Machine Learning Based Flow Simulator: Flow Around an Airfoil with Vortex Generators

Abstract

Controlling the flow structure around an airfoil is crucial for increasing lift and reducing drag. Delaying flow separation improves aerodynamic performance, especially in aircraft and wind turbines. In recent years, artificial intelligence and machine learning methods have emerged as fast and cost-effective alternatives to traditional approaches in fluid mechanics. In this study, we aimed to control the flow around the NACA (National Advisory Committee for Aeronautics) 4412 airfoil using vortex generators (VGs) and to develop a machine-learning-based flow simulator that predicts velocity components based on angle of attack, VG yaw angle, and spatial coordinates. Experimental measurements were conducted in an open-surface, closed-loop water channel at a Reynolds number of Re = 1.0 x 104 using a two-dimensional Particle Image Velocimetry (PIV) system. A total of 60,500 data points were collected per velocity component from 20 experimental cases within the range of alpha = 0 degrees-20 degrees and beta = 15 degrees-30 degrees. A Multilayer Perceptron (MLP) model implemented using TensorFlow was trained to predict the ensemble-averaged (u) and (v) velocity components. We analyzed the effects of hidden layer neuron count and mini-batch size, achieving the highest accuracy with 41 neurons and a batch size of 4, yielding R2 values of 0.978 for (u) and 0.950 for (v). The error distributions were symmetric and closely approximated a Gaussian distribution. Experimental results showed that VGs delayed early-stage flow separation at low alpha but became less effective at higher alpha. The MLP model successfully reconstructed major flow features, providing a reliable data-driven alternative to CFD-based methods. Future work will extend the model to various airfoils, VG designs, Reynolds numbers, and unsteady flows using time-resolved PIV data.