Unsupervised simulation of incompressible flows with physics- and equality- constrained artificial neural networks

This paper introduces an unsupervised physics- and equality-constrained neural network framework utilizing a pressure-Poisson objective and adaptive augmented Lagrangian method to successfully simulate high-Reynolds-number incompressible flows without labeled data, overcoming previous limitations in enforcing strict divergence-free constraints and boundary conditions.

The Gist

Imagine you are trying to teach a robot how to predict how water flows around a rock or inside a box. Usually, to teach a robot this, you need to show it thousands of videos of water flowing (labeled data) so it can learn by example. This is like teaching a child to ride a bike by showing them a million videos of other kids riding.

This paper introduces a new way to teach the robot. Instead of showing it videos, we just give it the rules of the universe (the laws of physics) and say, "Figure it out." The robot has to learn the flow purely by trying to obey these rules, without any prior examples. This is called "unsupervised learning."

However, there's a catch. When water moves fast (high speed), it gets chaotic and tricky. Previous attempts by robots to learn these fast flows using only rules often failed. They would get confused, and the water would magically disappear or behave in impossible ways.

The Problem: The "Leaky Bucket"

In physics, water is incompressible, meaning you can't squeeze it into a smaller space. If water flows into a room, an equal amount must flow out. If your robot's prediction doesn't balance this perfectly, it's like a bucket with a hole in the bottom; the math breaks down.

Old methods tried to force the robot to follow the rules, but they were too loose. The robot would say, "I'm mostly following the rules," and that wasn't good enough for fast, complex flows.

The Solution: A Strict Teacher with a Special Scorecard

The authors built a new system called PECANN. Think of this as a very strict teacher who uses a special grading system.

1. The Scorecard (The Objective): Instead of just asking the robot to follow the basic flow rules, the teacher gives it a specific, hard-to-get-right test: the Pressure Poisson Equation.

   • Analogy: Imagine you are trying to balance a stack of plates. The basic rules say "don't drop them." But the Pressure Poisson Equation is like a specific rule that says, "The stack must be perfectly flat, or the whole thing collapses." The robot's main goal is to minimize the "wobble" of this stack. If the stack wobbles, the robot knows it's wrong.

2. The Strict Teacher (The Constraints): The robot isn't allowed to just get close to the answer. It must hit the target exactly. The authors use a method called CA-ALM (Conditionally Adaptive Augmented Lagrangian Method).

   • Analogy: Imagine a robot trying to walk a tightrope. Old methods let the robot sway a little and say, "That's close enough." This new method is like a coach who yells, "Stop! You are 1 millimeter off! Fix it immediately!" The coach adjusts the pressure on the robot's feet dynamically until it is perfectly balanced.

3. The Training Wheels (Adaptive Viscosity): When the robot starts learning fast flows, it gets shaky and might fall over. To help, the authors add a temporary "training wheel" called Adaptive Vanishing Entropy Viscosity.

   • Analogy: This is like adding a little bit of honey to the water to make it flow slower and smoother while the robot learns the basics. Once the robot gets the hang of it, the honey is magically removed, and the water flows naturally again. The robot learns the fast flow without the honey, but the honey helped it get started.

What Did They Prove?

The team tested this new "Strict Teacher" system on three famous challenges:

• The Moving Lid (Cavity Flow): Imagine a box where the top lid slides back and forth, dragging the water inside. They tested this at very high speeds (Reynolds numbers up to 7,500).
  • Result: The robot predicted the swirling vortices (eddies) perfectly, matching the best traditional computer simulations, even without seeing any training videos.
• The 3D Twist (Beltrami Flow): A complex, twisting 3D flow that has a known mathematical answer.
  • Result: The robot was much more accurate than previous AI methods, getting the pressure and speed right with very little error.
• The Cylinder (Flow Past a Rock): Water flowing past a cylinder. At a certain speed, the water stops flowing smoothly and starts shedding vortices (swirls) in a rhythmic pattern (like a flag flapping in the wind).
  • Result: This is the "holy grail." The robot started with a random guess and spontaneously figured out that the water would start flapping and shedding swirls, without anyone telling it to do so. It captured the exact rhythm of the flapping.

The Bottom Line

The paper claims that by changing what the robot tries to minimize (focusing on the pressure balance) and how strictly it enforces the rules (using the strict teacher method), they finally solved the problem of simulating fast, complex water flows using only the laws of physics.

They did this without using any pre-recorded data or "cheating" with known answers. The robot learned the flow from scratch, just by trying to obey the rules of physics perfectly. This is a big step toward using AI to replace traditional, heavy-duty computer simulations for fluid dynamics.

