In neural networks, Delaunay triangulation is used to construct a graph data structure, where points (e.g., point cloud elements or feature vectors) become the vertices of the graph, and edges are determined from the triangulation results.