Enhancement of Low Fidelity Fluid Simulations using Machine Learning

We introduce a machine learning method to estimate discrepancies between low fidelity potential flow and high fidelity Navier-Stokes solution velocity fields. The Scikit-Learn implementation of random forest is used to correlate local flow features of potential flow solutions and deviations from the corresponding Navier-Stokes solutions. The fidelity of potential flow solutions for similar problems can then be enhanced with an additive correction term provided by the machine learned model. The method is demonstrated in an attached flow problem around an elliptical cylinder. A machine learning model trained using a single reference problem at 30 degree angle of attack (AoA) achieves less than 3.5% training error and the prediction of Navier-Stokes velocity fields within 10% over a range of AoAs, between 0 and 45 degrees. Use of multiscale features leads to a slight improvement during the training. The training process using two reference problems suffers from mixing of data and overfitting; feature vectors at a higher dimension and a larger data size may be required to address this issue.