Deep learning accelerated solutions of incompressible Navier-Stokes equations on non-uniform Cartesian grids

This paper introduces an enhanced HyDEA framework that utilizes Mesh-Conv operators and a novel multi-level distance vector map strategy to extend deep learning-accelerated solutions of the pressure Poisson equation to non-uniform Cartesian grids, enabling robust and generalizable simulations of incompressible flows around complex immersed boundaries.

The Gist

Imagine you are trying to solve a massive, complex puzzle where every piece represents a tiny drop of water or air moving around an object, like a submarine or a bird's wing. This is what scientists call Computational Fluid Dynamics (CFD). To solve this puzzle, they use a set of mathematical rules (the Navier-Stokes equations).

However, there's a huge problem: one specific part of the puzzle—the **Pressure Poisson Equation **(PPE)—is like the "bottleneck" of a highway. It's the part that slows everything down to a crawl because it requires solving a gigantic, messy math problem over and over again.

The Old Way vs. The New Way

**The Old Way **(The "Uniform Grid" Problem)
Previously, the authors created a smart assistant called HyDEA. Think of HyDEA as a super-fast GPS that helps a hiker (the computer) find the shortest path through a forest.

• How it worked: The forest was laid out in a perfect, uniform checkerboard pattern. The GPS knew exactly how far every step was.
• The Limitation: Real life isn't a perfect checkerboard. When you get close to a rock (a solid object like a ship hull), you need tiny, detailed steps to see the cracks. Far away, you can take giant steps. This is a non-uniform grid (a grid where the spacing changes).
• The Failure: The old HyDEA GPS got confused on these uneven grids. It tried to use the same "step size" rules everywhere, which led to wrong turns and slow progress. It was like trying to use a map designed for a city grid to navigate a winding mountain trail.

**The New Solution **(MConv)
In this paper, the authors upgraded HyDEA with a new tool called **Mesh-Conv **(MConv).

• The Analogy: Imagine the old GPS just looked at the map. The new MConv GPS has a smart sensor that feels the ground under its feet.
• How it works: When the grid gets tight (near the rock), the sensor says, "Hey, the steps here are tiny, so I need to adjust my calculations." When the grid is loose (far away), it says, "Okay, we can take bigger steps."
• The Magic: They built a special "distance map" that tells the AI exactly how far apart the grid points are at every single level of the puzzle. This allows the AI to understand the shape of the terrain perfectly, even when it's uneven.

Why This Matters: The "One Size Fits All" Trick

Usually, if you want a computer to learn how to navigate a specific mountain, you have to train it on that specific mountain. If you give it a different mountain, it fails.

But this new HyDEA is a chameleon.

• The authors trained the AI not on specific rivers or ships, but on the math of the grid itself.
• The Result: Once trained, this AI can handle a circular cylinder, an oval egg shape, a complex submarine profile, or even a flapping wing, without needing to be retrained. It just looks at the grid spacing and adapts instantly. It's like teaching a driver to drive on any road by teaching them how to read the road signs, rather than memorizing every single street in the city.

The Results: Speed and Accuracy

The team tested this new system on several difficult scenarios:

1. Flow past a cylinder: Like water flowing around a pipe.
2. Flow past a submarine shape: A complex, curved object.
3. A flapping wing: Like a bird or insect wing moving up and down.

The Outcome:

• Speed: The new method solved the "bottleneck" problem 3 to 8 times faster than the traditional methods.
• Accuracy: It didn't just go fast; it was just as accurate as the slow, traditional methods. The flow patterns (like swirling vortices behind the object) looked exactly right.
• Efficiency: It reduced the number of "steps" (iterations) the computer had to take to find the answer by a huge margin.

The Bottom Line

Think of this paper as upgrading a car engine. The old engine (standard math) was powerful but slow on rough terrain. The new engine (HyDEA with MConv) has a suspension system that automatically adjusts to the bumps and dips of the road (the non-uniform grid).

This means engineers can now simulate complex fluid flows—like designing a more efficient ship or a quieter wind turbine—much faster and with less computing power, opening the door for more realistic and detailed simulations in the real world.

Technical Summary

1. Problem Statement

Incompressible flow simulations using fractional-step methods rely heavily on solving the Pressure Poisson Equation (PPE). This step typically involves solving large-scale linear systems iteratively, which constitutes the primary computational bottleneck, especially in engineering applications requiring non-uniform Cartesian grids (e.g., for local grid refinement near solid boundaries in fluid-structure interaction).

Previous hybrid deep learning approaches, such as HyDEA (Hybrid Deep lEarning line-search directions and iterative methods for Accelerated solutions), successfully accelerated PPE solutions on uniform grids by combining classical iterative solvers (like Conjugate Gradient) with Deep Learning Line-Search Methods (DLSM). However, these methods failed on non-uniform grids because standard Convolutional Neural Network (CNN) operators assume spatially shared kernel parameters and uniform grid spacing. Applying them to non-uniform data either requires interpolation (distorting boundary features) or fails to capture local resolution variations, leading to poor convergence.

2. Methodology

The authors propose an enhanced MConv-based HyDEA framework designed to solve the PPE on non-uniform Cartesian grids while maintaining the efficiency of hybrid iterative solvers.

A. Numerical Framework: Decoupled Immersed Boundary Projection Method (DIBPM)

The study utilizes the DIBPM to handle fluid-structure interaction (FSI). This method decouples the pressure, velocity, and forcing terms, reducing computational cost compared to standard Immersed Boundary Methods (IBM). The core challenge remains solving the linear system $M\delta p = S$ (Eq. 6 in the paper) at each time step.

