Automatic method for enabling SMP for AD

[aeslaughter]

Reason

When AD objects are used you need to have the following in the input file, but this isn't intuitive.

[Preconditioning]
  [smp]
    type = SMP
    full = true
  []
[]

[Executioner]
  type = Steady 
  solve_type = NEWTON
[]

Design

I propose the following that will automatically add the correct Preconditioning if it is not provided.

[Executioner]
  type = Steady 
  solve_type = AD_NEWTON
[]

Impact

This will make using MOOSE with AD easier and less prone to errors.

[aeslaughter]

@idaholab/moose-team Are people OK with this or have a better idea?

[lindsayad]

I personally think that any time a user specifies solve_type = NEWTON, the default behavior should be to add SMP with full = true regardless of whether AD is being used or not. However the Jacobian is being computed, NEWTON requires a complete (and accurate) Jacobian in order to be successful

[fdkong]

I kind of agree with @lindsayad. If we use NEWTON, we should trigger ``full=true" by default. But I think we also should reserve an option so that users can turn it off if they want. For example, if the equation is very simple, we do not need to form the world.

[tophmatthews]

If I remember correctly...NEWTON isn't really explicitly newton right? I've used it in the past with a close enough jacobian. Also, a full SMP is used sometimes with PJFNK. How about

[Executioner]
  type = Steady 
  solve_type = PJFNK/NEWTON
  SMP = none/partial/full
[]

[rwcarlsen]

I'm with @fdkong and @lindsayad. I have been converting a bunch of stuff to AD and it is really easy to forget to set solve_type=NEWTON along with a full=true SMP preconditioner in input files. I frequently forget to set one or more of those things.

[rwcarlsen]

To be clear - solve_type=AD_NEWTON would be an improvement over the status quo, but I'd still might forget to update files to use that that already have solve_type=NEWTON.

[lindsayad]

If I remember correctly...NEWTON isn't really explicitly newton right?

It depends on whether you're using a line search or not. If you're using line_search = basic (which really means no line search), then you are conducting pure Newton. If you're using a line search (e.g. a globalization strategy), then yes in terms of jargon, you're using a Quasi-Newton method.

I've used it in the past with a close enough jacobian

You can definitely achieve convergence with a non-perfect Jacobian using Newton (or Quasi-newton), but you'll be pretty unlikely to achieve a quadratic convergence rate at any point in the non-linear solve.

[Executioner]
type = Steady
solve_type = PJFNK/NEWTON
SMP = none/partial/full
[]

I'm not quite sure what you mean here.

[tophmatthews]

Thanks for the replies @lindsayad. I was suggesting having an SMP input parameter as an enum with options none, partial, full, similar to how the solve_type is an enum. You can even throw a warning if solve_type=NEWTON without SMP=full which should fail tests, but allow you to set it to none if you wish.