Equivariance in Neural Fluid Surrogates: A Study on CFD Acceleration
This paper investigates the role of group-equivariant architectures in neural surrogates for computational fluid dynamics (CFD). Neural surrogates can accelerate CFD simulations by orders of magnitude, but their real-world use requires handling large meshes and limited data. Equivariance provides inductive biases but may harm performance when symmetry is broken by distributional alignment. The study covers automotive aerodynamics and hemodynamics, assessing when equivariance improves generalization.
Key facts
- Neural surrogates accelerate CFD simulations by orders of magnitude.
- Group-equivariant architectures introduce inductive biases.
- Equivariance can be detrimental when symmetry is broken.
- Study covers automotive aerodynamics and blood flow (hemodynamics).
- Published on arXiv with ID 2605.18816.
- Announce type is cross.
- Focus on scalability to large meshes and limited training data.
- Systematic assessment of equivariance across tasks with varying distributional alignment.
Entities
Institutions
- arXiv