MUSA-PINN: Multi-scale Weak-form Physics-Informed Neural Networks for Fluid Flow in Complex Geometries

The paper introduces MUSA-PINN, a multi-scale weak-form Physics-Informed Neural Network that reformulates PDE constraints as integral conservation laws over hierarchical control volumes to overcome convergence pathologies and significantly improve accuracy and mass conservation in fluid flow simulations within complex geometries like Triply Periodic Minimal Surfaces.

The Gist

Imagine you are trying to predict how water flows through a very complicated, twisting maze of pipes. This isn't just a straight pipe; it's a 3D labyrinth with curves, branches, and tiny nooks, like the inside of a high-tech heat exchanger or a liquid-cooling plate for a supercomputer.

Traditionally, engineers use a method called CFD (Computational Fluid Dynamics). Think of this as building a physical model of the maze out of millions of tiny Lego bricks (a mesh). You simulate the water moving through these bricks. It's accurate, but it's incredibly slow and expensive, especially if you want to test 100 different maze designs.

Enter PINNs (Physics-Informed Neural Networks). This is a newer, "mesh-free" approach. Instead of building Lego bricks, you give a smart AI (a neural network) the rules of physics (like "water can't disappear" and "it can't flow through walls") and ask it to guess the flow pattern. It's like asking a genius to visualize the water flow in their head without needing a physical model.

The Problem: The "Local" Blind Spot
The paper argues that standard AI (PINNs) has a major flaw when the maze gets too twisty.

• The Analogy: Imagine you are trying to learn the layout of a giant, winding cave system by only looking at one small spot at a time. You check a spot, see the water is flowing right, and say, "Okay, that's good." Then you move to the next spot.
• The Failure: Because the AI only checks tiny, isolated points, it doesn't realize that the water it saw 10 feet ago is supposed to connect to the water 10 feet away. In a twisting maze, these small, isolated checks add up to a big mess. The AI might think water is disappearing or appearing out of nowhere because it can't "see" the whole path. It gets confused, the math breaks, and the prediction fails.

The Solution: MUSA-PINN (The "Control Volume" Detective)
The authors propose a new method called MUSA-PINN. Instead of checking tiny, isolated points, they change the strategy to check entire neighborhoods at once.

Here is how it works, using simple metaphors:

1. The "Control Volume" (The Bubble Check)

Instead of asking, "Is the water flowing correctly at this single pixel?", MUSA-PINN draws a bubble (a sphere) around a section of the pipe.

• The Rule: It applies a fundamental law of physics: Conservation. Whatever water flows into the bubble must flow out of it. Nothing can vanish inside the bubble.
• The Magic: By checking the surface of the bubble (the skin), the AI ensures that the water is balanced globally. It's like a bank teller counting the money coming in and going out of a vault, rather than just checking a single bill. This forces the AI to respect the "big picture" rules of physics, even in a twisting maze.

2. The "Multi-Scale" Strategy (The Three-Layered Net)

A single bubble size doesn't work for a complex maze. You need different sizes to catch different problems. The authors use three layers of bubbles:

• The Big Bubbles (Long-Range): These are huge bubbles that span long distances. They act like a macro-manager, ensuring that the water entering the start of the maze eventually reaches the end. They stop the AI from getting "lost" over long distances.
• The Medium Bubbles (The Skeleton): These bubbles are placed right along the "spine" or centerline of the pipes. They act like a guide dog, following the winding path to make sure the water stays in the channel and doesn't leak into the walls.
• The Small Bubbles (The Details): These are tiny bubbles that zoom in on sharp corners and tight turns. They act like a microscope, fixing small errors and ensuring the water flows smoothly right next to the walls.

3. The "Two-Stage" Training (Warm-up then Sprint)

Training the AI is like teaching a child to ride a bike.

• Stage 1 (The Training Wheels): First, the AI is told to focus only on making sure the water doesn't disappear (Mass Conservation). It learns to keep the flow balanced.
• Stage 2 (The Sprint): Once the flow is balanced, the AI is allowed to focus on the harder stuff: how the water pushes and pulls on itself (Momentum). This prevents the AI from getting overwhelmed by trying to solve everything at once.

The Result

When they tested this on complex, twisting 3D shapes (called TPMS, which look like intricate biological structures), the old AI methods failed miserably, with errors up to 90%. The new MUSA-PINN method reduced errors by up to 93%.

In Summary:
If standard AI is like a tourist trying to navigate a maze by looking at a single step at a time, MUSA-PINN is like a drone flying over the maze, checking entire sections at once to ensure the path makes sense. By using "bubbles" of different sizes to check the rules of physics, it solves the problem of complex fluid flow much faster and more accurately than ever before, without needing to build millions of tiny Lego bricks.

Technical Summary

1. Problem Statement

The paper addresses the limitations of standard Physics-Informed Neural Networks (PINNs) when solving steady, incompressible Navier-Stokes equations in topologically complex geometries, such as Triply Periodic Minimal Surfaces (TPMS) used in heat exchangers and liquid cooling plates.

• The Core Issue: Standard PINNs rely on point-wise residual minimization (strong-form). In complex, tortuous channels with high aspect ratios, this approach suffers from:
  • Local-Global Inconsistency: Point-wise constraints fail to propagate global physical information (like mass conservation) through winding paths.
  • Convergence Pathologies: The optimization landscape becomes unstable, leading to gradient issues and violations of fundamental conservation laws (mass and momentum).
  • Mesh Generation Bottlenecks: While traditional CFD requires expensive mesh generation for such geometries, standard PINNs fail to provide accurate mesh-free alternatives for these specific industrial cases.

