[petsc-users] Non-linear solve: DIVERGED_LINE_SEARCH

[Matthew Knepley]

On Thu, Sep 7, 2023 at 8:49 AM Karthikeyan Chockalingam - STFC UKRI via
petsc-users <petsc-users at mcs.anl.gov> wrote:

> I just noticed that the line search diverges even for the linear Poisson
> equation (where the Jacobian is a constant). I am not sure, what I am doing
> wrong.
>

You need to test the Jacobian. We have facilities for this. Run using

  --snes_test_jacobian

This will tell you at each step whether the Jacobian matches the finite
difference approximant. It looks like your Jacobian at the second step is
wrong. Is it possible that you are not zeroing out the matrix before adding
in updates? That would make the second step wrong.

  Thanks,

     Matt


>
>
> Best,
>
> Karthik.
>
>
>
> *From: *Chockalingam, Karthikeyan (STFC,DL,HC) <
> karthikeyan.chockalingam at stfc.ac.uk>
> *Date: *Thursday, 7 September 2023 at 12:19
> *To: *Stefano Zampini <stefano.zampini at gmail.com>
> *Cc: *petsc-users at mcs.anl.gov <petsc-users at mcs.anl.gov>
> *Subject: *Re: [petsc-users] Non-linear solve: DIVERGED_LINE_SEARCH
>
> I am trying to solve the non-linear scalar problem
>
> -\nabla \cdot \big( (1+u^2)\nabla u\big)=f
>
> Using finite element formulation.
>
>
>
> The linear iteration converged – does that mean my Jacobian is correct??
>
>
>
> I am hoping the non-linear residual goes down to 1-e11.
>
>
>
> Kind regards,
>
> Karthik.
>
>
>
>
>
> *From: *Stefano Zampini <stefano.zampini at gmail.com>
> *Date: *Thursday, 7 September 2023 at 11:49
> *To: *Chockalingam, Karthikeyan (STFC,DL,HC) <
> karthikeyan.chockalingam at stfc.ac.uk>
> *Cc: *petsc-users at mcs.anl.gov <petsc-users at mcs.anl.gov>
> *Subject: *Re: [petsc-users] Non-linear solve: DIVERGED_LINE_SEARCH
>
> The solver did not diverge.
>
> It was the line search that was not able to make further progress in
> minimizing the 2-norm of the residual.
>
> This is common in nonlinear solvers. It would help if you tell us what you
> are trying to solve.
>
> Note that at the first step, your residual norm is already 1.e-6. What
> kind of accuracy do you want?
>
>
>
>
>
> Il giorno gio 7 set 2023 alle ore 13:08 Karthikeyan Chockalingam - STFC
> UKRI via petsc-users <petsc-users at mcs.anl.gov> ha scritto:
>
> Hello,
>
>
>
> The non-linear solution diverged. The final solution seems right and I
> believe the Jacobian is correct (not 100% certain).
>
>
>
> I am not sure if I doing something wrong in the solver setting.
>
>
>
>
>
>  0 SNES Function norm 3.890991210938e-03
>
>     0 KSP Residual norm 9.037762538598e+00
>
>     1 KSP Residual norm 2.120375403775e-01
>
>     2 KSP Residual norm 5.155439334511e-03
>
>     3 KSP Residual norm 1.394364169369e-04
>
>     4 KSP Residual norm 9.233543407204e-06
>
>   Linear solve converged due to CONVERGED_RTOL iterations 4
>
>
>
>       Line search: Using full step: fnorm 3.890991210938e-03 gnorm
> 7.701511565083e-06
>
>   1 SNES Function norm 7.701511565083e-06
>
>     0 KSP Residual norm 5.630229687829e-03
>
>     1 KSP Residual norm 1.030475601271e-04
>
>     2 KSP Residual norm 2.576454714319e-06
>
>     3 KSP Residual norm 6.669316846898e-08
>
>     4 KSP Residual norm 3.215810984829e-09
>
>   Linear solve converged due to CONVERGED_RTOL iterations 4
>
>
>
>       Line search: gnorm after quadratic fit 1.805500533481e-05
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 2.563759884284e-05 lambda=3.0804668685096816e-02
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 3.332721829751e-05 lambda=3.0804668685096817e-03
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 4.102754045833e-05 lambda=3.0804668685096822e-04
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 4.872893294880e-05 lambda=3.0804668685096822e-05
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 5.643043250787e-05 lambda=3.0804668685096822e-06
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 6.413194279696e-05 lambda=3.0804668685096827e-07
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 7.183345417492e-05 lambda=3.0804668685096828e-08
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 7.953496567312e-05 lambda=3.0804668685096829e-09
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 8.723647719173e-05 lambda=3.0804668685096831e-10
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 9.493798871875e-05 lambda=3.0804668685096832e-11
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 1.026395002516e-04 lambda=3.0804668685096835e-12
>
>       Line search: Cubic step no good, shrinking lambda, current gnorm
> 1.103410117889e-04 lambda=3.0804668685096835e-13
>
>       Line search: unable to find good step length! After 12 tries
>
>       Line search: fnorm=7.7015115650831560e-06,
> gnorm=1.1034101178892401e-04, ynorm=6.2052357872955976e-03,
> minlambda=9.9999999999999998e-13, lambda=3.0804668685096835e-13, initial
> slope=-5.9313318983096354e-11
>
> Nonlinear solve did not converge due to DIVERGED_LINE_SEARCH iterations 1
>
>
>
>
>
> Thank you for your help.
>
>
>
> Kind regards,
>
> Karthik.
>
>
>
> --
>
> *Dr. Karthik Chockalingam*
>
> Senior Research Software Engineer
>
> High Performance Systems Engineering Group
>
> Hartree Centre | Science and Technology Facilities Council
>
> karthikeyan.chockalingam at stfc.ac.uk
>
>
>
>  [image: signature_3970890138]
>
>
>
>
>
> --
>
> Stefano
>


-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener

https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20230907/4d9271b8/attachment.html>