Do not average unfilled material properties.

[gassmoeller]

I found a bug in the average_property function that would cause floating point exceptions if a material output property is not filled. Our current convention for unfilled output properties is no longer that they are set to size 0, but that they are left unfilled at their default values NaN. I left the old condition untouched so that in case a user material model sets the size to 0 it still works the same way as before. I also moved the condition behind the disabling of the floating point exceptions, because isnan(NaN) seems to sometimes throw floating point exceptions (at least it did for me) even though as far as I understand it shouldnt.

[jdannberg]

Thank you for fixing this! This works for me, and I made a pull request to your branch adding a simple test case (that crashed for me on the current main branch). The problem here was actually the GMG solver.

I don't know if it would be worth investigating in more detail why that averages viscosity values that have not been filled by the material model in the Stokes assembly.

[jdannberg]

Suggested change

	all_entries_NaN(const std::vector<double> &values)
	all_entries_NaN (const std::vector<double> &values)

[gassmoeller]

I don't know if it would be worth investigating in more detail why that averages viscosity values that have not been filled by the material model in the Stokes assembly.

I am pretty sure that is because the GMG computes only the right-hand side of the Stokes equation in the assembly step. The viscosity is evaluated separately. Therefore it doesnt request the viscosity from the material model.

I merged your test and updated the results to pass. This should be ready for another look.

[jdannberg]

Oh, I see. That makes sense then!