Change material model logic for determining initial time step number

[naliboff]

Related to #2362.

@bangerth - Let me know if this is what you had in mind. A similar logic was already in drucker_prager.

diffusion_dislocation did not have a reference to the tilmestep at all in setting the strain rate during time step 0 or initiation stage. Rather, it just checks in.strain_rate.size() before calling a function to compute the viscosity.

[tjhei]

Can we create a function in SimulatorAccess instead? Maybe something like is_initial_time_step()?

[bangerth]

Exactly what I had in mind. Unfortunately, this now fails several tests -- but I can't see why because @tjhei 's server doesn't want us to see the output diffs :-(

[bangerth]

There is also a merge conflict you'll have to address.

[naliboff]

@tjhei - I think having that something like is is_initial_time_step() is a good idea. Would this apply to only after the model setup or during the first set of set of solves (e.g., time step "0" now)? Until then, any objections to the logic introduced here?

@bangerth - I think I know what the issue is with the tests. I'll rebase and see if that fixes the problem.

[naliboff]

Ok, got it down to two tests failing. @tjhei - When you have a chance, can you take a look at the tester permissions? The tests diffs are currently not accessible.

[tjhei]

The tests diffs are currently not accessible.

that means that the upload to the webserver hasn't completed yet. It is available now but it might take up to an hour to complete.

[bangerth]

On 06/28/2018 12:29 PM, Timo Heister wrote: Can we create a function in SimulatorAccess instead? Maybe something like is_initial_time_step()?

It's not the initial time step, but instead *before* the initial time step. I could definitely store a flag for this, as well as an accessor function. Would you like that?

[gassmoeller]

I do not think we need to store a separate flag, we could just move the logic that is already present in many plugins into a separate function SimulatorAccess::is_during_initialization would then check for timestep_number == numbers::invalid_unsigned_int. I am sure this is used in a number of plugins (at least in some of my own that are not in master).

[tjhei]

I do not think we need to store a separate flag

agreed.

[bangerth]

@naliboff and @anne-glerum: We discussed this before :-) I just looked and there is no such is_during_initialization() function. I'll add one next.

[bangerth]

@naliboff -- #2972 was just merged. Want to give that a try instead of the approach you have here?

[naliboff]

@bangerth 👍

[naliboff]

@MFraters - The visco_plastic Newton derivative tests are failing with the proposed changes. Can we take a look at this together at some point?

[MFraters]

hmm, this is interesting. So I think the reason it is failing has to do with how the test is set up. For the test I needed to trick the material model in using not the reference strain-rate, but the strain-rate which I provided in the test. So we need a way to set this->simulator_is_past_initialization() to true for the test.

[naliboff]

oofta ;) Is the strain-rate in the test constant throughout the model domain?

[MFraters]

no, I have a different strain-rate tensor for each 'quadrature' point.

[MFraters]

hmm, I will have to look into this a bit more tomorrow. Adding a function to set the variable simulator_is_past_initialization through simulator access wont work because it only has a const pointer to simulator (a pointer which can't change the simulator class). So it would have to be a function directly in the simulator class which I would somehow have to get access to in the file tests/visco_plastic_derivatives_2d.cc. This feels quite intrusive, so I do need to think about it a bit more. Help with this problem is greatly appreciated.

[bangerth]

Can you explain why the test now fails? Are you saying that the branch you end up in in the computation of e_dot_ii is wrong?

[MFraters]

yes, it should compute second invariant of the strain-rate tensor instead of using the reference strain-rate.

[naliboff]

@MFraters - Any chance we can come back and take a look at how to fix these tests?

As I understand it, the issue is that current logic in visco_plastic.cc allows one to calculate the standard strain rate invariant prior to initialization. With the change in this PR, we not only use the reference strain rate on the first iteration of time step 0, but now also before initialization is finished. This then breaks the derivative tests as it passes visco_plastic a strain rate tensor prior to initialization.

Could we somehow trick simulator_is_past_initialization() to return true just for this test?

[naliboff]

Closing this as there does not seem to be an easy path for adjusting the Newton test setup.

[MFraters]

ah, sorry. I totally missed your last reply. If you still want to discuss this at some point or want me to have a look at it again let me know.