How realistic are FEA tests?

FEA test accuracy depends on too many things:

- The Physical properties of the material defined in the software, re-check them. You would be surprised how incomplete some material libraries are, I guess the software manufacturers don't want to take the responsability of defining resistance properties.

- The precision of the mesh, and how it is generated, especially around holes, corners and small details.

- Are the models meshed as shell or blocks?

- How are the contacts, constraints and loads defined, are they in accordance to the real scenery?

- Are loads applied in a single point or distributed in a specific area?

- Are loads defined as stress or pressure?

- Does your software calculates only elastic deformation analysis or can it go post elastic?

- …

I think that the secret for accuracy is to try experiment with multiple settings and if possible, using more than one software, results will converge to the more accurate values.

RF response was quite complete. The last item on his list is fairly important in many dynamic situations.

The setup is often as important as the software.

Here's a start: Do a simplified, real-world static test. Evaluate the results and then reproduce the expirament in your software, and review the outcome.
If your setup is really good, and the results are way off, there is likely a problem. Verify your material with other sources as well.

Results from FEA/CFD is always discutable. Be ready to make a live part and test it if a part has important role in machinery. Law and lawyers is more expensive from producting a real model . In life on product , appearing forces is not always predictable and this is a main reason why we used a safety factor. Estimated errors is between 1 to 25 % and most works in FEA is based to predict a errors between real an imaginery results. FEA results is always virtual results.