SMART: Transformer-based model slashes aerodynamic simulation costs by skipping meshes

SMART (Scalable Mesh-free Aerodynamic Surrogate with a Transformer) is a new machine learning model that predicts aerodynamic properties like pressure and drag from raw 3D geometries, without ever building a simulation mesh. Traditional numerical solvers first require a high-quality mesh around the object (e.g., a car body), a process that can take hours or days. Mesh-based surrogate models also still need that mesh as input. SMART eliminates this step entirely by working with a simple point cloud—a set of 3D coordinates sampled from the object's surface. It uses a transformer encoder to fuse geometric and simulation parameter information into a shared latent space. A physics decoder then attends to encoder layers at multiple depths via cross-attention, jointly updating geometric features and the evolving physical field. This design allows SMART to answer queries at arbitrary spatial locations, not just at mesh points.

Extensive experiments on standard aerodynamic benchmarks show that SMART is competitive with—and often outperforms—existing surrogate models that require the simulation mesh as input. The model was accepted at ICML 2026, a top machine learning conference. For industries like automotive, aerospace, and wind energy, SMART could drastically reduce the time and cost of early design iterations. Engineers could explore thousands of shape variations in minutes instead of days, using only the CAD model. The mesh-free nature also means it can handle complex, non-watertight geometries that are problematic for meshing algorithms.