Integrating Charge Equilibration with Equivariant Machine-Learning Interatomic Potentials.

Area of Science:

• Materials Science
• Computational Chemistry
• Physics

Background:

• Machine-learning interatomic potentials (MLIPs) excel in modeling atomic interactions but struggle with long-range electrostatic effects and charge transfer.
• Existing MLIPs often fail to accurately capture phenomena crucial for materials properties, such as nonlocal electronic behavior.

Purpose of the Study:

• To augment the Multi-Atomic Cluster Expansion (MACE) potential with a charge equilibration (QEq) framework for self-consistent charge redistribution.
• To evaluate the capabilities and limitations of ML-enhanced QEq in modeling charged defects and transferable potentials.

Main Methods:

• Integration of a charge equilibration (QEq) framework into the equivariant Multi-Atomic Cluster Expansion (MACE) potential.
• Application of the ML-enhanced QEq model to charged oxygen vacancies in wurtzite ZnO and a transferable water potential.

Main Results:

• The ML-enhanced QEq model accurately reproduced charge state-dependent relaxations and migration pathways for ZnO defects in small to medium supercells.
• Limitations of the quadratic QEq formalism were observed in larger ZnO systems, leading to charge delocalization and loss of distinct charge states.
• Initializing long-range water models from pretrained short-range representations improved data efficiency and transferability from gas-phase clusters to bulk liquid.

Conclusions:

• ML-enhanced QEq shows promise for capturing complex defect physics and enabling transferable potentials.
• The study highlights the fundamental constraints of QEq formalisms in large systems and the importance of physically informed architectures.
• Pretrained representations are crucial for extending ML potentials to systems dominated by long-range electrostatics and nonlocal charge response.