DeePMD-kit v2: A software package for deep potential models.

Area of Science:

• Computational Physics
• Materials Science
• Quantum Chemistry

Background:

• Machine learning potentials (Deep Potential models) are increasingly used for atomistic simulations.
• Existing software packages require continuous updates to incorporate new features and improve usability.

Purpose of the Study:

• To provide an overview of the current major version of the DeePMD-kit package.
• To highlight its advanced features, technical details, and usability improvements.
• To present a comprehensive procedure for molecular dynamics simulations using DeePMD-kit.

Main Methods:

• Overview of DeePMD-kit's architecture and key components.
• Description of advanced features: DeepPot-SE, attention-based descriptors, tensile property fitting, DP-range correction, DP long range, and GPU support.
• Demonstration of a molecular dynamics simulation workflow and benchmarking of model accuracy and efficiency.

Main Results:

• The current DeePMD-kit version offers enhanced features including DeepPot-SE, improved descriptors, and GPU acceleration.
• Usability is significantly improved through better documentation, GUIs, and APIs.
• Benchmarks demonstrate the accuracy and efficiency of various Deep Potential models within the package.

Conclusions:

• DeePMD-kit is a powerful, versatile, and user-friendly tool for machine learning-based molecular dynamics simulations.
• The package supports a wide range of applications in physics, chemistry, biology, and materials science.
• Ongoing developments promise further advancements in computational atomistic modeling.