Scipy optimizer module

[keceli]

PR Type

• Breaking change
• Feature
• Patch

Brief Description

This is the initial PR to add a module for optimization. It is based on single point energies and uses scipy.optimize.minimize function. There are two similar modules, one satisfies simde.AOEnergy and the other simde.Energy.

Not In Scope

Optimizer modules that make use of the gradient and geomeTRIC, PyBerny, and ASE will follow up this one.

PR Checklist

• Is documented
• Is tested
• Adheres to applicable organization standards

TODOs

• Task 1
• Task 2

[ryanmrichard]

Overall looks pretty good.

[ryanmrichard]

Should be SimDE not NWX.

[keceli]

I want to be able use NWChemEx driver modules to get CCSD, MP2 energies once available.

[ryanmrichard]

It should still be SimDE. Requiring the user to build NWChemEx adds a lot of build overhead and makes your plugin dependent on literally the entire stack.

That said, building of other plugins is useful for testing, but should be limited to integration testing only. I'll warn you that, to my knowledge, no one has yet tried to use NWChemEx as a dependency for integration testing, so there may be hiccups.

[keceli]

It should still be SimDE. Requiring the user to build NWChemEx adds a lot of build overhead and makes your plugin dependent on literally the entire stack.

Could you clarify? I assume the main role of this repo is to provide geometry optimization, which would require energies and gradients. NWChemEx is the repo that provides access to all plugins that can give those properties. How would I access them if I just use SimDE?

[keceli]

I understand that you mean those dependencies would only be required during testing, but I think a user of this repo would expect to access them directly.

[jwaldrop107]

The NWChemEx repo is the one that pulls together all of the functionalities. It's not meant to be used as a dependency of our plugins. This repo will more likely be a dependency of NWChemEx, actually. That way, when someone builds NWChemEx, they will get the Integrals, SCF, ChemCache, the geometry optimization modules here, and any other plugins we associate with it down the line.

[ryanmrichard]

I feel like we should be creating a new property type for optimizers. Something like Optimize<X,Y...> which optimizes X by varying Y.... So Optimize<AOEnergy, Nuclei> would be the property type. I'd set it up so the inputs are the union of the inputs to X as well as the Y... objects and the returns are the return of X and the optimized values of the Y. So the API of the module would essentially be energy, optimized_nuclei (AOs, ChemicalSystem, Nuclei of chemical system)