11
$\begingroup$

I have never used any AI driven calculation package before and to be honest don't fully understand the ins and outs of it. To be more specific I'm looking for something that can find local minima for molecules along with an absolute minima. I would like to get some quick exposure to some available packages (maybe some pros/cons) so that I can get a better idea of what exactly would be the ideal package for me.

Please let me know what is available for this, and what exactly I should be looking out for (ie: ''Make sure you find AI driven not Monte Carlo based'').

I would like to avoid calculating the whole PES, and would like to arrive at the minima more directly.

Improve this question
$\endgroup$
11
  • $\begingroup$ +1. Only a few more questions/answers and you fulfill commitment! The question might need some clarification. You want to avoid actually calculating the PES point-by-point in say, MOLPRO? $\endgroup$
    – Nike Dattani
    Oct 19, 2020 at 3:30
  • $\begingroup$ CREST by XTB is my newest favorite algorithm for exploring PESs. They also have metadynamics which can explore the PES (not sure how it is different from CREST). $\endgroup$
    – Cody Aldaz
    Oct 19, 2020 at 4:27
  • $\begingroup$ @NikeDattani Yes, the main goal is to avoid point-by-point calculations. I can edit the question to clarify it some more if you think that will help. $\endgroup$
    – Cavenfish
    Oct 19, 2020 at 4:41
  • 1
    $\begingroup$ I also don't understand why you're mentioning AI? I am aware of efforts to do global minimization using machine learning, but these are ongoing and there is very little published research in this area. $\endgroup$
    – jheindel
    Oct 20, 2020 at 23:36
  • 1
    $\begingroup$ @B.Kelly - we just published a pretty thorough evaluation of over 600 molecules against DLPNO-CCSD(T) energies: doi.org/10.1002/qua.26381 (open access) Suffice to say that PM7 is not good, but GFN2 is an excellent balance of speed and accuracy. I would definitely not trust HF since dispersion interactions are critical for conformed energies. $\endgroup$
    – Geoff Hutchison
    Dec 31, 2020 at 1:57

2 Answers 2

Reset to default
9
$\begingroup$

The Atomic Simulation Environment has two nice implementations of global optimization algorithms. The first is a basin hopping algorithm from a 1997 paper by Wales and Doye in J. Phys. Chem. A. The second is a minima hopping algorithm from a 2004 paper by Goedecker in J. Chem. Phys. There are several nice example use-cases here.

If you're looking at relatively simple organic molecules, I agree with a prior comment that CREST is a nice code for sampling lots of conformations to find low-energy structures, which uses the semi-empirical xTB code as a driver.

I should clarify though that finding a global minimum is easier said than done, and the success of any method will depend a lot on the various degrees of freedom for your system.

Improve this answer
$\endgroup$
1
  • 2
    $\begingroup$ +1. Lots of useful information in there, as usual! I'm quite glad this question got answered actually, as it seemed simple yet got left without answers for months. Also I'm especially happy to see that 1997 paper by Wales and Doye mentioned, since Jon Doye was the internal examiner for my doctoral thesis at Oxford! $\endgroup$
    – Nike Dattani
    Dec 28, 2020 at 7:34
4
$\begingroup$

Bayesian Optimization

There are some nice options for exploring potential energy surfaces using Bayesian optimization. This has the advantage of using Gaussian Process regression to build a surrogate to the potential energy surface

Bayesian optimization works extremely well for "expensive" functions (e.g., minutes to hours per point) in which the overhead of performing the GPR is offset by efficiently exploring the potential energy surface - perfect for first principles calculations.

My group has some efficient methods for finding the global minima, but not exploring the entire PES.

You might want to check out BOSS by the Aalto group: https://cest-group.gitlab.io/boss/tutorials.html e.g. "Efficient Cysteine Conformer Search with Bayesian Optimization"

They find a >10x speedup over a genetic algorithm.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Awesome. This answer has a lot of useful information indeed! $\endgroup$
    – Nike Dattani
    Dec 31, 2020 at 3:59

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .