Add demo of nonlinear scipy solver #3468
Conversation
jorgensd
requested review from
jhale,
michalhabera,
garth-wells,
chrisrichardson,
mscroggs,
francesco-ballarin and
IgorBaratta
Thanks, Nate! Co-authored-by: Nate <[email protected]>
Add header Fix functional for complex run.
Co-authored-by: Nate <[email protected]>
|
We do already have a demo that uses a SciPy solver: https://github.com/FEniCS/dolfinx/blob/main/python/demo/demo_types.py |
We only show how to solve linear problems. |
|
I at least think we should make it clearer from the demo title that it shows how to use scipy. |
There was a problem hiding this comment.
Not withstanding issue on where to put this raised by @garth-wells, here are my comments.
|
|
||
| # If the Jacobian is independent of uh, we can avoid re-assembling the matrix | ||
|
|
||
| is_nonlinear = uh in ufl.algorithms.expand_derivatives(jacobian).coefficients() |
There was a problem hiding this comment.
@jhale I've here used `ufl to figure it out.
I think it is a quite nice point to make, as some people prefer always using a non-linear solver, but could benefit alot from not having to reassemble J at every step of the Newton solver (even if it is going to be just 1).
As we now support DOLFINx without PETSc, it is good to have a show-case of how to interface with scipy to solve problems.