Enhancing web-based CFD post-processing using machine learning and augmented reality

Two approaches for improving web-based post-processing tools for databases of computational fluid dynamics solutions are presented. The first uses Machine Learning to automatically classify the computational results and detect features of interest in the solutions. The second uses Augmented Reality to display three-dimensional models, on a tablet or phone screen, so that they appear as real objects to the user. Two examples of the application of Machine Learning are given. In the first, a convo-lutional neural network (CNN) is used to detect three-dimensional separations (“corner separations”) in the exit flow field of axial compressor blades. The CNN is trained on an artificially generated training set of two-dimensional scalar fields and shown to be accurate when classifying either experimental or computational results. The second example is the use of a variant of the ‘YOLO’ (You Only Look Once) CNN that can detect objects in an image and report their position and size. The network is trained, again using artificially generated data, to detect vortices from two-component, two-dimensional velocity fields. The network is shown to be able to to detect the tip vortex shed by computations of the Onera M6 wing. A multi-user Augmented Reality ‘room’ is demonstrated that provides an interactive, collaborative environment for the analysis of computational results. Selected geometry or flow data can be extracted from the database and rendered in the web browser of a device (phone or tablet) on top of a live feed from the device’s camera. The view of the computation responds to the movement of the device, such that a team of engineers can interrogate the solution as if they were viewing the computed results superimposed on a real object