... so definitely more than 4.4 but less than 4.6 million? /s

source?

Just expressing my support for typst as well! It’s mature enough to typeset a PhD thesis, has a very mellow learning curve, and has a great community.

Here I was thinking I’d have a hard time convincing people RPA is empirical.

Do quantum monte carlo techniques have true potential or are we stuck with decades-old approximations invented by highly noncomputational scientists?

simple Python API; to drive a single 'master' job which then does everything else!

a golden (😂) PES

Actually, from that perspective, even a 1000x slowdown could be acceptable since it would be used less for super long MDs and more for building models above and beyond atomic-level MD...

From a distance, and this is probably controversial, but it feels like AF has made so much progress that would have otherwise required decades of atomic simulations ?

For drug discovery, do you think more accurate atomic interactions are the way to go, or will people gradually abandon bottom-up atomic-level simulations?

So as long as all possible low-density environments are included in training, putting a limit at fixed cutoff makes sense?

Hmm, yeah, and maybe the fixed neighbors thing is just a trick to speed up training and improve performance on the synthetic benchmarks

At the same time: beyond a “threshold” number of neighbors there is maybe so much screening that the required # neighs to include becomes a constant?

I should really get out of my matsci cave because I am so unaware of these things 😂

I was imagining OpenMM with a bunch of custom force expressions, PME, and anisotropic pressure control (often needed in solid state) at 1 ms / step — and maybe 100 ms / step for MACE for a similar system, approx..

at least in my experience!

Although matbench leading entries usually truncate the number of neighbors to consider to a fixed number, such that the cost of a single message passing layer no longer scales with density …

Right sorry, wasn’t counting bio! Though I think it’s more the increased density rather than shear system size that widens the performance gap?

In matsci / catalysis, with proper enhanced sampling, the main worry is not so much the achievable time scales rather than the accuracy of the QM data…

100x because that’s how much slower an “optimized” ML potential for any particular system would be. I might be optimistic here but a small MACE network and the right training data have always gotten me below ~1 meV/atom and ~50 meV/A errors.

link: arxiv.org/abs/2404.03777