Faster than Hookes law? — Wesley 2 days ago

@Wesley It's hard to beat Hooke's law, but I would also say that you can't do any meaningful calculations with purely Hooke's law. You at least have to move to Lennard-Jones or something, which is a bit more computationally expensive. — Tristan Maxson 2 days ago

I tend to agree with Tristan. Especially the Reactive force fields have quite a few additional calculations that are computationally more expensive — lcdumort 2 days ago

Floating point arithmetic isn't likely to be on ML's side yet but its prospects are very interesting. Nobody really thinks of reaxFF when discussing FF's... I have no idea how well optimized the codes are for reaxFF, that is worth alot on its own — Wesley 2 days ago

@Wesley 1) OP didn't say it is faster. 2) Bond stretching (I guess you mean that by "Hooke's law") is not enough to describe the majority of materials, as anharmonic terms, three-, four-body terms, intermolecular terms, Coulomb interaction, etc will all be there. 3) Far biggest part of MM calculation time is the calculation of Coulomb interactions. — Greg 2 days ago

@Wesley "Floating point arithmetic isn't likely to be on ML's side" would you elaborate on this? Most ML algorithms can work with single or half-precision, well optimized for GPUs. How is this worse than any classical MD codes? — Greg 2 days ago

My argument is based on EAM potential vs ML potential. For large solid system, I found EAM to have much faster speed. ML/NN potential are still based on zero kelvin calculation hence some of the finite temperature properties are difficult to get. Final thing optimization of semi-emperical potential such as MEAM/EAM is easy to tweak. — pranav kumar 2 days ago

@Greg count the FLOPS, and wish ML good luck for me. — Wesley 2 days ago

@Wesley While it is impossible to beat Hooke's law in a single force evaluation, in a complete MD simulation one can potentially beat an all-atom Hooke's law simulation, by on-the-fly coarse-graining. One may run a fully atomistic simulation for 1 ps, train a coarse graining model on that, and run coarse grained dynamics from then. When required, one can regenerate fully atomistic snapshots from coarse-grained snapshots of the trajectory with some generative model. — wzkchem5 2 days ago

I am not insinuating doing a simulation with only hooke's law, I am just using as an example, the force applied to one atom, based on an atom bonded directly to it is just hookes law., same with angle. Torsions are slightly harder. Intermolecular, harder, but still, even with LJ and coulomb, pretty simple and not many FLOPS required. I just don't see how ML is going to get forces with fewer FLOPS... If you are training a model on 1ps of all atom, your ML no longer has accuracy supremacy — Wesley 2 days ago