Minimize objective function with ROL and Panzer packages.

[delcmo]

Hello,

I have two questions for the Panzer and ROL developers.

We are trying to use a gradient descent method to minimize an objective function defined by an ODE using ROL package within Vertex-CFD. Vertex-CFD is built on Panzer.

I had a look at the examples and tutorials available in the ROL package and found some example solving for PDEs with Intrepid (Trilinos/packages/rol/example/PDE-OPT).

Before I dive into the source code, I would appreciate it if any developers could answer the following questions:

1. I would like to know if there is an example that uses Panzer instead of Intrepid to solve for an ODE using a ROL solver?
2. Is there a documentation for solver options available in ROL?

Thanks,

Marco

[github-actions]

Automatic mention of @aj463

[dridzal]

We don't have a Panzer example. There are ODE examples (possibly even including ROL) in MrHyDE, which uses the Panzer DOF manager (but no other functionality). Which underlying Panzer capability are you actually using when solving and optimizing ODEs? If this is related to Tempus and Thyra model evaluators, then we may have something useful in ROL. In any case, the most important task would be to implement a ROL objective function, including value() and gradient(), feed the objective function into an instance of ROL::Problem, say problem, and then run problem.check(). If the finite difference checks pass, we can provide further instructions of what algorithms to select.

[rppawlo]

There is a ROL example in Panzer at:
Trilinos/packages/panzer/adapters-stk/example/main_driver/main_driver_opt.cpp

This relies on finite difference for derivatives. There is a branch where we were recovering the derivative capability at: #14059 The changes in this PR get the example running but break a number of other panzer unit tests. In April I will be working on getting these changes merged in.

[delcmo]

Perfect! I did not think of looking for a ROL example in Panzer. I will have a look at it and will get back to you two if I have any questions. Thanks @dridzal and @rppawlo.

Marco