Machine learning approaches for repositories of numerical simulation results

Simulations are used intensively in the developing process of new industrial products and have achieved a high degree of detail. In that workflow often up to thousand finite element model variants, representing different product configurations, are simulated within a few days. Currently the decision process for finding the optimal product parameters involves the comparative evaluation of large finite element simulation bundles by post-processing each one of those results using 3D visualization software. This time consuming process creates a severe bottleneck in the product design and evaluation workflow. To handle the data we investigate an analysis approach based on nonlinear dimensionality reduction to find a low dimensional parametrization of the data. A core functionality of the approach is the determination of a suitable distance measure between numerical simulation result taking the geometrical deformation of the discretization grid into account. In the obtained lower dimensional representation, similar model variants are organized in clusters and for example input or output variables can be investigated along such a parametrization. We demonstrate the application of this approach to a realistic and relevant industrial example for robustness analysis of the bumper location in a frontal crash simulation.