DeePMD-GNN: A DeePMD-kit Plugin for External Graph Neural Network Potentials

Area of Science:

• Computational chemistry and materials science
• Development of advanced simulation tools
• Machine learning in scientific modeling

Background:

• Machine learning potentials (MLPs) offer efficient and accurate atomic interaction predictions, impacting drug discovery, catalysis, and materials design.
• Current MLP software lacks interoperability, hindering consistent benchmarking and requiring separate interfaces with molecular dynamics (MD) software.

Purpose of the Study:

• To introduce DeePMD-GNN, a plugin extending DeePMD-kit to support external graph neural network (GNN) potentials.
• To enable seamless integration of GNN models (NequIP, MACE) within DeePMD-kit.
• To facilitate the use of GNN models in combined quantum mechanical/molecular mechanical (QM/MM) applications.

Main Methods:

• Development of the DeePMD-GNN plugin for the DeePMD-kit framework.
• Integration of popular GNN potentials (NequIP, MACE) into the DeePMD-kit ecosystem.
• Implementation of the range-corrected ΔMLP formalism for QM/MM applications.

Main Results:

• DeePMD-GNN successfully integrates external GNN potentials within DeePMD-kit.
• The plugin supports GNN models in QM/MM simulations.
• Benchmark calculations were performed for NequIP, MACE, and DPA-2 models under consistent training conditions.

Conclusions:

• DeePMD-GNN enhances the interoperability of machine learning potentials in molecular simulations.
• The plugin provides a unified framework for GNN models and QM/MM applications.
• This work facilitates more consistent benchmarking and broader application of advanced MLPs in scientific research.