DeePMD-kit v3: A Multiple-Backend Framework for Machine Learning Potentials

Area of Science:

• Computational Physics
• Materials Science
• Chemistry

Background:

• Machine learning potentials (MLPs) are crucial for molecular dynamics (MD) simulations.
• Existing software packages often rely on single machine learning frameworks (e.g., TensorFlow), limiting interoperability.
• Previous DeePMD-kit versions faced integration challenges due to framework specificity.

Purpose of the Study:

• Introduce DeePMD-kit version 3 with a multibackend framework.
• Demonstrate enhanced versatility and interoperability for MLPs.
• Facilitate integration with diverse machine learning frameworks and differentiable force fields.

Main Methods:

• Developed a multibackend architecture for DeePMD-kit.
• Integrated support for TensorFlow, PyTorch, JAX, and PaddlePaddle.
• Showcased seamless backend switching and integration capabilities.

Main Results:

• DeePMD-kit version 3 supports multiple ML frameworks, overcoming previous limitations.
• The multibackend architecture enables easy integration with other MLP packages.
• Demonstrated successful integration of differentiable molecular force fields.

Conclusions:

• The multibackend framework significantly enhances the flexibility and interoperability of DeePMD-kit.
• This advancement facilitates the development of complex, cross-framework scientific workflows.
• Broadens the applicability of MLPs in physics, chemistry, and materials science research.