Improving the unsteady aerodynamic performance of transonic turbines using neural networks

Summary A recently developed neural net-based aerody-namic design procedure is used in the redesign of atransonic turbine stage to improve its unsteadyaerodynamic performance. The redesign procedureused incorporates the advantages of both tradi-tional response surface methodology (RSM) andneural networks by employing a strategy calledparameter-based partitioning of the design space.Starting from the reference design, a sequence ofresponse surfaces based on both neural networksand polynomial fits are constructed to traverse thedesign space in search of an optimal solution thatexhibits improved unsteady performance. The pro-cedure combines the power of neural networks andthe economy of low-order polynomials (in terms ofnumber of simulations required and network train-ing requirements). A time-accurate, two-dimen-sional, Navier-Stokes solver is used to evaluate thevarious intermediate designs and provide inputs tothe optimization procedure. The procedure yieldeda modified design that improves the aerodynamicperformance through small changes to the refer-ence design geometry. These results demonstratethe capabilities of the neural net-based design pro-cedure, and also show the advantages of includinghigh-fidelity unsteady simulations that capture therelevant flow physics in the design optimizationprocess.A patent application that covers some of the original ideas inthis report has been filed by NASA. This report has been sub-mitted for review toward presentation at the 38th AIAA Aero-space Sciences Meeting and Exhibit, Jan. 10-13, 2000. Reno,NV.*Riverbend Design Services, Palm Beach Gardens. Florida.