GPUMD

Official Resources

• Homepage: https://gpumd.org/
• Source Repository: https://github.com/brucefan1983/GPUMD
• Documentation: https://gpumd.org/
• License: GPL-3.0

Overview

GPUMD (Graphics Processing Units Molecular Dynamics) is a highly efficient molecular dynamics package fully implemented on GPUs. It features the neuroevolution potential (NEP) approach for machine learning potentials, enabling accurate and fast simulations of thermal transport properties.

Scientific domain: Molecular dynamics, thermal transport, machine learning potentials
Target user community: Researchers studying thermal transport using MD with ML potentials

Theoretical Methods

• Classical molecular dynamics
• Neuroevolution potentials (NEP)
• Green-Kubo thermal conductivity
• Homogeneous non-equilibrium MD (HNEMD)
• Spectral decomposition of thermal conductivity
• Heat current autocorrelation

Capabilities (CRITICAL)

• GPU-accelerated MD simulations
• NEP machine learning potentials
• Thermal conductivity calculations
• Spectral thermal conductivity
• HNEMD method
• Green-Kubo method
• Phonon participation ratio
• Modal analysis

Key Strengths

GPU Acceleration:

• Fully GPU-native implementation
• Orders of magnitude speedup
• Large system sizes feasible
• Efficient memory usage

NEP Machine Learning:

• Neuroevolution potential training
• Near-DFT accuracy
• Fast evaluation
• Active learning support

Thermal Transport:

• Multiple methods (GK, HNEMD)
• Spectral decomposition
• Modal contributions
• Accurate predictions

Inputs & Outputs

• Input formats:

  • xyz structure files
  • NEP potential files
  • run.in control file

• Output data types:

  • Thermal conductivity
  • Heat current
  • Spectral properties
  • Trajectory files
  • Thermodynamic properties

Interfaces & Ecosystem

• NEP training: Built-in potential training
• LAMMPS: Some compatibility
• ASE: Python interface available
• calorine: Python package for NEP

Advanced Features

• NEP training: Built-in neuroevolution potential development
• HNEMD method: Homogeneous non-equilibrium MD for thermal conductivity
• Spectral decomposition: Frequency-resolved thermal conductivity
• Modal analysis: Phonon mode contributions to transport
• Multi-GPU support: Scalable to large systems
• Active learning: Efficient training data selection

Computational Cost

• NEP training: Hours to days (one-time cost)
• MD simulations: Very fast on GPU (millions of atoms)
• Thermal conductivity: Minutes to hours depending on convergence
• Overall: Orders of magnitude faster than DFT-MD

Performance Characteristics

• Speed: Extremely fast on GPU
• System size: Millions of atoms feasible
• Accuracy: NEP provides DFT-level accuracy
• Parallelization: Multi-GPU support

Limitations & Known Constraints

• Requires NVIDIA GPU
• NEP training requires expertise
• Classical MD limitations apply
• Learning curve for NEP development

Application Areas

• Thermal conductivity of complex materials
• Disordered and amorphous systems
• Nanostructured materials
• High-temperature properties
• Materials with strong anharmonicity

Comparison with Other Codes

• vs LAMMPS: GPUMD is fully GPU-native; LAMMPS has GPU packages but CPU-centric design
• vs Phono3py/ShengBTE: GPUMD uses MD-based methods; others use perturbation theory
• vs DeepMD-kit: Both support ML potentials; GPUMD has built-in NEP, DeepMD uses DP
• Unique strength: Integrated NEP training + GPU MD + thermal transport in one package

Best Practices

NEP Training:

• Use diverse training dataset
• Include thermal expansion data
• Validate phonon dispersions
• Test force/energy predictions

MD Simulations:

• Equilibrate system thoroughly
• Use sufficient system size
• Run long enough for convergence
• Monitor temperature stability

Thermal Conductivity:

• Use both GK and HNEMD methods
• Check size convergence
• Validate with experiments
• Analyze spectral contributions

Community and Support

• Open-source GPL-3.0
• Very active development (Zheyong Fan)
• Comprehensive documentation
• Tutorial examples included
• Growing user community
• Regular updates and new features

Verification & Sources

Primary sources:

1. GitHub: https://github.com/brucefan1983/GPUMD
2. Z. Fan et al., Comput. Phys. Commun. 218, 10 (2017)
3. Z. Fan et al., Phys. Rev. B 104, 104309 (2021) - NEP

Confidence: VERIFIED

Verification status: ✅ VERIFIED

• Source code: OPEN (GitHub, GPL-3.0)
• Documentation: Comprehensive
• Active development: Very active
• Academic citations: >500