feat: averaging of stress components using quadrature

[ryar9534]

Implementation of a new averaged stress output per cell, where the averaging for each component is done using the quadrature rule of the finite element.

The new field is averageStress (6 components per cell) and is registered/owned by the solid mechanics solver, like the average strain (which I have made some minor changes to in order to unify with this new stress). This average stress is stored in addition to the stress at the quadrature points, but the stress owned by the constitutive law is no longer output (using the arithmetic mean).

An annoying rebaseline will be required, as this is gonna change all of the mechanics/poromechanics.

[CusiniM]

we should make the stress NO_PLOT.

[ryar9534]

we should make the stress NO_PLOT.

The new average stress or the old stresses?

[CusiniM]

we should make the stress NO_PLOT.

The new average stress or the old stresses?

the old ones should disappear I guess.

[ryar9534]

we should make the stress NO_PLOT.

The new average stress or the old stresses?

the old ones should disappear I guess.

So are we ok with the overall increased memory footprint? The advantage of the old stresses is that the average is never stored explicitly

[CusiniM]

we should make the stress NO_PLOT.

The new average stress or the old stresses?

the old ones should disappear I guess.

So are we ok with the overall increased memory footprint? The advantage of the old stresses is that the average is never stored explicitly

well, yes, but it's also wrong and we certainly don't want the output to depend on FE.

[codecov]

Codecov Report

Attention: Patch coverage is 5.55556% with 17 lines in your changes missing coverage. Please review.

Project coverage is 56.45%. Comparing base (42966e9) to head (6a5f2a7).

Files	Patch %	Lines
...ers/solidMechanics/SolidMechanicsLagrangianFEM.cpp	0.00%	9 Missing ⚠️
...idMechanics/kernels/StressStrainAverageKernels.hpp	0.00%	6 Missing ⚠️
...icsSolvers/solidMechanics/SolidMechanicsFields.hpp	0.00%	2 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #3279      +/-   ##
===========================================
- Coverage    56.46%   56.45%   -0.01%     
===========================================
  Files         1059     1059              
  Lines        89458    89468      +10     
===========================================
- Hits         50512    50510       -2     
- Misses       38946    38958      +12     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[castelletto1]

This file contains two functions quadraturePointKernel once for the strain and one for the stress. In my opinion only one function should exist to perform the averaging of a second order symmetric stress tensor -- and probably of any scalar (rank 0), vector (rank 1), ... . The class AverageOverQuadraturePointsBase should be responsible for this task, and the quantity to be averaged should either provided directly or with a lambda function that evaluates such property.

[CusiniM]

we could generalize it and come up with a single kernel. We could try to make it accept a list of views instead of a specific one. It would also mean that we do all averaging in a single kernel launch which it is indeed better. We could use something similar to the elementViewAccessors and create an array1d of all scalars, one for all vectors and one for all tensors to be averaged and have the averagingKernel only ask for the keys of the fields that need to be averaged.
The strain probably remains a bit of a special case though and i don't know that we do any averaging for other quantities.

[castelletto1]

In the long term, such computations should be managed by a numerical quadrature library (i.e. --> Shiva)

[castelletto1]

Here we divide by the cell element volume (m_elementVolume[k]) to compute the increment to the average value. In AverageOverQuadraturePointsBase the cell element contribution is directly included in the weight definition. Both correct, but it would be nice to be consistent

[CusiniM]

IMO we should divide only once at the end.

[castelletto1]

Generally speaking, I would argue that m_elementVolume[k] should not even be required and computed directly in the quadrature process. For this specific case:

real64 cellVolume{0.0}
for( int icomp = 0; icomp < 6; ++icomp )
{
  m_avgStress[k][icomp] += detJxW*m_stress[k][q][icomp];
  cellVolume += detJxW;
}
m_avgStress[k][icomp] /= cellVolume;

[CusiniM]

what's the status of this PR?

[ryar9534]

what's the status of this PR?

@CusiniM I just didnt know if this is still something folks want in develop. I originally did it to see if it helped with some stress oscillations in NL but it didnt.

[rrsettgast]

@herve-gross @jhuang2601
Can we figure out if this is needed?