B. Hybrid Solver Architecture (HyDEA)

The solver alternates between:

1. Classical Iterative Methods: Specifically, Preconditioned Conjugate Gradient (PCG) variants (ICPCG, JPCG, MGPCG) to handle high-frequency errors.
2. Deep Learning Line-Search Method (DLSM): A neural network predicts an approximate error vector based on the residual vector, acting as a line-search direction to rapidly eliminate low-frequency errors.

C. Key Innovation: Mesh-Conv (MConv) Operator

To adapt the DLSM to non-uniform grids, the authors replace standard CNN operators in the U-Net branch network with Mesh-Conv (MConv) operators.

• Mechanism: MConv explicitly incorporates local grid spacing information into the convolution operation. It calculates a local weight function $\bar{\eta}_{p,q}$ based on the Euclidean distance between a target node and its four adjacent neighbors.
• Simplification: Since the grids are orthogonal Cartesian, the angular component of the original MConv is redundant; only distance vectors are used.
• Multi-Level Distance Vector Map Construction: A critical technical contribution is the strategy to handle the hierarchical nature of U-Nets (which involve downsampling/upsampling). The authors generate distance vector maps ($\Delta x, \Delta y$) for every level of the U-Net by applying average pooling to nodal coordinates and computing coordinate differences. These maps are concatenated and fed into the MConv operator at each specific network level, ensuring the convolution kernels "know" the local grid resolution.

D. Training Strategy

• Data-Independent Training: The model is trained exclusively on fabricated linear systems derived from the coefficient matrix of the PPE (using Lanczos iteration to generate spectral-biased random vectors).
• Generalization: Because the training data is based on the matrix structure rather than specific flow physics, the trained model generalizes to diverse obstacle geometries (cylinders, ellipses, complex profiles) and moving boundaries without retraining.

3. Key Contributions

1. Extension to Non-Uniform Grids: Successfully adapted the HyDEA framework from uniform to non-uniform Cartesian grids, a significant hurdle for standard CNN-based solvers.
2. Novel Multi-Level Distance Mapping: Developed a strategy to construct grid-spacing maps compatible with the hierarchical structure of U-Nets, enabling MConv to function correctly across different feature map resolutions.
3. Geometry-Agnostic Generalization: Demonstrated that a single set of fixed neural network weights can solve PPEs for various complex geometries (stationary, inclined, and oscillating obstacles) and flow conditions without retraining.
4. Hybrid Efficiency: Proved that combining MConv-based DLSM with classical PCG methods significantly outperforms standalone PCG and standard convolution-based HyDEA on grids with high resolution ratios ($\Delta_{max}/\Delta_{min}$ up to 41).

4. Results and Benchmarks

The method was validated on three benchmark cases involving 2D incompressible flows with local grid refinement:

• Case 1: Lid-driven cavity with a stationary cylinder ($\Delta_{max}/\Delta_{min} \approx 2.3$).
  • HyDEA reduced iterations by ~10x compared to ICPCG.
  • MConv and standard Conv performed similarly here, as grid variation was mild.
• Case 2: Flow past various obstacles (Circular cylinder, Elliptical cylinder, DARPA SUBOFF profile, Oscillating cylinder).
  • Grid ratio $\Delta_{max}/\Delta_{min} \approx 41$.
  • Critical Finding: Standard convolution-based HyDEA failed to reduce iterations compared to standalone PCG. MConv-based HyDEA achieved massive acceleration (e.g., reducing iterations from ~90 to ~6-20 depending on the time step).
  • The model maintained high accuracy for complex shapes (DARPA SUBOFF) and moving boundaries (oscillating cylinder) using the same trained weights.
• Case 3: Flapping elliptical wing ($\Delta_{max}/\Delta_{min} \approx 28$).
  • HyDEA reduced computational time by a factor of ~3 compared to ICPCG.
  • Predicted global coefficients (Lift/Drag) matched literature values with high accuracy.

Performance Metrics:

• Acceleration: Wall-time acceleration ratios ranged from 3x to 8x depending on the preconditioner and grid non-uniformity.
• Convergence: MConv-based HyDEA consistently reached the residual tolerance ($10^{-6}$) in significantly fewer iterations than standalone methods, particularly when grid spacing varied drastically.

5. Significance and Future Work

• Significance: This work bridges the gap between high-fidelity CFD (requiring non-uniform grids) and AI acceleration. It demonstrates that deep learning can be effectively integrated into classical iterative solvers for complex, real-world geometries without the "spectral bias" or generalization issues of purely data-driven models.
• Limitations:
  • Currently restricted to 2D simulations due to GPU memory and computational overhead.
  • Limited to non-uniform Cartesian grids; arbitrary unstructured meshes would require Graph Neural Networks (GNNs), which are computationally heavier.
  • Requires retraining if the entire grid topology changes drastically (though it generalizes well to new geometries on similar grid structures).
• Future Directions: Extending the framework to 3D simulations, optimizing code efficiency for larger scales, and adapting the architecture to handle arbitrary unstructured grids.

In conclusion, the MConv-based HyDEA represents a robust, generalizable, and highly efficient solution for accelerating incompressible flow simulations on non-uniform grids, offering a promising path toward real-time or high-throughput CFD applications in engineering design.