Unified differentiable learning of electric response

Area of Science:

• Computational Materials Science
• Machine Learning
• Condensed Matter Physics

Background:

• Predicting material responses to stimuli is crucial but limited by computational costs in current methods.
• Existing simulations often struggle with scaling for large-scale material analysis.

Purpose of the Study:

• To develop an equivariant machine learning framework for accurate prediction of material responses to electric fields.
• To overcome the computational limitations of traditional simulation methods for materials science.

Main Methods:

• Implemented an equivariant machine learning framework based on differential relationships between potential functions and external fields.
• Unified model enforces physical constraints, symmetries, and conservation laws for predicting electric enthalpy, forces, polarization, Born charges, and polarizability.
• Applied the framework to alpha-quartz (α-SiO2) and ferroelectric barium titanate (BaTiO3).

Main Results:

• Demonstrated accurate prediction of vibrational and dielectric properties for α-SiO2.
• Enabled large-scale dynamics simulations under arbitrary electric fields with high accuracy.
• Successfully captured temperature, frequency, and time-dependent ferroelectric hysteresis in BaTiO3, revealing domain switching mechanisms.

Conclusions:

• The equivariant machine learning framework significantly advances the scale and accuracy of computational materials science simulations.
• The method provides unprecedented insights into ferroelectric domain switching and other electric field-driven phenomena.
• This approach paves the way for more efficient and comprehensive material property prediction.