Rémi Delaporte-Mathurin

I'm still struggling to understand exactly what would this convergence plot show. Is it: at each time step, we record the final error (absolute and relative) of the Newton solver...

This sounds very good to me and I don't think it would be super hard to implement. Maybe we can start with a prototype in fenics to see how it's...

@KulaginVladimir this is very interesting! (I think on your second plot you meant to label the y axes with absolute error and relative error right?) Do we really need to...

I understand now that the NewtonSolver object created on call in NonLinearVariationalSolver cannot be retrieved outside the class as it's not stored as an attribute. Your solution seems to work...

Sure!

Does it compare well in terms of performance?

I don't think there's a need to test performance accurately here. Do you know how would one setup different solver parameters with this method? @allentro was interested

Any parameter really. But I guess it's the same as when you do ```python solver.parameters["error_on_nonconvergence"] = False solver.parameters["absolute_tolerance"] = self.settings.absolute_tolerance solver.parameters["relative_tolerance"] = self.settings.relative_tolerance solver.parameters["maximum_iterations"] = self.settings.maximum_iterations solver.parameters["linear_solver"] = self.settings.linear_solver ```...

Currently in your code, the solver is redefined at every `solve_once` call but it could be defined only once (maybe outside of the class) and reused everytime.

I agree that this problem isn't ideal for testing alternative solvers and preconditioners. Both mumps and UMFPACK are direct linear solvers, I'd be interested to investigate pre-conditioned iterative linear solvers....