Complex Nozzle Optimization Techniques Using Machine Learning

The desire to develop faster, stealthy next-generation aircraft has led to the use of complex nozzles for aircraft integration. These nozzles contain multiple high-velocity streams that exit from non-axisymmetric areas. Due to the complexity of these nozzles, acoustic experiments and simulations are computationally and time expensive, making them not ideal for the design process. This study employs an artificial neural network (ANN) as a tool for rapid noise pre- diction of far-field acoustics based on geometric and flow parameters. The ANN is configured to predict the input features that minimize resulting far-field noise. To achieve this, the optimization strategies Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) were used. The model was restricted to a third stream nozzle pressure ratio (NPR3) of 1.89 due to a low noise bucket that has been shown to occur, which correlates to perfectly expanded flow leaving the third stream. For this constraint, the model predicted that the lowest noise configuration occurs with a rectangular deck plate which extends as far as allowed by the network. With ESPL predictions within .15 dB of measured values, this study shows that an ANN can be utilized as a rapid noise prediction tool in the design process to determine low noise configurations when provided flow and geometry parameters of a supersonic jet nozzle.