Design-Variable Hypernetworks for Flowfield Emulation and Shape Optimization of Compressor Airfoils

Abstract

Deep-learning-based flow emulators are used to predict the flowfield around parametrically defined airfoils and then used in place of Reynolds-averaged Navier–Stokes solvers in design optimization. The flow emulators are based on a) decoder convolutional neural networks, which generate solution snapshots in the computational domain, and b) design-variable hypernetworks, which provide pointwise predictions in physical space. The flow emulators are used to predict parametric subsonic and transonic compressor flows in an industrial design use case with baseline geometry corresponding to the NASA rotor 37. Both methods are effective in representing unseen subsonic airfoil flowfields, with mean errors less than 1%. The hypernetwork-based method generalizes more effectively under transonic conditions and is used in place of computational fluid dynamics (CFD) to drive shape optimization at varying rotor speeds. Under transonic conditions and at nominal speed, the emulator-driven optimization achieves the same optimal design as CFD in a reduced number of iterations at a fraction of the online computational cost while providing similarly performing designs at off-nominal conditions. It is remarked that once the emulator is trained once offline, it can be used online to conduct many different design optimizations, e.g., with different objective functions, constraints, and tradeoffs. These results establish the utility of design-variable hypernetworks as a viable emulation and optimization tool in practical industrial design.