Adapting turbofan noise modelling tool using neural networks

Abstract

System-level multidisciplinary design and optimization (MDO) of turbofans, incorporating integrated noise and emissions predictions, can require substantial computational resources. Noise assessments typically consume significantly more compared to 1-dimensional simulations of turbofan performance and correlation-based emissions predictions, making them the bottleneck of the entire process. In this study, the use of neural networks in adapting an existing semi-empirical noise modelling tool for MDO applications is presented. The aim is to reduce the computation load while keeping the accuracy of the predictions sufficiently close to the original models. Deep neural networks (DNN) and its combination with K-nearest neighbors (KNN) make up stacking model have been applied for the purpose. The dataset for neural network training comes from the noise model developed by Chalmers Noise Code (CHOICE), an open-source framework with the capability to predict the source noise level, from individual airframe, engine components and the entire aircraft. For the aircraft approach phase, the neural network selects the important design parameters at the engine design point as inputs and outputs the noise of each engine component. Overall, the practical use of neural networks proves beneficial which could achieve noise prediction quickly and efficiently with high accuracy. The combined use of DNN and KNN could improve the accuracy of the trained models significantly.