DP Compress: A Model Compression Scheme for Generating Efficient Deep Potential Models.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine-learning interatomic potential energy surface (PES) models offer speed advantages over traditional electronic structure methods.
• Deep neural network-based PES models, like Deep Potential (DP), face challenges with computational cost and efficiency compared to empirical force fields.

Purpose of the Study:

• To introduce DP Compress, a model compression scheme designed to improve the performance of DP models.
• To accelerate DP-based molecular dynamics simulations and reduce their memory footprint.

Main Methods:

• DP Compress is an efficient postprocessing technique applied after DP model training (DP Train).
• The scheme integrates several DP-specific compression strategies.
• The method was tested on Cu, H2O, and Al-Cu-Mg systems.

Main Results:

• DP Compress achieves an order of magnitude speedup in DP-based molecular dynamics simulations.
• The compression scheme reduces memory consumption by an order of magnitude.
• The accuracy of DP Compress was validated across various physical properties for the tested systems.

Conclusions:

• DP Compress effectively enhances the efficiency of Deep Potential models without compromising accuracy.
• The compression technique is applicable to both CPU and GPU computing environments.
• DP Compress is a valuable tool for accelerating molecular modeling research and is publicly available.