Universal machine learning for the response of atomistic systems to external fields.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning potentials enable efficient molecular simulations for closed systems.
• External fields significantly alter chemical structure and reactivity but are often excluded from ML models.

Purpose of the Study:

• To develop a universal machine learning model that incorporates external field effects into atomic interaction potentials.
• To enable accurate simulations of molecular and periodic systems under external electric fields.

Main Methods:

• Introduction of a universal field-induced recursively embedded atom neural network (FIREANN) model.
• Integration of pseudo field vector-dependent features into atomic descriptors for rotational equivariance.
• Training the model using atomic forces to address polarization issues in periodic systems.

Main Results:

• FIREANN successfully represents system-field interactions and correlates response properties (dipole moment, polarizability) with field-dependent potential energy.
• The model demonstrates suitability for spectroscopic and dynamics simulations in both molecular and periodic systems.
• FIREANN overcomes the multiple-value polarization issue in periodic systems by training on forces alone.

Conclusions:

• The FIREANN method offers a universal and efficient approach for first-principles modeling of complex systems in strong external fields.
• This model significantly advances the capability of machine learning in simulating systems under external field influence.