The fully Lagrangian stress history has not been run on the cylinder benchmark before, and it does not run now: with Lagrangian_Swarm (cells or nodes proxy) as Unknowns.DFDt, the first Navier-Stokes solve reports DIVERGED_FUNCTION_NANORINF after 0 iterations, so the residual itself is NaN. The swarm data holds no NaN at that point.
What works, same code, same environment:
- test_0070 (Maxwell box, Stokes, quad mesh) passes.
- The same box with
UnstructuredSimplexBox (regular and irregular) reproduces the Stokes curve.
- The same box with
uw.systems.NavierStokes (SUPG momentum) and the swarm history set on Unknowns.DFDt runs without NaN.
What fails: the DFG cylinder script (~/+Simulations/navier_stokes_supg/scripts/dfg_cylinder.py, -uw_stress_transport particle), a gmsh mesh with the cylinder hole, inlet parabola ramped from rest, at Wi 1, resolution 5, unrefined, seven steps. NaN at the first solve with: fill_param 3 and 8; DEVSS on and off; cells proxy and nodes proxy; SUPG momentum and trace-back (SLCN) momentum. The swarm is advected before each solve with the previous velocity (zero at step 1), as in test_0070, with population_control = dict().
Not bisected further. The remaining differences from the passing box case are the gmsh mesh with a hole and its boundary labels, the ramped inflow, the outlet treatment, and the script's stress diagnostics. Reproducer: smoke_p_*.log under ~/+Simulations/stress_transport/cylinder/; the box probe is particle_simplex_probe.py (quad|simplex|simplex_regular, stokes|ns_supg).
Context: the other three stress histories run this benchmark to Wi 10 (#735, #737). The particle history is the one with no Courant floor and no per-step resampling, which is why it is wanted here.
The fully Lagrangian stress history has not been run on the cylinder benchmark before, and it does not run now: with
Lagrangian_Swarm(cells or nodes proxy) asUnknowns.DFDt, the first Navier-Stokes solve reportsDIVERGED_FUNCTION_NANORINFafter 0 iterations, so the residual itself is NaN. The swarm data holds no NaN at that point.What works, same code, same environment:
UnstructuredSimplexBox(regular and irregular) reproduces the Stokes curve.uw.systems.NavierStokes(SUPG momentum) and the swarm history set onUnknowns.DFDtruns without NaN.What fails: the DFG cylinder script (
~/+Simulations/navier_stokes_supg/scripts/dfg_cylinder.py,-uw_stress_transport particle), a gmsh mesh with the cylinder hole, inlet parabola ramped from rest, at Wi 1, resolution 5, unrefined, seven steps. NaN at the first solve with: fill_param 3 and 8; DEVSS on and off; cells proxy and nodes proxy; SUPG momentum and trace-back (SLCN) momentum. The swarm is advected before each solve with the previous velocity (zero at step 1), as in test_0070, withpopulation_control = dict().Not bisected further. The remaining differences from the passing box case are the gmsh mesh with a hole and its boundary labels, the ramped inflow, the outlet treatment, and the script's stress diagnostics. Reproducer:
smoke_p_*.logunder~/+Simulations/stress_transport/cylinder/; the box probe isparticle_simplex_probe.py(quad|simplex|simplex_regular, stokes|ns_supg).Context: the other three stress histories run this benchmark to Wi 10 (#735, #737). The particle history is the one with no Courant floor and no per-step resampling, which is why it is wanted here.