Non-linear Finite Element Structural Analysis

Introduction

In most professional structural design practice, prediction of the response of structures during an earthquake is carried out by the analysis of simple linear elastic finite element models and quite a bit of "hand waving" to predict the non-linear response from linear elastic results.  Linear elastic models are used because they are computationally inexpensive and most importantly, they are fast.  However, exact non-linear finite element structural analysis methods have been around for years, but have long been considered merely an academic interest because they are notoriously slow and computationally intensive.  With the advent of parallel computing and the ever increasing speed of processors, exact non-linear analyses will no longer be considered only an academic interest.
 

A Terse Treatment of Non-linear Structural Analysis

At a given load step in the linear elastic analysis of a structure, the state determination for the structural elements is trivial and does not require iteration to satisfy any form of equilibrium.  For a load step in a non-linear analysis, there are three levels of iteration required.  Iteration is required to ensure global, or overall structural equilibrium, i.e. the balance between applied loads and internal resistance.  This internal resistance is provided by the elements, where a second level of iteration is required to ensure that the loads transferred to the elements are equilibrated by the elements' resisting forces.  These element resisiting forces are provided by the elements' constitutive modeling, for example, steel or reinforced concrete.  The constitutive model provides a third level iteration where the internal element deformations are required to satisfy a prescribed stress-strain relationship.
 

Why Parallelize?

In forming the global equations of equilibrium for a structure, looping over all elements in the finite element model is required.  This looping is done so that elements can be queried for their tangent stiffness and their resisting forces.  In a sequential computing environment, elements are queried one by one for their tangent stiffness, which is then assembled into the global system of equilibrium equations.  After the equations have been solved, the elements are asked to provide, one by one, their resisting forces.  This process of forming tangent stiffness, solving equations, and forming resisting forces is repeated until global equilibrium is satisfied for the current load step.  In an earthquake analysis, there will be thousands of load steps, so in a sequential computing environment, this analysis will take quite some time, even with current computer processing capabilities.  By parallelizing the element operations of forming tangent stiffness and resisting forces, run-time will be significantly reduced.
 

Other Areas of Interest

Dynamic load balancing of a non-linear finite element analysis is required.  As elements begin to experience non-linear response, the amount of computation required to form tangent stiffness and resisting forces increase as more iterations will be required to satisfy equilibrium.  During an earthquake analysis, it is not known before which elements will exhibit non-linear behavior, therefore, an efficient form of dynamic load balancing is required to maintain equity among processors.  In addition, substructure domain decomposition techniques may improve data locality and take advantage of architecture specific memory features.

Current Work

This subject has been studied rather intensively by many researchers.  Significant work in parallelizing the finite element method for structural analysis is ongoing here at the University of California under the direction of  Dr. Frank McKenna.
 

References

Cook, Robert D., David S. Malkus and Michael E. Plesha, Concepts and Applications of Finite Element Analysis, Third Edition, John Wiley & Sons, 1989.

McKenna, Francis, Object-Oriented Finite Element Programming: Frameworks for Analysis, Algorthms, and Parallel Computing, Ph.D. Thesis, University of California, 1997.


Michael H. Scott
mhscott@ce.berkeley.edu

January 26, 2000