gradients in ADNodalKernel

[heinono1]

Check these boxes if you have followed the posting rules.

• Q&A General is the most appropriate section for my question
• I have consulted the posting Guidelines on the Discussions front page
• I have searched the Discussions forum and my question has not been asked before
• I have searched the MOOSE website and the documentation does not answer my question
• I have formatted my post following the posting guidelines (screenshots as a last resort, triple back quotes around pasted text)

Question

Is there a more-or-less ready-made way described how to access gradients of coupled variables in an ADNodalKernel? I am working on an AD nodal kernel that needs the gradient of a coupled variable. I know such gradients (nor gradients of Functors) are not readily accessible in ADNodalKernel so I am trying to figure out how best to do so, One path took me deep into the guts of libMesh and PETSc trying to identify the element and nodes that belong to each process - I got it working on 2 cores but invariably got segmentation fault on > 2 cores and I could not figure out why. One path I have been thinking of is to calculate the gradient separately as ADMaterials and then pass to the kernel and evaluate at _qp(=0). I obviously want to leverage the whole AD framework so I don't have to mess with Jacobians.

Thanks,
Olle

[GiudGiud]

Hello

Is there a more-or-less ready-made way described how to access gradients of coupled variables in an ADNodalKernel?

What type (finite element family and order) of variable are you taking the gradient of? As you are likely aware, the gradient of a linear lagrange variable at a node is not continuous so we don't offer this API.

invariably got segmentation fault on > 2 cores and I could not figure out why

Likely a lack of ghosting. You need algebraic ghosting to compute gradients near process boundaries.
You can look at ./userobjects/NodalPatchRecoveryBase.C to see a relationship manager (in charge of ghosting values) in action. It is also a good class to see how a nodal gradient patch-'average' can be defined

[GiudGiud]

hello

Yes it does though if you are using AD I would suggest adding a few APIs in the NodalPatchRecovery to return AD numbers to make things simpler.

The creation of an AD number "from scratch" is not that difficult, you just need to:

• do the same work as the non-AD version for computing the gradient
• whenever you using a variable value in that work, first create an AD number (ADReal) with the variable value and the index into the solution vector (with "1" for dv/ddof)

[GiudGiud]

@lindsayad or @hugary1995 there is no AD nodal patch recovery at this time right?

[heinono1]

Thanks, Guillaume. I think I can do that.

[lindsayad]

There is not, but NodalPatchRecoveryMaterialProperty inherits from the templated IndexableProperty so you could template NodalPatchRecoveryMaterialProperty easily and then work with AD material properties. This would require casting the variable gradient into a vector-valued material property, but that's not difficult

[heinono1]

Thanks, Alex and Guilaume. I did dig into how to do a patch-approximation with proper ghosting for gradients at nodes. At the end of the day, though, I concluded that sticking with my full-blown weak formulation was more efficient and more accurate.