Finite elements for higher-order problems in isogeometric analysis and statistical inference
Repository URI
Repository DOI
Change log
Authors
Abstract
Despite recent advances, there are still challenges in solving higher-order partial differential equations using the finite element method. This thesis focuses on finite element techniques for higher-order problems with applications to isogeometric analysis and statistical inference. In both areas, the efficient solution of higher-order partial differential equations is essential.
The construction of isogeometric smooth spline basis functions is crucial for solving higher-order partial differential equations on meshes with arbitrary topologies. We introduce a simple blending approach that produces smooth blended B-splines, referred to as SB-splines, on unstructured quadrilateral and hexahedral meshes. We first define a set of mixed smoothness quadratic B-splines that are
The statistical finite element method coherently sythesises data and finite element models using Bayesian inference. However, the Gaussian process inference using covariance matrices is computationally expensive and memory-intensive. Consequently, the size of the finite element model is significantly limited. To this end, we introduce a sparse precision formulation by specifying the Gaussian process priors as Matérn random fields. Instead of kernel parametrisation, the Matérn random fields are parametrised as solutions of a higher-order stochastic partial differential equation. Finite element discretisation of the corresponding weak form results in sparse precision matrices. Specifically, the sparse precision matrices for arbitrary Matérn smoothness are obtained using recursive finite element computation and rational approximation. In addition to achieving improved scalability, we demonstrate the extension of the stochastic partial differential equation approach to non-Euclidean domains and the modelling of non-stationary and anisotropic Matérn random fields.