Density-based long-range electrostatic descriptors for machine learning force fields.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning force fields (MLFFs) are crucial for simulating materials.
• Existing short-range descriptors struggle to capture long-range electrostatic interactions.
• Incorporating long-range physics into MLFFs is essential for accurate predictions.

Purpose of the Study:

• To develop a novel long-range descriptor for MLFFs.
• To ensure the descriptor maintains translational and rotational symmetry.
• To integrate long-range electrostatic interactions into atom-centered descriptors.

Main Methods:

• The descriptor is based on an atomic density representation.
• It is designed for straightforward integration into existing machine learning schemes.
• Performance is benchmarked against the long-distance equivariant (LODE) descriptor and message-passing networks.

Main Results:

• In a toy electrostatic model, the descriptor achieved <0.1% error.
• For liquid and rock salt NaCl, it reduced errors by 2-3x compared to short-range descriptors.
• No improvement was observed for solid zirconia, unlike message-passing networks.

Conclusions:

• The proposed descriptor effectively captures long-range electrostatic interactions in certain materials.
• It offers a promising alternative to existing methods for specific applications.
• Further research is needed to address limitations in complex material systems.