Learning Electronic Polarizations in Aqueous Systems.

Area of Science:

• Computational Materials Science
• Statistical Learning in Physics
• Quantum Chemistry

Background:

• The polarization of periodic systems is a discontinuous function of atomic positions, hindering statistical learning.
• Accurate modeling of dielectric properties requires robust polarization prediction methods.

Purpose of the Study:

• To compare two distinct approaches for building statistical learning models of bulk polarization.
• To evaluate model performance for both bulk aqueous systems and the challenging air-water interface.
• To develop a computationally efficient data-driven preprocessing protocol for polarization modeling.

Main Methods:

• Comparison of a point charge model preprocessing method with a Wannier center prediction method.
• Testing models on bulk aqueous systems and the air-water interface.
• Development of a data-driven preprocessing protocol using low-level theory derivatives.

Main Results:

• Both methods perform comparably on bulk systems, with point charge preprocessing slightly superior but requiring more effort.
• The Wannier center approach accurately predicts polarization at the air-water interface without modification.
• The data-driven preprocessing protocol achieves similar accuracy to existing methods without needing explicit point charge models.

Conclusions:

• The proposed data-driven preprocessing strategy effectively models polarization for complex systems and interfaces.
• These training strategies facilitate the construction of accurate polarization models for studying dielectric properties.
• The methods offer a pathway to achieve ab initio accuracy in the study of realistic materials.