Built on Graph Convolutional Networks (GCNs) tailored for unstructured CFD meshes
Uses a multi-resolution autoencoder to compress and reconstruct aerodynamic fields
Tested on the NASA Common Research Model wing/body configuration
Predicts pressure and shear-stress coefficients across nonlinear transonic regimes
Impact
Drastically reduced computational cost compared to traditional CFD
Enables fast, scalable aerodynamic analysis for design and optimization
Opens new possibilities for real-time simulation and control in aerospace engineering
Benefits Summary
Lightweight: while other physics AI approaches require top-spec high-memory GPUs, our lightweight implementation runs on your laptop, with a GPU memory as lows as 6 GB.
Efficiency: AI models learn complex flow patterns, reducing simulation time from hours to seconds
Scalability: Easily adaptable to different aircraft geometries and flight conditions
Insight: Reveals hidden aerodynamic relationships through learned representations
Integration: Ideal for embedding into digital twins, design loops, and autonomous systems
