Problem: Cluster a large, dense, three-dimensional CFD simulation result into homogeneous areas that can be used for graph based simulation and visualization.
My work included custom pre-processing, the research and implementation of existing algorithms, as well as the development of a new algorithm. Due to the research-typical, open formulation of the problem statement, there is no closed, complete solution of the problem. Nevertheless I pin-pointed the best existing clustering algorithm for the graph based simulation. In addition I developed a useful graph based visualization technique for exploring CFD datasets that utilizes my own clustering algorithm.
The abstract of my thesis captures the essence of my solution. The full thesis can be downloaded at the end of this post.
Results from computational fluid dynamics (CFD) simulations are generally complex and difficult to understand.
This work proposes a new method that computes from a given simulation result, e.g., the underhood flow of air around a car engine, a sparse directed graph network with a few hundred nodes.
The goal is a graph that preserves the essential properties of the flow in such way that it is suitable for applications ranging from information visualization to flow simulation.
The algorithm finds bundles of similar streamline segments, which are then mapped back to the original dataset in order to produce a complete partition.
A flow graph is derived from this partition by integration over the CFD cells.
By utilizing a custom-built simulation framework, the proposed method is shown to produce meaningful graphs, which can be used within the mentioned application areas.