Wisevolve Potential Engine: Machine-Learning Molecular Dynamics at 100 Million Atom-Steps per Second on a Single GPU
Abstract
General-purpose molecular dynamics software supports diverse interatomic potentials and simulation requirements. Within a simulation stage, the loaded potential model and the required calculations typically remain fixed as the atomic configuration evolves. We present the Wisevolve Potential Engine (WPE), which uses this information from GPUMD to select and reuse a GPU execution policy tailored to the model and simulation stage. GPUMD retains simulation control and ownership of the simulation data. On a single NVIDIA RTX 5090, WPE-enabled GPUMD (GPUMD-WPE) achieves a median throughput of 109 million atom-step/s across 106 neuroevolution potential (NEP) workloads. For a 15.1-million-atom diamond Si system, it sustains 142 million atom-step/s. Relative to GPUMD, geometric-mean speedups are 3.49x across these NEP workloads, 3.61x across 16 workloads using NEP with dynamic charges (qNEP), 1.24x across 28 classical-potential workloads, and 10.5x across 56 path-integral molecular dynamics (PIMD) configurations. End-to-end calculations of water structure, lattice thermal conductivity, and PIMD thermal expansion show close agreement with corresponding GPUMD results and published simulation results. A separate 12.5-million-atom Cu shock-compression simulation demonstrates its applicability to large-scale nonequilibrium dynamics. These results support model- and stage-specific specialization as a practical way to accelerate GPU potential evaluation in general-purpose molecular dynamics workflows.
Related
- Sorption Molecular Dynamics of Point Defects in 2D Materials with a Machine Learning Potential
- Machine-Learning Molecular Dynamics Reveals Grotthuss Transport in Concentrated Electrolytes
- Machine learning kinetics from molecular dynamics data
- Strain-Induced BCC-to-FCC Transformation in Tungsten Uncovered by Molecular Dynamics Simulations with Machine Learning Potentials
- Transition from Homogeneous to Domain-Wall-Mediated Polarization Switching in BaTiO3: A Machine-Learning Molecular Dynamics Study
BibTeX
@misc{bu2026wisevolve,
title = {{Wisevolve Potential Engine: Machine-Learning Molecular Dynamics at 100 Million Atom-Steps per Second on a Single GPU}},
author = {Hekai Bu and Zheyong Fan},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.17366},
eprint = {2610.17366},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.17366},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Wisevolve Potential Engine: Machine-Learning Molecular Dynamics at 100 Million Atom-Steps per Second on a Single GPU
AU - Hekai Bu
AU - Zheyong Fan
PY - 2026
DA - 2026/10/12
PB - arXiv
DO - 10.48550/arxiv.2610.17366
UR - https://arxiv.org/abs/2610.17366
AB - General-purpose molecular dynamics software supports diverse interatomic potentials and simulation requirements. Within a simulation stage, the loaded potential model and the required calculations typically remain fixed as the atomic configuration evolves. We present the Wisevolve Potential Engine (WPE), which uses this information from GPUMD to select and reuse a GPU execution policy tailored to the model and simulation stage. GPUMD retains simulation control and ownership of the simulation data. On a single NVIDIA RTX 5090, WPE-enabled GPUMD (GPUMD-WPE) achieves a median throughput of 109 million atom-step/s across 106 neuroevolution potential (NEP) workloads. For a 15.1-million-atom diamond Si system, it sustains 142 million atom-step/s. Relative to GPUMD, geometric-mean speedups are 3.49x across these NEP workloads, 3.61x across 16 workloads using NEP with dynamic charges (qNEP), 1.24x across 28 classical-potential workloads, and 10.5x across 56 path-integral molecular dynamics (PIMD) configurations. End-to-end calculations of water structure, lattice thermal conductivity, and PIMD thermal expansion show close agreement with corresponding GPUMD results and published simulation results. A separate 12.5-million-atom Cu shock-compression simulation demonstrates its applicability to large-scale nonequilibrium dynamics. These results support model- and stage-specific specialization as a practical way to accelerate GPU potential evaluation in general-purpose molecular dynamics workflows.
ER -