Application of artificial neural networks to the design of turbomachinery airfoils

The feasibility of applying artificial neural networks to the aerodynamic design of turbomachinery airfoils is investigated. The design process involves defining a target pressure distribution, computing several flows to adequately populate the design space in the vicinity of the target, training the neural network with this data, and, finding a design that has a pressure distribution that is closest to the target. The last step is carried out using the network as a function evaluator. This design process is tested using an established how simulation procedure, a simple two-layer feedforward network and a conjugate gradient optimization technique, Results are presented for some validation tests as well as a complete design effort where the pressure distribution from a modern Pratt and Whitney turbine was used as a target. These results are very encouraging and clearly warrant further development of the process for full three-dimensional design.