A Parameterized Prediction Method for Turbulent Jet Noise based on Physics-informed Neural Networks

Abstract Turbulent noise prediction is integral to fluid equipment design, and multiple simulations or experiments are often required for noise distribution under varying operating conditions during the design optimization process, which could be expensive. Recently, the Physics Informed Neural Networks (PINNs) method has emerged as an efficient machine learning method for solving parameterized partial differential equations with geometric shapes, boundary conditions, and equation parameters as variable parameters through a single training session without data. In this study, a parameterized prediction method is developed to predict turbulent jet noise based on PINNs without any training datasets. Both two-dimensional (2D) and three-dimensional (3D) jet flow problems are solved. The 2D problem is solved with the Reynolds number as a variable parameter, and the 3D problem is solved with the Reynolds number and nozzle eccentricity as variable parameters. The predicted results are in good agreement with those from conventional computational fluid dynamics (CFD), with average errors of 3% and 6% for the 2D and 3D flow and acoustic power fields, respectively. In terms of computational efficiency, the time required by the method for the three-dimensional problem with two variable parameters is only one-seventh of that of the traditional CFD method. This study demonstrates that for engineering noise scenarios with varying parameters, the method based on PINNs offers a more efficient parameterized predicting approach and is promising for future applications.