# FEM Convergence Testing

### Introduction

A fundamental premise of using the finite element procedure is that the body
is sub-divided up into small discrete regions known as finite elements.
These elements defined by nodes and interpolation functions.
Governing equations are written for each element and these elements are
*assembled* into a global matrix.
Loads and constraints are applied and the solution is then determined.
### The Problem

The question that always arises is:
*How small do I need to make the elements before I can trust the
solution?*
### What to do about it...

In general there are no real firm answers on this.
It will be necessary to conduct convergence tests!
By this we mean that you begin with a mesh discretization and then observe
and record the solution.
Now repeat the problem with a finer mesh (i.e. more elements) and then
compare the results with the previous test.
If the results are nearly similar, then the first mesh is probably good
enough for that particular geometry, loading and constraints.
If the results differ by a large amount however, it will be necessary to
try a finer mesh yet.
### The Consequences

Finer meshes come with a cost however: more calculational time and large
memory requirements (both disk and RAM)!
It is desired to find the minimum number of elements that give you a
converged solution.
### Beam Models

For beam models, we actually only need to
define a single element per line unless we are applying a distributed
load on a given frame member.
When point loads are used, specifying more that one element per line will
not change the solution, it will only slow the calculations down.
For simple models it is of no concern, but for a larger model,
it is desired to minimize the number of elements, and thus calculation
time and still obtain the desired accuracy.
### General Models

In general however, it is necessary to conduct convergence tests on your
finite element model to confirm that a fine enough element discretization has
been used.
In a solid mechanics problem, this would be done by creating several models
with different mesh sizes and comparing the resulting deflections and
stresses, for example.
In general, the stresses will converge more slowly than the displacement,
so it is not sufficient to examine the displacement convergence.