Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MDcraft -- a modern molecular dynamics simulation package with machine learning potentials support

I. S. Galtsov, R. V. Muratov, G. V. Vyskvarko, S. A. Murzov, S. A. Dyachkov, P. R. Levashov

TL;DR

MDcraft presents a modular, scalable molecular dynamics platform that unifies classical and machine-learning interatomic potentials behind a high-level Python API and a performant C++ core. It advances HPC MD through hybrid MPI+thread parallelization and a Voronoi-based dynamic domain decomposition, enabling efficient moving-window simulations and large-scale shock studies. The work provides concrete integrations of BPNN, MTP, and DeePMD potentials, with supporting tooling and examples, and demonstrates capabilities via Argon and Aluminum shock Hugoniot calculations, moving-window workflows, and cross-validation against established codes and experiments. Collectively, MDcraft offers an accessible, scalable framework for integrating cutting-edge ML potentials into MD workflows, broadening the practical reach of accurate atomistic simulations on modern clusters.

Abstract

Molecular dynamics is widely used to study various phenomena, such as diffusion, shock wave propagation, and plasma dynamics. A wide range of software packages supports the expanding scope of molecular dynamics applications. However, the quality of simulations depends on force field approximations, ranging from simple models to direct quantum solutions. Recently, machine learning approaches for constructing accurate interatomic potentials have received significant attention. In MDcraft, we integrate these advances into a scalable, physically accurate framework. MDcraft is a comprehensive, modern molecular dynamics platform. It offers a high-level Python API with a user-friendly, script-based interface. The core simulation algorithms are implemented in C++ to ensure robustness and computational efficiency. MDcraft is built for high-performance computing on modern clusters and supports dynamic domain decomposition and load balancing via the Message Passing Interface (MPI) for scalable parallelization. Additionally, MDcraft leverages multithreading within nodes through standard C++ parallelism, enabling efficient use of heterogeneous architectures. We demonstrate the code's capabilities through several examples, including the shock response in aluminum, the shock Hugoniot in argon, and the cold curve of copper.

MDcraft -- a modern molecular dynamics simulation package with machine learning potentials support

TL;DR

MDcraft presents a modular, scalable molecular dynamics platform that unifies classical and machine-learning interatomic potentials behind a high-level Python API and a performant C++ core. It advances HPC MD through hybrid MPI+thread parallelization and a Voronoi-based dynamic domain decomposition, enabling efficient moving-window simulations and large-scale shock studies. The work provides concrete integrations of BPNN, MTP, and DeePMD potentials, with supporting tooling and examples, and demonstrates capabilities via Argon and Aluminum shock Hugoniot calculations, moving-window workflows, and cross-validation against established codes and experiments. Collectively, MDcraft offers an accessible, scalable framework for integrating cutting-edge ML potentials into MD workflows, broadening the practical reach of accurate atomistic simulations on modern clusters.

Abstract

Molecular dynamics is widely used to study various phenomena, such as diffusion, shock wave propagation, and plasma dynamics. A wide range of software packages supports the expanding scope of molecular dynamics applications. However, the quality of simulations depends on force field approximations, ranging from simple models to direct quantum solutions. Recently, machine learning approaches for constructing accurate interatomic potentials have received significant attention. In MDcraft, we integrate these advances into a scalable, physically accurate framework. MDcraft is a comprehensive, modern molecular dynamics platform. It offers a high-level Python API with a user-friendly, script-based interface. The core simulation algorithms are implemented in C++ to ensure robustness and computational efficiency. MDcraft is built for high-performance computing on modern clusters and supports dynamic domain decomposition and load balancing via the Message Passing Interface (MPI) for scalable parallelization. Additionally, MDcraft leverages multithreading within nodes through standard C++ parallelism, enabling efficient use of heterogeneous architectures. We demonstrate the code's capabilities through several examples, including the shock response in aluminum, the shock Hugoniot in argon, and the cold curve of copper.

Paper Structure

This paper contains 16 sections, 26 equations, 13 figures, 1 table.

Table of Contents

  1. MDcraft package description
  2. Data structures
  3. Hybrid parallel algorithms
  4. Dynamic domain decomposition with load balancing
  5. Neighbor lists construction
  6. Molecular dynamics
  7. Potentials support
  8. Performance
  9. Molecular dynamics with machine learning potentials
  10. Molecular dynamics with Behler-Parrinello neural networks
  11. Software implementations
  12. Applications
  13. Moving window simulation
  14. Argon shock Hugoniot: system of isotherms and a series MW simulations
  15. Comparison of aluminum shock Hugoniot in the MW using different EAM potentials
  16. ...and 1 more sections

Figures (13)

  • Figure 1: (a) Basic Atom data structure suitable for a large number of applications. (b) C++ implementation of the Atom data structure. (c) Access fields of the Atom data structure using Python.
  • Figure 2: (a) A distributed memory system scheme consisting of 4 computing nodes communicating via MPI. (b) A shared memory system scheme within a computing node with 2 multicore CPUs. (c) The process manager (thread pool) usage scheme for handling array elements in parallel.
  • Figure 3: (a) Decomposition using a Voronoi diagram with data exchange layers. (b) Voronoi domain containing atoms handled by a single MPI process.
  • Figure 4: Neighbor list construction. Virtual Cartesian grid covers a computational domain.
  • Figure 5: Performance benchmark results: exectution time on a single node with varying numbers of MPI-tasks (left), and scalability from 1 to 8 nodes (right).
  • ...and 8 more figures