Run with -snes_monitor and -ksp_monitor_true_residual and send the output
> On Sep 20, 2023, at 9:34 AM, Harry-Arthur Cousin <harry-art...@hotmail.fr> > wrote: > > Hello, > > I'm on petsc version 3.4.5. > > My program has a time loop where I have to solve a non-linear system of the > form F(X) = D*L*J + I + kappa(X) where X is an unknown vector, D and L are > diagonal square matrices, J is a vector that can be calculated by solving a > linear system with X and kappa is a non-linear application. I use the > "-snes_mf" option to, as I understand it, approximate the Jacobian using > finite differences. > > A first version of my problem could be solved, so I made it more complex by > changing the values of the matrix L, which used to have the same value on the > whole diagonal and now has values of the same order of magnitude but > different. This second version takes longer and doesn't converge after a > while. I get the message: Nonlinear solve did not converge due to > DIVERGED_LINEAR_SOLVE. However, when I look at the faq > (https://petsc.org/release/faq/#why-is-newton-s-method-snes-not-converging-or-converges-slowly), > it indicates that it's probably a Jacobian problem. What should I do? Do I > have to calculate the explicit Jacobian? The explicit Jacobian requires the > inversion of a matrix (due to the linear system for solving J). > > > Kind regards, > > Harry Cousin
