Desired Force Field / Optimization Methods

One of the frequent requests for Avo2 was that the optimization should work with different tools (e.g., optimizing crystals with GULP or the XTB / GFN2 method, etc.)

Much like I did recently with electrostatics, I’d like to get some feedback about methods or programs you’d like to use for interactive optimization.

I mention interactive, because these should all be fast, like ~1s per step or faster (i.e., not DFT methods … unless you know something new. :slight_smile: )

Right now, the initial list will include:

  • Open Babel:
    • MMFF94
    • GAFF
    • UFF
  • XTB
    • GFN-FF
    • GFN0 (including crystals)
    • GFN1
    • GFN2

Ideally, I’d like to plug into OpenKIM or another interatomic library, but I don’t know much about them.

Other programs? What’s on your wish list?

I almost forgot… I’ll have a download for TorchANI, so ANI2x, ANI1x, and ANI1ccx will be available for suitable molecules.

1 Like

On one hand, it is nice to have multiple tools (here: multiple geometry optimization algorithms) at hand. Yet, the more options in the shelf, the more likely Joe and Jane who like to pick up Avogadro as a light weight, intuitive application will be puzzled facing the choice to make except either

a) their level of study of chemistry is advanced enough to recognize pros & cons (like computational cost, atoms considered by the parametrization)

b) though more interested e.g., in synthetic chemistry (continue to) have exchange with a group active in quantum chemistry as a source of advice and counsel, or

c) are at least guided by Avogadro’s documentation by a white list (perhaps a black list in addition) in lines like “MMFF94 usually is fine for small organic molecules” (or indicating a synopsis, e.g. as table) and “be aware of folding proteins, don’t use algorithm x”. The later because I saw successes by DeepMind and AlphaFold may mislead the younger user to the idea “here I have a tool (i.e., Avogadro), I started on my own, but now wonder why it takes so much time/the result differs so much from what I anticipated.” (Because e.g., the optimization is about the isolated molecule in vacuum, no solvent, no salts around.)

Thus my reluctance to amend the wish list as cases a) to c) actually are a sort sequence.

Personally, I’m usually fine with what MMFF94s returns as a result (e.g., for an illustration for a poster) if the molecular structure isn’t compared with results of quantum chemistry/a diffraction analysis) and by “chemical intuition” is fair enough. Perhaps though one could add “a wiggler”*, e.g. a button/an entry to launch OpenBabel’s /systematic/ conformer search, serving as a check if the suggested conformation not only is /an/ energetic optimum, but the global one (within the theory’s limitation).

*) I borrowed the term from synchrotron X-ray diffraction. The explanation once given of what it does still sticks to my mind.

We’ve almost always had “automatic” default choices. If you pick up any version of Avogadro and default to “optimize” or “AutoOptimize” it will Just Work™

That’s going to continue.

For example, with the charge models, the code has a list of supported elements. So electrostatics models that only support CHONS, etc. won’t even show up in the menu.

As far as documentation, that’s always a good idea, and also the purpose of the forum.

The main point of the thread isn’t that the defaults will change.

Avo2 makes it easy to write, download and install plugin scripts. So I wanted to know besides the list above, are there other programs / packages people would like to see integrations.

The balance is always tricky - to have a program that’s easy enough to use for beginning students but also supports integrating more advanced tools. Which is why community feedback is important. :wink:

At some point, I’ll be integrating a more complete conformer search toolbox. But as my research group’s recent papers have shown, force field methods are not reliable for accurate conformational search.

Also, we’ve shown that systematic searches are problematic beyond the smallest molecules because most conformers are high energy and thus not relevant for Boltzmann statistics.

The vision is that in the not-so-distant future, you’d be able to run a search and get the more-or-less complete ensemble. While it’s not perfect, at the moment, I would recommend the CREST program using either GFN-FF / GFN2 “hybrid” search or GFN2 directly.

I very much like your suggestions! I also would love to see integration of semiempirical methods, like AM1 or modern PM7 (MOPAC).

Now that OpenMOPAC is on GitHub and conda-forge that should be much easier to do.

It probably won’t be my first priority, but once the pieces are in place, it’s fairly easy to add more methods.

2 Likes