Technical Summary

Technical Summary: Unsupervised Simulation of Incompressible Flows with Physics- and Equality-Constrained Artificial Neural Networks

Problem Statement
While Physics-Informed Neural Networks (PINNs) have demonstrated promise in solving partial differential equations, their application to incompressible flows, particularly at high Reynolds numbers, has been limited. Existing approaches often rely on auxiliary labeled data, supervised pretraining, or analytical reference solutions to achieve convergence. A purely unsupervised method—comparable to conventional finite-difference or finite-volume solvers in its ability to enforce strict physical constraints without external data—has not been successfully demonstrated. The authors attribute this gap to the lack of a robust mechanism for enforcing the divergence-free constraint and boundary conditions to strict tolerances, a capability fundamental to traditional incompressible flow solvers.

Methodology
The proposed method utilizes the Physics- and Equality-Constrained Artificial Neural Network (PECANN) framework combined with a Conditionally Adaptive Augmented Lagrangian Method (CA-ALM). The core innovation lies in the formulation of the optimization problem:

1. Pressure-Poisson-Based Objective: Unlike baseline approaches that minimize momentum residuals directly, this method adopts the residual of the pressure Poisson equation as the objective function ($J$). This objective is minimized subject to the momentum equations, the continuity (divergence-free) equation, and boundary/initial conditions, all enforced as strict equality constraints.
2. Constraint Enforcement (CA-ALM): The CA-ALM algorithm assigns independent Lagrange multipliers and penalty parameters to each constraint, updating them adaptively to ensure heterogeneous constraints (PDEs, boundary conditions, initial conditions) are satisfied to strict tolerances.
3. Adaptive Vanishing Entropy Viscosity: To stabilize training for advection-dominated, high-Reynolds-number flows without influencing the final solution, an adaptive vanishing entropy viscosity is introduced. This artificial viscosity is active only during early training stages and vanishes once the optimization converges, ensuring the final solution reflects the true Reynolds number.
4. Network Architecture: The method employs Fourier feature mappings to enhance the network's ability to capture high-frequency components, concatenated with original inputs to preserve low-frequency behavior.

Key Contributions and Ablation Studies
The paper establishes that the choice of objective function is critical. A baseline formulation that minimizes momentum residuals directly (omitting the pressure Poisson equation) proved ineffective even with the same constraint enforcement machinery. The pressure-Poisson-based formulation ensures that both velocity and pressure fields are physically consistent by construction.

Furthermore, the study identifies critical conditions for outflow boundaries in external flow problems. Through ablation studies on flow past a cylinder, the authors demonstrate that enforcing global mass-flux conservation at the outlet is essential for recovering physical solutions. A zero-Neumann condition on the pressure gradient, combined with this mass-flux constraint, yielded the best agreement with reference data.

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

• 2D Lid-Driven Cavity (LDC): Simulations were performed up to $Re = 7,500$. The proposed formulation achieved excellent agreement with reference data (Ghia et al., Erturk et al.), capturing corner vortices and velocity profiles accurately. Baseline formulations failed to produce acceptable results beyond $Re = 400$.
• 3D Unsteady Beltrami Flow: The method achieved an order-of-magnitude reduction in error compared to previous NSFnets results, with pressure predictions being two orders of magnitude more accurate.
• Flow Past a Circular Cylinder:
  • Steady ($Re=40$): The method successfully recovered physical solutions with drag coefficients matching literature values, identifying the necessity of mass-flux constraints at the outlet.
  • Unsteady ($Re=100$): Using a discrete-time formulation with BDF2, the method captured the spontaneous onset of periodic vortex shedding (von Kármán street) starting from a randomly initialized network with zero initial velocity and no external perturbations. The predicted Strouhal number and lift coefficients aligned closely with Ansys Fluent simulations.

Significance
The paper claims that this work provides the first purely unsupervised neural-network-based solver for incompressible flows that rivals conventional solvers in enforcing physical constraints. By rigorously enforcing the divergence-free condition and boundary conditions through equality constraints, the method overcomes the primary limitation of previous PINN approaches. The ability to capture spontaneous vortex shedding in unsteady flows without external perturbations or labeled data demonstrates a significant step toward meshless, data-free computational fluid dynamics. While the training cost remains higher than conventional solvers, the authors suggest this gap can be mitigated through high-performance frameworks and domain decomposition, providing a foundation for future extensions to complex moving boundaries and inverse problems.