Why MESH convergence in DNVGL – RP – F112 is dangerous​

  • Does not give consequent results
  • Can easily be manipulated without breaking any code rules
  • In several cases give confidence to bad mesh and bad results
  • Adds 0 value and increase runtime
  • Favorites bad engineering by Lower runtime and better results  

Mesh convergence setup from the code:

Non-linear Analysis:
A mesh convergence is considered reached with a convergence error of 5% on the total strain when a halving of the local element size is applied (ref. DNVGL-RP-F112 Appendix A1.1).

ANSYS Setup:

stressman

Refinement Depth of 2 equals halving of the element size.

Start Mesh 1

stressman

Start Mesh 2

stressman

End Mesh 1

stressman

End Mesh 2

stressman

Convergence History

stressman

Final strain value 2.3832e-003

Convergence History

Final strain value 3.7323e-003

57% Calculation error between 1 and 2, but both converged in accordance with DNVGL – RP – F112

What do you do?

•Use common engineering sense
•Ensure that the MESH captures discontinuities
•Run some test and look for a change in results with different MESH
•For piping it is often a good start to have 3 elements across wall thickness and a mesh size of OD X 0.12
Example: This gives a 6” pipe a mesh of 20mm
•Be critical when evaluating results
Share on facebook
Facebook
Share on twitter
Twitter
Share on linkedin
LinkedIn