25
$\begingroup$

Machine learning is an increasingly common tool for developing force fields for molecular dynamics simulations. It's not totally clear what should be considered a machine-learning potential, but let's just say a machine learning potential is one which has many, many parameters which do not have any obvious physical interpretation. Any kind of neural network, Gaussian process, or permutationally invariant polynomial model, I would consider as an ML potential for this question.

I have had quite a few conversations in which people say things along the lines of "machine learning potentials are slower than molecular mechanics potentials." I've observed this to be true for certain ML potentials I've used, but I haven't seen any real benchmarking of this.

Are there references which provide good comparisons of the speed of ML potentials to comparable classical force fields? I have seen a few papers comparing the accuracy of, for instance, classical versus ML water models, but not a corresponding comparison of the speeds of these models.


Here is a recent paper which augments a classical force field with an ML force field to kind of get the best of both worlds (speed and accuracy) to a certain extent. I still don't see a real time comparison though.

$\endgroup$
1
  • 1
    $\begingroup$ +10. A very interesting question indeed! And welcome to the site !!! $\endgroup$ Sep 1, 2020 at 19:57

1 Answer 1

12
$\begingroup$

We performed some timing benchmarks as part of our recent paper, albeit not on molecular dynamics:

"Assessing conformer energies using electronic structure and machine learning methods" Int J Quantum Chem. 2020; 121:e26381

enter image description here

It was a bit controversial, since we compared single-core CPU times and not in batch mode. Once the ML method runs the model, it's faster - although we saw speedups of ~70-100x for both ML and force field methods.

No doubt ML methods are faster on GPU, but tuned force fields are as well.

It seems hard to imagine an ML method that's truly faster than a good implementation of a force field. Most force field terms are intentionally designed to use only a few arithmetic operations (e.g. bonding terms require distance and a harmonic potential). As such, they can be highly optimized for both GPU and CPU implementations.

On the other hand, ML methods can clearly do a more accurate job with empirically fitting (e.g, non-bonded interactions). I suspect more hybrid methods will appear over time.

$\endgroup$
5
  • $\begingroup$ Before I ask a formal question on the site, do you know of any comparisons between polarizable force-fields and machine learning forcefields? I am interested in what is a better use of my time... looking into polarizable ff development, or just cutting to the chase and going ML $\endgroup$
    – B. Kelly
    Jan 17 at 14:41
  • $\begingroup$ I don't, but we'd be happy to help use a polarizable force field with our benchmark set to see a comparison. My initial guess is that ANI is probably more accurate, but a polarizable force field is probably faster. $\endgroup$ Jan 17 at 14:43
  • $\begingroup$ I will shoot you an email when I am in a better position to take you up on that offer, I am interested. The struggle I have with polarizable is it has the same kind of dead end that fixed charge models have. You can only parameterize certain physics. With ML, you basically ignore physics, but that allows you, with extreme care, to capture all of the physics $\endgroup$
    – B. Kelly
    Jan 17 at 14:51
  • 1
    $\begingroup$ I suspect there will be classes of models with, e.g., classical bond and angle terms, and then ML non-bonded terms. For some systems (e.g., water) there are v. careful, accurate physical models. The advantage of a polarizable model is that they include a great deal more than a fixed-charge model. Not perfect (e.g., anisotropy) but generating general, transferrable models without QM is hard. $\endgroup$ Jan 18 at 4:23
  • $\begingroup$ The AIQM1 (nature.com/articles/s41467-021-27340-2) looks about right. Semi-empirical + neural network + dispersion correction. Too bad it is currently too slow (in my opinion) for large systems, but I think AIQM2 will probably be faster $\endgroup$
    – B. Kelly
    Jan 18 at 13:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.