Flow-field Emulation and Shape Optimization of Compressor Airfoils using Design-Variable Hypernetworks

Deep-learning-based flow emulators are used to predict the flow field around parametrically-defined airfoils. These flowemulators are then used in place ofReynolds-AveragedNavier-Stokes (RANS) solvers in design optimization. The flowemulators are based on a) Decoder Convolutional Neural Networks (DCNN), which generate solution snapshots in the computational domain, and b) Design-Variable Hypernetworks (DVH) which provide pointwise predictions in physical space. The flow emulators are used to predict parametric subsonic and transonic compressor flows with the transonic baseline designs corresponding to the NASA rotor 37 at 70% span. The emulators are then deployed in a practical industrial design use case. The first design exploration consists of solely geometric shape parameters, and is later expanded in the transonic case to account for varying rotor speed and thus varying flow conditions. Hypernetwork-based models are seen to outperform DCNN-based models on the transonic problem and are used in place of CFD to drive shape optimization at varying rotor speed. At nominal speed, the emulator-driven optimization achieves the same optimal design as CFD-driven optimization in a reduced number of iterations at a fraction of the online computational cost, while providing similarly-performing designs at off-nominal conditions. These results establish the utility of Design-Variable Hypernetworks as a viable emulation and optimization tool in practical industrial design.