Automatic differentiation

[rebpoli]

Hi

What is the current status of AD in libmesh?
I can see that Moose reports that in their webpage. Is there a standard way to use it?

Thank you
Renato

[roystgnr]

We don't use AD in libMesh itself, but Moose uses MetaPhysicL for AD, which is an optional libMesh dependency (because we use it for constructing projection operators as sparse matrices for GMG). If you look at MOOSE ADRealForward.h, their standard types for it are defined there; you can also see examples in a few of the unit tests like sparse_derivs_unit.C in the MetaPhysicL test directory.

[rebpoli]

Thank you. Is that the way to go? Any other library I should consider? Renato

…

On Tue, Apr 8, 2025 at 2:06 PM roystgnr ***@***.***> wrote: We don't use AD in libMesh itself, but Moose uses MetaPhysicL for AD, which is an optional libMesh dependency (because we use it for constructing projection operators as sparse matrices for GMG). If you look at MOOSE ADRealForward.h, their standard types for it are defined there; you can also see examples in a few of the unit tests like sparse_derivs_unit.C in the MetaPhysicL test directory. — Reply to this email directly, view it on GitHub <#4132 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMOVER66IZTGRXX2G7BEO7T2YQMZXAVCNFSM6AAAAAB2UZCBHKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDOOBXGQYTOOBZG4> . You are receiving this because you authored the thread.Message ID: ***@***.***> *roystgnr* left a comment (libMesh/libmesh#4132) <#4132 (comment)> We don't use AD in libMesh itself, but Moose uses MetaPhysicL for AD, which is an optional libMesh dependency (because we use it for constructing projection operators as sparse matrices for GMG). If you look at MOOSE ADRealForward.h, their standard types for it are defined there; you can also see examples in a few of the unit tests like sparse_derivs_unit.C in the MetaPhysicL test directory. — Reply to this email directly, view it on GitHub <#4132 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMOVER66IZTGRXX2G7BEO7T2YQMZXAVCNFSM6AAAAAB2UZCBHKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDOOBXGQYTOOBZG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[roystgnr]

IIRC there's a ton of AD libraries around these days. The big advantages you'd get with MetaPhysicL are that it's already in contrib/ and that we've got facilities that ease its integration with MPI and with some of our numerics classes. The big disadvantage is that it only does forward-mode AD, which is usually appropriate when your number of calculated outputs is similar to or greater than your number of inputs; if you have far more input variables than outputs then you might want adjoint AD (also called reverse AD, and I think the ML cool kids call it "backpropagation") and you'd need to look elsewhere.