Intelligent Prediction of Separated Flow Dynamics using Machine Learning

Abstract

Understanding separated flow dynamics is crucial for implementing effective flow control techniques. These techniques help mitigate adverse effects on vehicle performance and environmental pollution. This research aims to improve flow control strategies by predicting separated flow dynamics solely through wall pressure measurements using artificial intelligence and numerical data. Initially, we identify numerical models that accurately replicate separated flow dynamics. Notably, the Detached Eddy Simulation (DES) model strongly agrees with experimental data, particularly in the turbulent regime at Reh= 89100, downstream of backward facing steps (BFS). Subsequently we conducted a correlational analysis that revealed a significant relationship between various wall pressure points and the velocity field, leading to the adoption of deep learning techniques such as Recurrent Neural Networks with Long Short-Term Memory (LSTM). These neural networks, tailored for time-dependent data, demonstrate high accuracy of low MSE of 13.48% using ten wall pressure points in predicting velocity magnitude contour over (BFS). To enhance predictions, Proper Orthogonal Decomposition (POD) is utilized to reduce system complexity while retaining essential dynamics, resulting in a lower MSE of 5.07%. Additionally, we identify the ideal wall pressure measurement region that accurately captures the entire dynamic behavior, achieving an acceptable MSE of 23.48% for predicting low order vorticity, with only three wall pressure points. This research aids in developing efficient flow control strategies with limited pressure data and offers valuable insights for closed-loop flow control applications.

Keywords

• flow separation
• backward facing step
• instabilities
• lstm recurrent neural network
• dynamics prediction
• machine learning