A matrix-free isogeometric Galerkin method for Karhunen-Lo\`eve approximation of random fields using tensor product splines, tensor contraction and interpolation based quadrature

TL;DR

This paper introduces a matrix-free, parallel Galerkin method using tensor product splines and interpolation quadrature to efficiently compute the Karhunen-Loève expansion of random fields, overcoming computational challenges of traditional approaches.

Contribution

The work develops a novel matrix-free, parallel approach for solving the eigenvalue problem in KLE using tensor product splines and interpolation quadrature, reducing computational costs.

Findings

Achieves high accuracy with fewer quadrature points.

Demonstrates excellent scalability and robustness in 3D benchmarks.

Significantly reduces computational time and memory requirements.

Abstract

The Karhunen-Lo\`eve series expansion (KLE) decomposes a stochastic process into an infinite series of pairwise uncorrelated random variables and pairwise $L^2$-orthogonal functions. For any given truncation order of the infinite series the basis is optimal in the sense that the total mean squared error is minimized. The orthogonal basis functions are determined as the solution of an eigenvalue problem corresponding to the homogeneous Fredholm integral equation of the second kind, which is computationally challenging for several reasons. Firstly, a Galerkin discretization requires numerical integration over a $2d$ dimensional domain, where $d$, in this work, denotes the spatial dimension. Secondly, the main system matrix of the discretized weak-form is dense. Consequently, the computational complexity of classical finite element formation and assembly procedures as well as the memory requirements of direct solution techniques become quickly computationally intractable with increasing polynomial degree, number of elements and degrees of freedom. The objective of this work is to significantly reduce several of the computational bottlenecks associated with numerical solution of the KLE. We present a matrix-free solution strategy, which is embarrassingly parallel and scales favorably with problem size and polynomial degree. Our approach is based on (1) an interpolation based quadrature that minimizes the required number of quadrature points; (2) an inexpensive reformulation of the generalized eigenvalue problem into a standard eigenvalue problem; and (3) a matrix-free and parallel matrix-vector product for iterative eigenvalue solvers. Two higher-order three-dimensional benchmarks illustrate exceptional computational performance combined with high accuracy and robustness.