosity νa at epoch e. p. 7

Where the paper says it

Following the same procedure as CA-ALM, (Fig. 1), we maintain an exponential moving average of the squared entropy norm (||r||∞) 2 and use its unbiased root to update the entropy viscosity ν a a at epoch e: κ (e) ← ζκ(e−1) + (1 − ζ)(||r||∞) 2 , ν a(e) a ← min(ν a(e−1) a , √ κ e 1 − ζ e L 2 U2 ||r||∞), if ν a(e−1) a > 1 4Re , otherwise 0, (10) where U and L are characteristic velocity and length scales used in defining the Reynolds number, respectively.

..

This adaptive vanishing entropy viscosity is applied exclusively to high-Reynolds-number cases, stabilizing early training without influencing final predictions

Performance Across Benchmarks

The method was assessed on several benchmark problems without labeled data or supervised pretraining

**(3) The simulation of high-Reynolds-number incompressible flows over complex geometries has long been, and continues to be, a central focus of computational fluid dynamics (CFD). **

The method captures the spontaneous onset of periodic vortex shedding in unsteady cylinder flow without external perturbations starting from a randomly initialized network The proposed pressure-Poisson-based formulation is shown to yield excellent agreement with reference data for 2D lid-driven cavity flow up to Re = 7,500

Conclusion and Generalization

The study demonstrates that the baseline naïve formulation proves ineffective even with constraint-enforcement machinery The proposed method constitutes a meshless flow solver where flow variables are represented entirely by neural networks trained on randomly distributed collocation points The method's robustness is confirmed across 3D Beltrami flow and both steady and unsteady two-dimensional cylinder flows The method captures the spontaneous onset of periodic vortex shedding in unsteady cylinder flow without external perturbations starting from a randomly initialized network The training cost remains roughly two orders of magnitude higher than that of conventional solvers The proposed method constitutes a meshless flow solver in which flow variables are represented entirely by neural networks trained on randomly distributed collocation points, providing a robust foundation for future extensions toward complex moving boundaries and inverse problems in fluid flow

--- Page 30 ---

Acknowledgments

This material is based upon work supported by the National Science Foundation under Grant No.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved constraint enforcement for incompressible flows by adopting a pressure-Poisson-based objective function, which ensures that both the pressure and velocity fields are physically consistent by construction. The system can now solve incompressible flow equations without auxiliary labeled data or supervised pretraining.

  2. Enhanced stability for advection-dominated, high-Reynolds-number flows through an adaptive vanishing entropy viscosity that stabilizes the early stages of training without influencing the final predictions. This allows the AI system to train on challenging regimes like Re up to 7500 reliably from random initialization.

  3. Increased network expressive capacity and handling of input variations via a Fourier feature mapping, which is described as projecting input coordinates into a higher-dimensional space via sinusoidal transformations, thereby enhancing the expressive capacity of MLP networks. This enables better capture of high-frequency components in the input space for complex flow structures.

  4. Rigorous constraint satisfaction using the conditionally adaptive augmented Lagrangian method (CA-ALM), which enforces all constraints to strict tolerances by assigning independent Lagrange multipliers and penalty parameters to each constraint. This provides a principled means of enforcing these heterogeneous constraints to the strict tolerances required for physically consistent predictions.

  5. Robust handling of complex boundary conditions in external flow scenarios by enforcing global mass-flux conservation at the outlet, as demonstrated by Figure 9(b), which shows a physical solution is recovered when global mass-flux conservation is enforced at the outlet as a constraint.

Abstract

We present a physics-constrained neural solver for incompressible flow that requires neither labeled data nor supervised pretraining. The solver enforces incompressibility and boundary conditions (BCs) to strict tolerances. We attribute the limited Reynolds numbers of existing physics-informed approaches to insufficient enforcement of these constraints, on which conventional finite-difference and finite-volume solvers depend. We therefore formulate incompressible flow as a constrained optimization problem within the physics- and equality-constrained artificial neural network (PECANN) framework: the pressure Poisson residual is minimized subject to the momentum equations, continuity equation, and BCs imposed as equality constraints on the primitive variables, while a conditionally adaptive augmented Lagrangian method enforces each constraint. The pressure Poisson objective proves critical. Across all benchmarks, replacing it with the momentum residual fails to converge to accurate solutions under identical architecture and constraint enforcement, which we attribute to the global elliptic coupling of pressure and velocity that local momentum-residual minimization lacks. For advection-dominated flows, we introduce an adaptive vanishing entropy viscosity that stabilizes training while leaving the converged solution unchanged. The solver predicts lid-driven cavity flow up to Re = 7,500 with centerline velocity profiles matching benchmark data, and is further validated on flow over a backward-facing step at Re = 800, 3D unsteady Beltrami flow, and steady and unsteady flow past a circular cylinder under an inflow-outflow configuration, for which an ablation study identifies admissible outlet BCs for velocity and pressure. From a randomly initialized network and without external perturbations, the solver captures the spontaneous onset of periodic vortex shedding in unsteady cylinder flow.

Sources

Related papers