Steerable Implicit Neural Representations for Rapid Design Optimization in Fluid Systems

We present a data-driven framework for accelerating fluid dynamics–based design tasks across diverse geometry classes, including airfoils and mixing chambers. Our approach begins by generating high-fidelity simulation datasets tailored to each geometry class. We then train steerable implicit neural representations (INRs) that learn continuous mappings from geometry and flow parameters to physical quantities of interest, such as pressure, velocity, or lift and drag coefficients. These models enable rapid, differentiable inference of fluid behavior without repeated CFD solves. By embedding these INRs into optimization pipelines, we demonstrate efficient, accurate design iterations that maintain fidelity while significantly reducing computational cost. This work opens the door to interactive, physics-informed design tools across a wide range of fluid engineering applications.