Dealing with temperamental scientific models

Finite element models can be temperamental. I have a model at this moment that is producing incredibly accurate results. So good in fact that if they were someone else’s results I might be tempted to wonder if they were “post-processed”.

About an hour ago, the solution (quasi-Newton) was diverging half way through—now it soldiers right through the last solution step. The worst part is I have no idea what changed. I prefer a slightly more deterministic process.

In the words of one of the leaders in my specific niche of nonlinear finite element analysis to me at a recent conference:

NLFEA is an art, not a science.

This means building, calibrating, and validating models requires a lot of TLC. You’ll feel like it should just work, that you shouldn’t have to go through all this tedium. But today—in 2012—this is reality.