SST

[DavidAnderegg]

Purpose

I took the turbfix branch, updated it to master and started working on SST.

The biggest Problem with differentiating SST was the blending function f1 as it had some mpi-communication inside. I managed to get rid of this communication by extending d2wall to both layers of halo cells.

This is not production ready, but a bit of progress has been made. The following list explains what works and what doesn't in a random order:

• forward partials seem to be correct

• backwards partials seem to be correct apart from nodal forces

• backwards_fast partials are all over the place

• ANK/NK does work with ADPC, but early on DADI is more efficient

• ANK/NK does not really work using FDPC, i suspect SST is highly non-linear and maybe the step is poorly chosen for this model. Not sure

• Physicality check for SST-ANK probably needs adjustment as i sometimes encountered NaNs

• Adjoint using forward routines does converge and aerodynamic derivatives seem to be correct, the geometric ones are not

• Adjoint using backwards_fast routines does converge, but the derivatives are very wrong

• There is probably a bug in the backwards routines for the d2wall extension I introduced

Grid Convergence
I also run a grid-convergence for the 2D bump testcase here:

• The offset is probably because I used strain for turbulence production instead of vorticity

Solver Convergence
Convergence plot for a NACA0012 testcase:

• modified is the current state and base was the state before
• the first 6 orders of magnitude were converged using ANK with DADI for turbulence, after that, CSANK with ADPC did the rest.

Gradients
Relative tolerance of gradients (d Cl / d x) compared to CS with a stepsize of 1e-200 using the tutorial wing

• I am not sure why a step size of 1e-40 is so different to 1e-200. I think it is because SST appears to be highly-nonlinear
• I do consider the aerodynamic gradients obtained with the forward routines as correct although the difference to CS is quite big. I think this is also explainable by the high non-linearity. Tuning some options might also improve this.
• The number of cpus has an impact for geometric derivatives. This i why I think there is a bug in the wall-distance changes I introduced.

Expected time until merged

I dont know.

Type of change

• Bugfix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (non-backwards-compatible fix or feature)
• Code style update (formatting, renaming)
• Refactoring (no functional changes, no API changes)
• Documentation update
• Maintenance update
• Other (please describe)

Testing

I extended the basic tests such as test_jacVecProdBWDFast.py, test_jacVecProdFWD.py, test_functionals.py and test_adjoint.jpy. But as indicated above, most do not pass.

Checklist

• I have run flake8 and black to make sure the Python code adheres to PEP-8 and is consistently formatted
• I have formatted the Fortran code with fprettify or C/C++ code with clang-format as applicable
• I have run unit and regression tests which pass locally with my changes
• I have added new tests that prove my fix is effective or that my feature works
• I have added necessary documentation