Real-time stress analysis of three-dimensional boundary element problems with continuously updating geometry

Abstract

Computational design of mechanical components is an iterative process that involves multiple stress analysis runs; this can be time consuming and expensive. Significant improvements in the efficiency of this process can be made by increasing the level of interactivity. One approach is through real-time re-analysis of models with continuously updating geometry. In this work the boundary element method is used to realise this vision. Three primary areas need to be considered to accelerate the re-solution of boundary element problems. These are re-meshing the model, updating the boundary element system of equations and re-solution of the system.

Once the initial model has been constructed and solved, the user may apply geometric perturbations to parts of the model. A new re-meshing algorithm accommodates these changes in geometry whilst retaining as much of the existing mesh as possible. This allows the majority of the previous boundary element system of equations to be re-used for the new analysis.

Efficiency is achieved during re-integration by applying a reusable intrinsic sample point (RISP) integration scheme with a 64-bit single precision code. Parts of the boundary element system that have not been updated are retained by the re-analysis and integrals that multiply zero boundary conditions are suppressed. For models with fewer than 10,000 degrees of freedom, the re-integration algorithm performs up to five times faster than a standard integration scheme with less than 0.15% reduction in the L_2-norm accuracy of the solution vector. The method parallelises easily and an additional six times speed-up can be achieved on eight processors over the serial implementation.

The performance of a range of direct, iterative and reduction based linear solvers have been compared for solving the boundary element system with the iterative generalised minimal residual (GMRES) solver providing the fastest convergence rate and the most accurate result. Further time savings are made by preconditioning the updated system with the LU decomposition of the original system. Using these techniques, near real-time analysis can be achieved for three-dimensional simulations; for two-dimensional models such real-time performance has already been demonstrated.