brucefan1983/GPUMD

What it solves

GPUMD is a high-performance molecular dynamics (MD) package designed to perform large-scale atomistic simulations of materials. It addresses the need for high efficiency in MD simulations by fully implementing the software on graphics processing units (GPUs), allowing for the use of complex machine-learned potentials (MLPs) without the prohibitive computational cost typically associated with them.

How it works

The software provides two primary executables: gpumd for running simulations and nep for handling neuroevolution potentials (NEPs). It utilizes GPU acceleration via CUDA to compute atomic interactions. The package enables the training and application of NEPs—a specific class of machine-learned potentials that combine high accuracy with low computational overhead—to simulate the behavior of materials across various elements and structures.

Who it’s for

It is intended for researchers in materials science, chemistry, and physics who need to perform efficient, large-scale atomistic simulations using machine-learning-based potentials.

Highlights

• Fully GPU-implemented for maximum efficiency.
• Supports Neuroevolution Potentials (NEP) for high-accuracy, low-cost simulations.
• Compatible with Linux and Windows.
• Capable of handling a wide range of materials, including inorganic and organic materials across 89 elements (via NEP89).
• Supports advanced simulation methods such as path-integral molecular dynamics (PIMD) and hybrid Monte Carlo (MCMD).