qNEP: A Highly Efficient Neuroevolution Potential with Dynamic Charges for Large-Scale Atomistic Simulations.

Area of Science:

• Computational materials science
• Machine learning in physics
• Quantum chemistry

Background:

• Traditional methods for simulating electrostatics are computationally expensive.
• This limits large-scale, long-time simulations of phenomena driven by electrostatics.
• Existing machine learning potentials often struggle to incorporate electrostatics efficiently.

Purpose of the Study:

• To develop a computationally efficient machine learning framework for atomistic simulations incorporating electrostatics.
• To enable accurate prediction of dielectric properties, infrared spectra, and field-matter coupling.
• To extend the neuroevolution potential (NEP) to a charge-aware framework (qNEP).

Main Methods:

• Introduced explicit, environment-dependent partial charges into the NEP framework, creating qNEP.
• Represented each ionic partial charge using a neural network dependent on local descriptors.
• Derived consistent expressions for forces and virials accounting for charge position dependence.
• Implemented qNEP in the GPUMD package with Ewald summation and particle-particle particle-mesh support.

Main Results:

• qNEP accurately predicts Born effective charge tensors and polarization.
• Demonstrated accuracy and efficiency on water, Li7La3Zr2O12, BaTiO3, and a magnesium-water interface.
• Achieved scalable simulations of million-atom systems on nanosecond timescales using GPUs.

Conclusions:

• qNEP enables accurate atomistic simulations with explicit long-range electrostatics.
• The framework is suitable for studying dielectric response, infrared activity, and field-matter coupling.
• qNEP offers a computationally efficient alternative for large-scale electrostatic simulations.