Surface design based upon a combined mesh
The subject of this talk is the problem of surface design based upon a mesh that may contain both triangular and quadrangular domains. We investigate the cases when such a combined mesh occurs more preferable for bivariate data interpolation than a pure triangulation. First we describe a modification of the well-known flipping algorithm that constructs a locally optimal combined mesh with a predefined quality criterion. Then we introduce two quality measures for triangular and quadrangular domains and present the results of a computational experiment that compares integral interpolation errors and errors in gradients caused by the piecewise surface models produced by the flipping algorithm with the introduced quality measures. The experiment shows that triangular meshes with the Delaunay quality measure provide better interpolation accuracy only if the interpolated function is strictly convex, as well as a saddle-shaped function is better interpolated by bilinear patches within a combined mesh. For a randomly shaped function combined meshes demonstrate smaller error values and better stability in compare with pure triangulations. At the end we consider other resources for mesh improvement, such as excluding >bad< points from the input set for the mesh generating procedure. Because the function values at these points should not be lost, some linear or bilinear patches are replaced by nonlinear patches that pass through the excluded points.