A foundation machine learning potential with polarizable long-range interactions for materials modelling.

Area of Science:

• Computational chemistry
• Materials science
• Artificial intelligence in science

Background:

• Long-range interactions significantly influence chemical system behavior.
• Accurate modeling of these interactions is crucial for predicting material properties.
• Existing methods often struggle with capturing long-range physics and polarization effects.

Purpose of the Study:

• To develop a foundation framework integrating explicit polarizable long-range physics with an equivariant graph neural network potential.
• To create a versatile model trained across the periodic table for broad applicability in materials science.
• To enhance the accuracy and efficiency of computational materials modeling.

Main Methods:

• Integration of explicit polarizable long-range physics with an equivariant graph neural network (GNN) potential.
• Employment of a physically motivated polarizable charge equilibration scheme optimizing electrostatic interaction energies.
• Training the foundation model across the periodic table up to Plutonium (Pu).

Main Results:

• The model effectively captures challenging long-range interactions beyond traditional message-passing mechanisms.
• Accurate reproduction of polarization effects under external electric fields.
• Demonstrated strong performance in predicting mechanical properties, ionic diffusivity, ferroelectric transitions, and interface dynamics.

Conclusions:

• The developed foundation model offers a powerful and efficient approach to materials modeling.
• It successfully balances accuracy and computational efficiency for diverse chemical systems.
• The model can be efficiently finetuned for high-level accuracy on specific challenging materials science problems.