Many-body interactions and deep neural network potentials for water.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning in physics

Background:

• Machine learning potentials (MLPs) offer efficient alternatives to traditional force fields.
• Deep Potential Molecular Dynamics (DeePMD) is a prominent MLP framework.
• Accurate simulation of water properties is crucial across many scientific disciplines.

Purpose of the Study:

• To assess the accuracy and transferability of DeePMD potentials trained on MB-pol data for water systems.
• To identify limitations in DeePMD's ability to capture many-body interactions and physical principles in water.
• To contribute to the development of more robust MLPs for molecular simulations.

Main Methods:

• Detailed assessment of DeePMD potentials against MB-pol reference data.
• Analysis of bulk and interfacial properties of water.
• Evaluation of many-body interactions and adherence to the "nearsightedness of electronic matter" principle.

Main Results:

• DeePMD potentials exhibit limitations in reproducing MB-pol accuracy for various water systems.
• Inherent limitations in transferability and predictive accuracy were observed.
• Incomplete implementation of the "nearsightedness" principle and lack of long-range electric field representation contribute to these limitations.

Conclusions:

• DeePMD potentials face a "short-blanket dilemma," balancing computational efficiency with physical rigor.
• Current MLPs, including DeePMD, require improved methods for representing long-range interactions.
• The study provides insights for advancing ML models in simulating water and other condensed matter systems.