Latent Space Modelling of Unsteady Flow Subdomains

Abstract

Very complex flows can be expensive to compute using current CFD techniques. In this thesis, models based on deep learning were used to replace certain parts of the flow domain, with the objective of replacing well-known regions with simplified models to increase efficiency. To keep the error produced by the deep learning model bounded, a traditional CFD model and deep learning model were coupled using a boundary overlap area. In this overlap area, the flow computed by the traditional CFD model was used by the deep learning model as an input. It was demonstrated that since traditional CFD model continuously feeds in reliable information into the deep learning domain, the error remains bounded. Furthermore, it was found that the accuracy of the deep learning models depends significantly on the random initial weights. Therefore, deep learning models trained differently must be carefully compared.