betaFeedback

WorksPhysics and Astronomy

FiuMD: Heterogeneous Batching for Coarse-Grained Molecular Dynamics with Neural Network Potentials

Abstract

FiuMD is a GPU molecular dynamics engine that advances many independent biomolecular simulations as a single batch, even when the systems differ in size, topology, force-field parameters, periodic box, timestep, or thermostat parameters. All systems are concatenated into global tensors, so the classical interaction terms and a shared TorchMD-Net neural network potential are evaluated for the whole batch at once, while each system keeps its own integration and thermostat. FiuMD started as a fork of TorchMD and reworks its system representation, force evaluation, and integration to handle such heterogeneous batches. We test it with coarse-grained Cαα models of chignolin, Trp-cage, BBL, and λλ-repressor. In double precision, forces within a heterogeneous batch agree with isolated FiuMD and TorchMD evaluations to within the precision of repeated evaluations, with a maximum relative difference of 7×10−167\times10^{-16} across 60 systems. Over 1000 steps without a thermostat, batched trajectories deviate by at most 2.5×10−112.5\times10^{-11} Angstrom from isolated ones, even when periodic images lie within the neural network cutoff. With a different target temperature for each system of the ensemble, the mean and width of each system's kinetic temperature match the values expected from its own thermostat. We show that when individual simulations are too small to fill the GPU, batching can improve throughput several-fold over TorchMD. On a single GPU, batching four different protein systems achieves a 4.6-7.8-fold speedup per step compared with sequential TorchMD simulations. Compared with sequential homogeneous FiuMD batches, heterogeneous batching provides up to a 4-fold speedup for smaller batches while maintaining comparable performance for larger ones.

Paper details

vlnc:W_azmcn9fq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 5

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…