Finite relative residuals for 0th Newton iteration in time-dependent problem

[dokken]

@mcjbo

This seems very counterintuitive. Does a new solve/time step not indicate solving a new problem?

Feel free to file a pull request to dolfinx rectifying this. It would simply be adding _residual0 = 0.0; at:

FEniCS/dolfinx/blob/main/cpp/dolfinx/nls/NewtonSolver.cpp#L151

which would then consistently give you inf for iteration 0 using “residual” method.

@mcjbo

it seems that the relative residual is computed as r_\mathrm{rel} = \frac{r_\mathrm{abs}}{\Delta x}rrel=rabsΔxr_\mathrm{rel} = \frac{r_\mathrm{abs}}{\Delta x}, with \Delta xΔx\Delta x being the increment of the first Newton iteration. Thus, e.g. for the hyperelastic problem, the relative residual would have the unit \frac{N}{m}Nm\frac{N}{m}. Is this physical? What’s the interpretation of the relative residual?

You would have to ask the original author (@garth?). As i have already stated it is easy to implement your own Newton solver if you are not satisfied with the convergence criteria or logic.