This paper studies the use of adjoint CFD solvers in combination with surrogate modelling in order to reduce the computational cost of the optimization of complex 3D turbomachinery components. The method is applied to a previously optimized counter rotating turbofan, with a shape parameterized by 104 CAD parameters.
Through random changes on the reference design, a small number of design variations are created to serve as training samples for the surrogate models. A steady RANS solver and its discrete adjoint are then used to calculate objective function values and their corresponding sensitivities. Kriging and neural networks are used to build surrogate models from the training data. To study the impact of the additional information provided by the adjoint solver, each model is trained with and without the sensitivity information. The accuracy of the different surrogate model predictions is assessed by comparison against CFD calculations.
The results show a considerable improvement of the fitness function approximation when the sensitivity information is taken into account. Through a gradient based optimization on one of the surrogate models, a design with higher isentropic efficiency at the aerodynamic design point is created. This application demonstrates that the improved surrogate models can be used for design and optimization.