SMART: Scalable Mesh-free Aerodynamic Simulations from Raw Geometries using a Transformer-based Surrogate Model

SMART is a transformer-based neural surrogate model designed for aerodynamic simulations that operates without requiring a simulation mesh, utilizing only a point-cloud representation of geometries. It encodes geometry and simulation parameters into a shared latent space and employs a physics decoder to map spatial queries to physical quantities, allowing for efficient predictions at arbitrary locations. This approach significantly reduces computational costs associated with mesh generation while achieving competitive performance against traditional mesh-dependent models, making it a valuable tool for practitioners in fields requiring high-fidelity simulations.