Abstract
High-resolution computational fluid dynamics (CFD) simulations are critical for accurately capturing complex phenomena such as turbulent flows and shock-boundary layer interactions. However, the enormous computational cost of fine-resolution simulations often limits their practical applicability, driving the need for efficient surrogate approaches.
We propose PointSAGE, a novel mesh-independent framework that leverages the unordered, mesh-less nature of point clouds to learn complex fluid flow and directly predict fine-resolution CFD simulations from coarse-resolution data, completely disregarding mesh information. Unlike existing super-resolution methods such as UNets and GANs that are limited to structured data and fail on unstructured meshes, PointSAGE combines a Global Feature Extractor (inspired by PointNet) and a Local Feature Extractor (using SAGEConv/GraphSAGE) to capture both global and local inter-dependencies in fluid flow.
The model achieves significant speedups of 30X to 92X over traditional CFD simulations and can predict fine-resolution data across diverse point cloud sizes regardless of the training dataset dimensions, demonstrating true mesh independence and scalability.
Key Highlights
Architecture
PointSAGE integrates two core modules: a Global Feature Extractor inspired by PointNet that captures global flow patterns through shared multi-layer perceptrons and symmetric aggregation functions, and a Local Feature Extractor built on SAGEConv (GraphSAGE) that learns neighborhood-aware local representations by aggregating features from spatially proximate points. Together, these modules provide a comprehensive representation of both macro-scale flow structures and fine-grained local variations.
Key Contributions
- Mesh-Independent: First framework to perform CFD super-resolution using unstructured point clouds without mesh information.
- Dual Feature Extraction: Combines global features (PointNet-inspired) and local features (GraphSAGE) for comprehensive flow representation.
- Significant Speedup: Achieves 30X-92X computational acceleration over traditional CFD simulations.
- Scalability: Can predict on arbitrary point cloud sizes regardless of training dimensions, enabling flexible deployment across diverse simulation setups.
Results
PointSAGE demonstrates strong predictive accuracy across multiple CFD benchmark scenarios, closely matching high-fidelity simulation outputs while delivering substantial computational acceleration. The framework generalizes effectively to unseen point cloud resolutions and geometries.
