Measuring and improving the geometric accuracy of piece-wise polynomial boundary meshes

Highlights: • A disparity measure to optimize the geometric accuracy of a high-order mesh. • Optimized high-order meshes have geometric accuracy of order 2p. • Optimized high-order meshes approximate normals with order $2p-1$. • Non-interpolative and interpolative meshes feature the same accuracy. We present a new disparity functional to measure and improve the geometric accuracy of a curved high-order mesh that approximates a target geometry model. We have devised the disparity to account for compound models, be independent of the entity parameterization, and allow trimmed entities. The disparity depends on the physical mesh and the auxiliary parametric meshes. Since it is two times differentiable on all these variables, we can minimize it with a second-order method. Its minimization with the parametric meshes as design variables measures the geometric accuracy of a given mesh. Furthermore, the minimization with both the physical and parametric meshes as design variables improves the geometric accuracy of an initial mesh. We have numerical evidence that the obtained meshes converge to the target geometry (unitary normal) algebraically, in terms of the element size, with order 2p ($2p-1$, respectively), where p is the polynomial degree of the mesh. Although we obtain meshes with non-interpolative boundary nodes, we propose a post-process to enforce, if required by the application, meshes with interpolative boundary nodes and featuring the same order of geometric accuracy. In conclusion, we can obtain super-convergent orders, at least for sufficiently smooth parametric curve (surface) entities, for meshes of polynomial degrees up to 4 (3, respectively). In perspective, this super-convergence might enable using a lower polynomial degree to approximate the geometry than to approximate the solution without hampering the required geometric accuracy for high-order analysis.