Understanding solver parameters in NonlinearVariationalSolver

27 Jun 2021

      Hi everyone,

I am trying to understand the usage of solver parameters that are passed to
the NonlinearVariationalSolver in firedrake. In order to do some tests, I
decided to reproduce some of my previous calculations done in FEniCS for
nonlinear viscoelasticity. In old dolfin, I had the following for solver
parameters to be passed to PETSc's SNES:

snes_solver_parameters = {"nonlinear_solver": "snes",
                          "snes_solver": {"linear_solver": "lu",
                          "maximum_iterations": 50,
                          "report": True,
                          "line_search": 'basic',
                          "error_on_nonconvergence": False,
                          "relative_tolerance": 1e-7,
                          "absolute_tolerance": 1e-8}}
...
...
solver = NonlinearVariationalSolver(problem)
solver.parameters.update(snes_solver_parameters)

Now as I understand, in Firedrake the options passed to the solver are
consistent with the corresponding names in PETSc (correct me if I am
wrong), and therefore I did:

solver_params={
    "snes_type": "newtonls",
    "ksp_type": "preonly",
    "pc_type": "lu",
    "snes_max_it": 100,
    "snes_rtol": 1.e-7,
    "snes_atol": 1.e-8,
    "snes_view": None,
    "ksp_atol": 1.e-10, # not needed?
    "ksp_rtol":1.e-8,
    "ksp_view": None,
    "snes_linesearch_type": "basic",
    "snes_no_convergence_test": 1 # this has no effect
}
solver = NonlinearVariationalSolver(problem,
solver_parameters=solver_params)

However, for some reason, I am unable to get the solver to converge when
trying to solve the exact same problem as FEniCS. I experimented with
different quadrature degrees as well as tweaked the solver parameters for
SNES to best match what I am doing in dolfin. I understand that the form
compiler is different hence the precise code that is generated may be
different (I am not an expert on code generation) even though the python
scripts may look similar due to UFL syntax.

In fact the weird part is that the solver seems to converge fine till a
certain point in my input (loading). You can see this in the plot below
where I plot the final stress-stretch relation obtained at a point in my
domain from dolfin and firedrake..

[image: firedrake_vs_fenics.png]

I have marked the abscissa beyond which the solver fails to converge in
Firedrake. The complete codes are here:
https://github.com/bhaveshshrimali/comparingFiredrakeFEniCS

I tried setting the option `snes_no_convergence_test` to 1/True but that
had no effect as the SNES solver still diverges with `DIVERGED_DTOL`. In
the end I put the call to solve inside a try/catch block to at least be
able to see the solution till that point. And it surprisingly is accurate
till that point.

Would anyone have any clue of what could be going wrong? I wouldn't expect
that you go through the entire code(s) as that could be fairly long, but
any pointers would be helpful :)

I am using firedrake via Docker on windows.

Thanks

 --

Bhavesh Shrimali

Matthew Knepley

Ed Bueler

tags

participants (3)