dolfin team mailing list archive

[Matthew Knepley <knepley@xxxxxxxxx>]

In general, if a Newton iteration is not converging, you can

  a) Use a better line search (what is used now?)

  b) try a trust-region method

  c) use continuation in your parameter

  Matt

On Tue, Mar 17, 2009 at 10:38 AM, Patrick Riesen <priesen@xxxxxxxxxxxxxxx>wrote:

> Anders Logg wrote:
> > On Fri, Mar 13, 2009 at 07:48:39PM +0100, Patrick Riesen wrote:
> >> Patrick Riesen wrote:
> >>> Anders Logg wrote:
> >>>> On Wed, Mar 11, 2009 at 10:21:52AM +0100, Patrick Riesen wrote:
> >>>>>> hello
> >>>>>>
> >>>>>> So my Etafunction is now being correctly constructed and there are
> the
> >>>>>> correct values in the dof-vector(). i did this by a function class
> >>>>>> definition and a call to Eta as
> >>>>>>
> >>>>>>
> >>>>>> class Eta: public Function
> >>>>>> {
> >>>>>>    Eta (Mesh& mesh, Function& gamma, Form& form, ...) :
> Function(mesh,
> >>>>>> form, argument_id)
> >>>>>> { constructor....something as above }
> >>>>>>
> >>>>>> and then i call it as
> >>>>>>
> >>>>>> f_eta = new Eta(.....);
> >>>>>>
> >>>>>> in the main program.
> >>>>>>
> >>>>>> calling Eta as Function(mesh,form,id) i thought this should create a
> >>>>>> discrete function for f_eta, but the type of the f_eta function is
> >>>>>> 'user'. this gives me an error at the assembly, "missing eval() for
> >>>>>> user-defined function..."
> >>>>>> do i have to add an dummy eval() function to my Eta-class, or what
> do i
> >>>>>> have to change that no eva()-missing error is raised and my f_eta
> >>>>>> function is of type discrete?
> >>>>>>
> >>>>>> thanks for your help,
> >>>>>> patrick
> >>>>>>
> >>>>> sorry, i found an error, now the Eta is of type discrete (all 3
> argument
> >>>>> functions of my form are now discrete). but the eval() error at
> assembly
> >>>>> of the form is still present.
> >>>> Have you called vector() inside the constructor of Eta? That should
> >>>> make it discrete and the assembler should not complain. Perhaps there
> >>>> is some other function in you form that is missing an eval (or a
> >>>> vector).
> >>>>
> >>>   hi anders,
> >>> yes i did that, and now i found the error fortunately. i did not
> >>> initialize the bilinear form correctly with some placeholder function
> >>> for Eta. now the assembly is working, so i can go on.
> >>>
> >>> thank you & regards,
> >>> patrick
> >> hi all
> >> my code is now running but the Eta-funcs still give me some problems. i
> >> have initialized my bilinear/linear form with two constant functions as
> >>
> >> f_eta = new Function(mesh, 1.)
> >> f_Deta = new Function(mesh, 1.)
> >
> > Are you using an old version? This doesn't work with the current
> > interface.
> >
> >> the first newton iteration gives me a newtonian solution from which i
> >> take the velocity to compute the invariant in another form (gamma) and
> >> then i construct the discussed functions Eta, DEta, and renew f_eta,
> >> f_Deta as
> >>
> >> f_eta = new Eta(gamma,.....)
> >> f_Deta = new DEta(gamma,......)
> >>
> >>
> >> so i guess in the second iteration it will assemble with a variable
> >> viscosity as f_eta, f_Deta are different now but it just converges to
> >> the newtonian solution.
> >>
> >> is this a problem, that i constructed the forms with a constant
> >> function? does dolfin then not consider the values in the vector() of
> >> the new discrete functions for f_eta, f_Deta after renewing?
> >
> > Yes, if you construct your form with some Function, then that Function
> > will be used. The form won't change just because you create a new
> > Function later.
>
> hi,
> thanks again for the help so far,
> now i think i managed to construct my forms correctly and the code is
> running. the solution looks reasonable, starting from newtonian solution
> with n=1 and then changes when the powerlaw exponent is increased.
>
> however, the convergence is poor, e.g. with a powerlaw exponent of n=2
> it takes almost 30 iterations to converge (incremental criterium) and it
> does not converge beyond n=2 so far.
>
> to compute the variable viscosity terms in my form i followed the
> example of the quadrature elements implementation of the nonlinear
> poisson demo (sect. 6.10 in the ffc manual from nov 2008) and solve a
> linear system with another form-file to obtain the invariant of the
> strain tensor. then as discussed in previous mails (see above) i take
> power n-something of the invariant and compute the viscosity in the
> dof-vector adaptively in each iteration.
>
> is there a way to play around with the newtonsolver? now i use a
> lu-factorization with umfpack and ublas-backend, i just
> had
>
> NewtonSolver newtonsolver;
>
> e.g but i tried the same as
>
> NewtonSolver newtonsolver(lu,default_pc);
>
> i got an error:
>
> In file included from PowerlawStokesSolver.cpp:4:
> PowerlawStokesSolver.h:74: error: 'lu' is not a type
> PowerlawStokesSolver.h:74: error: 'default_pc' is not a type
>
> BTW: i use dolfin0.8.1 with dolfin-grade2
>
> regards,
> patrick
>
> _______________________________________________
> DOLFIN-dev mailing list
> DOLFIN-dev@xxxxxxxxxx
> http://www.fenics.org/mailman/listinfo/dolfin-dev
>



-- 
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