Generic Component-Based Software Architecture for the Simulation of Probabilistic ModelsA

Abstract

An uncertain behaviour is in the nature of many physical phenomena. This uncertainty has to be quantified for a meaningful prediction by a computer-aided simulation. A stochastic description of the uncertainty carries a physical phenomenon over to a probabilistic model, which is usually solved by numerical schemes. The present thesis discusses and develops models for challenging uncertain physical phenomena, efficient numerical schemes for a quantification of uncertainties (UQ), and a sustainable and efficient software implementation. Probabilistic models are often described by stochastic partial differential equations (SPDEs). The stochastic Galerkin method represents the solution of an SPDE by a set of stochastic basis polynomials. A problem-independent choice of basis polynomials typically limits the application to relatively small maximum polynomial degrees. Moreover, many coefficients have to be computed and stored. In this thesis new error-controlled low-rank schemes are presented, which in addition select relevant basis polynomials. In this manner the previously mentioned problems are addressed. The complexity of a UQ is as well reflected in the software implementation. A sustainable implementation relies on a reuse of software. Here, a software architecture for the simulation of probabilistic models is presented, which is based on distributed generic components. Many of these components are reused in different frameworks (and may also be used beyond a UQ). They can be instantiated in a distributed system many times and are interchangeable at runtime, where the generic aspect is preserved. Probabilistic models are derived and simulated in this thesis, which for instance describe uncertainties for a composite material and an aircraft design. Among other things, several hundred stochastic dimensions or a long runtime for simulations arise.