Uncertainty Quantification by Physics- and Data-Based Models for Mechanical Systems

Due to resource-limitations in the context of uncertainty quantification (UQ), full simulation models are not applicable. This is particularly true in the target application context of multi-phase flow in porous media or multiphysics muscle-modelling. Hence sophisticated surrogate modelling strategies are indispensable. This project has an accompanying project PN 6-2, which is requested in PN6 and is aiming at machine learning based surrogates. In contrast, the PN5 project will be aiming at adaptive reduction techniques that enable automatic adaption to external resource limitations (runtime, accuracy) and respect physical constraints. We will focus on projection-based model order reduction (MOR) and sparse grid (SG) approximation. We will consider parametric models that can both be treated in a deterministic as well as a stochastic context, and hence enable uncertainty quantification (UQ). In addition to parametric uncertainty, also the surrogate models themselves have inherent inaccuracy, and we will aim at controlling those errors by deterministic error bounds. In order to improve over UQ predictions by the reduced models, the learning results from the PN6 partner project will enable multifidelity strategies.