Response surface and neural network techniques for rocket engine injector optimization

The response surface methodology for rocket engine injector design optimiza tion for which only modest amounts of data may exist is examined. Two main aspects are emphasized: relative performance of quadratic and cubic polynom ial response surfaces and enhancement of the fidelity of the response surfa ce via neural networks. A data set of 45 design points from a semi-empirica l model for a shear coaxial injector element using gaseous oxygen and gaseo us hydrogen propellants is used to formulate response surfaces using quadra tic and cubic polynomials. This original data set is also employed to train a two-layered radial basis neural network (RBNN). The trained network is t hen used to generate additional data to augment the original information av ailable to characterize the design space. Quadratic and cubic polynomials a re again used to generate response surfaces for this RBNN-enhanced data set , The response surfaces resulting from both the original and RBNN-enhanced data sets are compared for accuracy. Whereas the cubic fit is superior to t he quadratic fit for each data set, the RBNN-enhanced data set is capable o f improving the accuracy of the response surface if noticeable errors from polynomial curve fits are encountered. Furthermore, the RBNN-enhanced data set yields more consistent selections of optimal designs between cubic and quadratic polynomials. The techniques de, eloped can be directly applied to injector design and optimization for rocket propulsion.