Deep learning based streamline visualization method on the 3D flow data

Abstract

Flow visualization is a subfield of data visualization, covering various science and engineering fields. It transforms the flow data defined by the velocity vector field into a meaningful visual representation. Even though there are many visualization methods for flow data, streamline is one of the most frequently utilized methods. Streamline is a geometric flow visualization technique that reveals the instantaneous particle trajectory along the velocity field. It describes global flow patterns to be able to understand flow streams intuitively. However, it is challenging to determine the seed locations generating streamlines that extract essential flow features. This thesis deals with the problem mentioned above using a deep learning model that learns global flow patterns and predicts the importance of streamlines across every grid point in the flow field. We presented a semi-automatic streamline generation method that reveals important flow features using the proposed deep learning model as well. More specifically, we proposed a deep regression model that learns global flows in the 3D velocity field to predict not only the importance score but also the spatial and shape properties of the streamlines across the entire flow field. Moreover, we presented the proposed model-based user interface, which shows the predicted information visually to extract important streamlines semi-automatically among all available streamlines. The proposed method was evaluated quantitatively and qualitatively by applying into 3D CFD data sets and performing an expert review.