2. Methodology: MUSA-PINN

The authors propose MUSA-PINN (Multi-scale Weak-form PINN), a hybrid framework that reformulates PDE constraints as integral conservation laws over hierarchical control volumes.

A. Weak-Form Formulation via Divergence Theorem

Instead of minimizing residuals at single points, MUSA-PINN enforces conservation over spherical control volumes ($V(c, r)$) intersected with the fluid domain.

• Derivation: The governing equations (Continuity and Momentum) are integrated over a control volume $V$. Using the Gauss Divergence Theorem, volumetric integrals are converted into surface flux integrals over the boundary $\partial V$.
• Residuals:
  • Continuity: $\oint_{\partial V} \mathbf{u} \cdot \mathbf{n} \, dS = 0$ (Net flux is zero).
  • Momentum: $\oint_{\partial V} (\mathbf{u} \otimes \mathbf{u} + p\mathbf{I} - \frac{1}{Re}\nabla \mathbf{u}) \cdot \mathbf{n} \, dS = 0$ (Flux balance).
• Implementation: These surface integrals are evaluated using Monte Carlo sampling on the control volume boundaries. This avoids volumetric meshing while providing a "globally informative" training signal that is more robust than point-wise residuals.

B. Multi-Scale Subdomain Placement Strategy

To handle the diverse scales of industrial channels (long-range transport, branching, and local details), the method employs a three-level strategy for placing control volume centers ($C_L, C_M, C_S$):

1. Large-Scale ($C_L$): Large radii placed randomly within the bounding box. These enforce long-range conservation and global flow organization, preventing large-scale drift.
2. Medium-Scale ($C_M$): Radii aligned with the geometric skeleton (extracted via Mean Curvature Flow). These ensure conservation signals follow the actual transport pathways, covering slender and branching channels effectively.
3. Small-Scale ($C_S$): Small radii sampled uniformly to refine local details and stabilize boundary layer behaviors.

C. Two-Stage Training Schedule

The optimization objective combines strong-form residuals (point-wise) with the multi-scale weak-form losses. A two-stage training schedule is introduced:

• Stage 1 (Warm-up): Prioritizes the weak-form continuity constraint (mass conservation). This quickly suppresses long-range mass imbalances.
• Stage 2 (Fine-tuning): Activates the weak-form momentum constraint to refine the velocity field and ensure physical consistency in complex passages.

3. Key Contributions

1. Weak-Form Conservation: Introduced a divergence-theorem-based formulation that enforces mass and momentum via surface flux integrals over spherical control volumes, providing a mesh-free, physically structured supervision signal.
2. Topology-Aware Multi-Scale Strategy: Designed a novel subdomain placement strategy (Large, Skeleton-aligned Medium, Small) that bridges global consistency and local fidelity without manual domain decomposition.
3. Robust Training Schedule: Developed a two-stage optimization protocol that prioritizes continuity before momentum, stabilizing convergence in complex geometries.
4. Benchmarking: Established a rigorous benchmark on TPMS heat-exchanger channels and industrial components (liquid cooling plates), demonstrating significant improvements over state-of-the-art baselines.

4. Experimental Results

The method was evaluated on four TPMS geometries (Primitive, Gyroid, Diamond, IWP) and industrial liquid cooling plates, comparing against baselines like PINN-PE, RoPINN, CoPINN, and MDPINN-GD.

• Accuracy: MUSA-PINN significantly outperformed all baselines.
  • On the challenging Diamond and IWP geometries, standard baselines often collapsed (relative errors >90%), while MUSA-PINN maintained stable convergence.
  • Relative $\ell_2$ error reduction: Up to 93% reduction in error compared to the best baseline.
  • On the Gyroid channel, MUSA-PINN achieved a relative $\ell_2$ error of 6.13% for velocity, compared to >83% for other methods.
• Physical Consistency:
  • Mass Conservation: Baselines exhibited significant mass leakage (flux imbalance) along the flow direction. MUSA-PINN maintained strict flux continuity, reducing cross-sectional mass error by over 97% compared to the best baseline.
  • Flow Structures: Visualizations showed MUSA-PINN successfully captured intricate secondary flows and boundary layer profiles that were smoothed out or lost by other methods.
• Scalability: The method demonstrated robustness at higher Reynolds numbers (Re=200, 400) and on complex industrial geometries (DualMS channels, liquid cooling plates).

5. Significance and Impact

• Solving the "Complex Geometry" Gap: This work bridges a critical gap in scientific machine learning, enabling PINNs to solve fluid dynamics problems in topologically complex, industrial-grade geometries where traditional point-wise PINNs fail.
• Mesh-Free Efficiency: By using Monte Carlo surface sampling on control volumes, it retains the mesh-free advantage of PINNs while overcoming the locality bias that plagues them in tortuous domains.
• Engineering Application: The ability to accurately simulate flow in heat exchangers and cooling plates without generating high-quality volumetric meshes offers a pathway for faster design exploration and optimization in thermal management systems.
• Future Directions: The authors note potential extensions to transient (time-dependent) regimes and conjugate heat transfer (multiphysics), suggesting a broader applicability for component-level flow simulation.

In summary, MUSA-PINN represents a significant advancement in physics-informed deep learning by shifting from local point-wise constraints to hierarchical integral conservation laws, effectively solving the stability and accuracy issues of fluid flow simulation in complex, real-world geometries.