MALAMUTE ROM Development: Jacobian Reduction

[raj-anant]

Dear MOOSE Team,

We are developing a ROM for MALAMUTE and are exploring different parts of the problem where we can achieve a reduction. In particular, we are interested in achieving reduction at the level of the Jacobian. However, we had a few concerns about the feasibility of the implementation of any such approach. Specifically, we had the following questions:

(i) Is it possible to read in an approximate base Jocobian for the problem through a file or generate it through a script inside MOOSE?
(ii) Is it possible to make small updates to the base Jacobian as the solution evolves using a script inside MOOSE? Essentially, we want to explore bypassing the computation of the Jacobian through AD at each non-linear iteration.
(iii) Is it possible to bypass PetSci for the solution of the non-linear iteration?
(iv) Is there a convenient way for dumping the Jacobian to a file every few time steps/iterations? Write now we are using "-ksp_view_mat" option, but that dumps the Jacobian directly into the text output file serially at every non-linear iteration, which is very difficult to post-process.
(v) Similarly, is there a way to dump the residuals to a file every few non-linear iterations? We are looking for the residuals at the start of the iteration, not just the final value after convergence.

We would appreciate any help on this.

Thanks,

Anant

[GiudGiud]

@cticenhour

[jiangwen84]

@GiudGiud This is more like framework/solver level question. @raj-anant worked with me, and I do not have sufficient knowledge to answer some of his question. We hope posting those questions here will provide us better directions to approach next.

[lindsayad]

The short answer to all of your questions is sadly no but we welcome contributions. You can write your own Executioner and SolveObject and FEProblemBase derived objects in which you can do almost anything you want. You could override the Jacobian computation routine from the latter for instance.

I'll give a little more detail on a few:

iii) You can write your own custom line search if that's what you mean by bypassing PETSc's nonlinear solution. However, you would still be using the built-in KSP in the PETSc SNES object
iv) You could hack in some condition MatView calls within the framework contingent on time step/iteration number etc.
v) similarly as iv)