Set different solver parameter for adjoint

[HaoLI-KU]

Hello, there. I am new to Firedrake.
I’m wondering if it’s possible to set different solver parameters for the forward and adjoint problems. As I understand it, if we use Firedrake’s own adjoint module to compute the gradient, the adjoint solver will automatically use the one for the forward problem. Is there a way to customise the adjoint solver?

[dham]

Yes. When you set up the forward NonLinearVariationalSolver you can pass an adj_kwargs argument whose value is a dictionary that you would like to use as the keyword arguments for the adjoint linear variational solver. To set the solver parameters you would use adj_kwargs={'solver_parameters':{...}}

[HaoLI-KU]

Thanks a lot, David!
It works for NonLinearVariationalSolver. But it seems not for LinearVariationalSolver. Am I correct?
On the other hand, linear PDEs don’t necessarily require it but is there a way to do so for them just out of curiosity?

[dham]

Ah yes, I see nobody has added the argument and passed it through. Probably because it's a pretty niche thing to do.

In Firedrake, LinearVariationalSolver is in fact a NonLinearVariationalSolver with snes_type:ksponly so you can achieve the same effect by simply formulating your problem as a nonlinear problem and setting the snes type.

[HaoLI-KU]

Thanks, David!