Optimized symmetry functions for machine-learning interatomic potentials of multicomponent systems.

Area of Science:

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

Background:

• Accurate potential energy landscapes are crucial for materials simulations.
• Current machine learning methods face computational scaling challenges with increasing chemical complexity.
• Efficient methods are needed for multicomponent systems.

Purpose of the Study:

• To develop a computationally efficient machine learning approach for potential energy landscapes.
• To address the unfavorable scaling of existing methods with the number of chemical species.
• To enable accurate simulations of complex multicomponent materials.

Main Methods:

• Utilized optimized symmetry functions to identify structural similarities in multicomponent systems.
• Achieved linear computational complexity.
• Combined symmetry functions with charge equilibration via neural network (CE-NN) for ionic materials.
• Applied the method to alkali-halide (MX) systems with six chemical species.

Main Results:

• The proposed method demonstrates linear computational scaling.
• Achieved good agreement with experimental and density functional theory (DFT) data.
• Accurately predicted physical and structural properties for various alkali-halide combinations.
• Successfully studied multicomponent systems with up to six chemical species.

Conclusions:

• The novel approach offers a computationally efficient and accurate way to model potential energy landscapes.
• Optimized symmetry functions combined with CE-NN are effective for ionic materials.
• This method significantly advances the simulation capabilities for complex multicomponent systems.