ePubs

The open archive for STFC research publications

Full Record Details

Persistent URL http://purl.org/net/epubs/work/42971
Record Status Checked
Record Id 42971
Title On the efficient solution of mixed finite element equations in geometrically multiscale thermal stress analysis
Contributors
Abstract In order to identify the best technique to solve poorly scaled systems of linear equations arising in a primal-mixed finite-element approach (FEMIX) in eometrically multiscale thermoelasticity, we examine the combination of HSL direct sparse solvers and matrix scaling routines. The criteria for optimality were robustness, accuracy and execution time. It will be shown that the present approach FEMIX-HSL enables the reliable solution of finite-element model problems where finite elements can differ in size by several orders of magnitude. In addition, it will be shown that by use of the HSL MA57 sparse solver and one of the scaling routines NC64 or MC30 during the factorization process, the execution time is at least two orders of magnitude faster than using a previous solver based on simple Gaussian elimination and that the accuracy of the solution is maintained even if the system matrix is poorly scaled. A number of pathological tests in elasticity and thermoelasticity are examined to test the robustness and execution time of the solvers. Model problems are examined in nanoindentation and microsized coating, where accurate and straightforward calculation of stress in the solid body and on surfaces of material discontinuity is of the utmost importance. The numerical experiments are reformed on a standard PC computing platform.
Organisation CSE , STFC
Keywords solid mechanics , time efficient , mixed finite element , thermo-mechanical coupling , multifrontal method , indefinite , multifield , multiscale , HSL , sparse , scaling
Funding Information
Related Research Object(s):
Language English (EN)
Type Details URI(s) Local file(s) Year
Report RAL Technical Reports RAL-TR-2008-009. 2008. dumiRAL2008009.pdf 2008
Showing record 1 of 1