Have you tried adjusting the numerical Jacobian h, thr finite difference perturbation ?
It defaults to 1.e-6, but should be lower if the variable being perturbed is small. Thanks. On Mar 30, 2016 7:01 AM, "David Knezevic" <[email protected]> wrote: > On Wed, Mar 30, 2016 at 7:47 AM, Paul T. Bauman <[email protected]> > wrote: > > > Hey David, > > > > On Tue, Mar 29, 2016 at 10:43 PM, David Knezevic < > > [email protected]> wrote: > > > >> I'm using FEMSystem for a problem, and I'm getting solid quadratic > >> convergence with the Newton solver. However, when I turn on the finite > >> difference jacobian check using "verify_analytic_jacobians = 1.e-6" it > >> reports an error: > >> "Relative error 0.133922 detected in analytic jacobian on element 0!" > >> > > > > There are circumstances where I get this behavior, that is quadratic > > convergence and the Jacobian is (partially) wrong. The one that pops to > > mind is my Rayleigh-damping implementation, where I haven't yet > implemented > > parts of the Jacobian. (Just to satisfy any curiosity, for material > > nonlinearities, C(i,j,k,l) has non-zero derivatives w.r.t. strain, so I > > haven't yet gone to the trouble to implement them, but I have the > Jacobian > > for all the other parts.) > > > > > >> I expected the code to pass this test, given that I'm getting good > >> convergence behavior. So before I do too much more bug-hunting, I just > >> wanted to check if there's a chance that I might be getting a "false > >> positive" with the finite difference jacobian check? > >> > > > > I have yet to find a case where I got a false positive. I find it helpful > > to get to a very small problem and compare the elements to zone in on the > > terms that differ. > > > > > Thanks for your comments. The problem I'm considering is plasticity, using > the radial return algorithm. As far as I can tell, the code matches the > text book, and it converges correctly. However, it doesn't match the finite > difference Jacobian from FEMSystem. So there are two possibilities: > > 1) Somehow the finite difference Jacobian is inconsistent with the radial > return algorithm. This doesn't seem impossible to me, given that the radial > return algorithm is highly path-dependent. > > 2) There is a bug somewhere in my analytical Jacobian. > > I'll look some more for a bug (I'll compare the elements, like you said). > But I was wondering if you think 1) is a possibility? > > David > > ------------------------------------------------------------------------------ > Transform Data into Opportunity. > Accelerate data analysis in your applications with > Intel Data Analytics Acceleration Library. > Click to learn more. > http://pubads.g.doubleclick.net/gampad/clk?id=278785471&iu=/4140 > _______________________________________________ > Libmesh-users mailing list > [email protected] > https://lists.sourceforge.net/lists/listinfo/libmesh-users > ------------------------------------------------------------------------------ Transform Data into Opportunity. Accelerate data analysis in your applications with Intel Data Analytics Acceleration Library. Click to learn more. http://pubads.g.doubleclick.net/gampad/clk?id=278785471&iu=/4140 _______________________________________________ Libmesh-users mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/libmesh-users
