Field Inversion Machine Learning Based Stochastic Noise Generation Model for Jet Noise Prediction

Predicting aerodynamic noise generated by turbulent flow is a challenging and expensive engineering task due to its unsteady nature. While some closure models have been developed for use in linearized equations to reduce the computational cost of aeroacoustic analyses, obtaining more accurate sound predictions still requires access to instantaneous turbulent velocity data. The stochastic Noise Generation (SNG) method provides a cost-effective way to generate synthetic turbulent velocity fields, but it faces challenges due to its high sensitivity to the model constants and limitations in accurately predicting high-frequency fluctuations. In this study, we propose to enhance the predictive capability of the SNG model by leveraging the Field Inversion Machine Learning (FIML) technique, a form of data assimilation. The inversion process of FIML is divided into two phases. The first phase aims to recover time-averaged turbulence statistics across the entire space by optimizing the spatial correction coefficients, while the second phase focuses on correcting time-dependent data by optimizing modal correction coefficients. Two corresponding machine learning models are trained to predict spatial and modal corrections by using 15 jet flow cases with varying inlet conditions. These trained models are subsequently tested on both seen and unseen flow cases. Results indicate that the machine learning-enhanced SNG model significantly outperforms the baseline model in terms of accuracy for both time-averaged and time-dependent turbulent flow statistics. The recovery of time-dependent flow statistics, as employed in this study, represents a novel application of FIML with a modal correction approach. Additionally, the utilization of a machine learning model to predict modal corrections from a steady solution is also innovative in this context.