Accelerating Fourth-Generation Machine Learning Potentials Using Quasi-Linear Scaling Particle Mesh Charge...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning potentials (MLPs) offer electronic structure accuracy at lower computational cost for atomistic simulations.
• Current MLPs often sum atomic energies based on local environments.
• Nonlocal phenomena like charge transfer necessitate charge equilibration (Qeq) in fourth-generation MLPs, increasing computational expense.

Purpose of the Study:

• To present a highly efficient formulation of charge equilibration (Qeq) for machine learning potentials.
• To overcome the computational bottleneck associated with traditional Qeq methods in large-scale simulations.
• To enable efficient calculation of energy derivatives that account for global structure-dependent atomic charges.

Main Methods:

• Developed a novel Qeq formulation that avoids explicit Coulomb matrix element computation.
• Implemented a quasi-linear scaling algorithm for the Qeq step.
• Enabled efficient calculation of energy derivatives considering Qeq-derived atomic charges.

Main Results:

• Achieved a quasi-linear scaling computational cost for the Qeq procedure.
• Significantly reduced the computational bottleneck for large systems in MLPs.
• Successfully computed energy derivatives incorporating global structure-dependent atomic charges.

Conclusions:

• The new Qeq formulation drastically improves the efficiency of MLPs requiring charge equilibration.
• This method accelerates atomistic simulations involving nonlocal phenomena like charge transfer.
• The approach is general and applicable to various force fields beyond MLPs.