EXPLORING PHYSICS-INFORMED MACHINE LEARNING FOR A CYLINDER IN CROSS FLOW

A cylinder in cross flow has been studied by many prior authors as a test case for computational fluid dynamics (CFD). We modified an existing case study for a cylinder in cross flow in an open source CFD program. Once the CFD model was working, we developed a physics-informed machine learning version for benchmarking. The machine learning code was validated using two existing sets of data from the literature for other systems. Our research question is focused on understanding how well a physics-informed neural net for the Navier Stokes equations represent the behavior of complex flow around a cylinder. After the neural net model was validated, we used data from the CFD tool for a cylinder in cross flow to test the performance of the neural net. We compared velocities and stream functions to the original CFD solution to assess performance. The results indicate the physics-informed machine learning model is computationally efficient and accurate for predicting the basic flow shapes. The number of training data points is very important for predicting major flow behavior. The machine learning model did not perform well for transfer learning in our study, failing to predict the key wakes. Future work will confirm how responsive the machine learning model is to variations in the input data, and different methods of providing a sparse training set.