Parameter-Free and Electron Counting Satisfied Material Representation for Machine Learning Potential Energy and...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Accurate prediction of material properties is crucial for discovering new materials.
• Developing efficient machine learning potentials requires large datasets and careful parameterization.

Purpose of the Study:

• To introduce a parameter-free volume element representation for machine learning potentials.
• To achieve accurate energy and force fitting for hexagonal boron nitride (BN) and silicon (Si) systems.
• To develop an efficient data generation scheme reducing computational cost.

Main Methods:

• A novel parameter-free volume element representation satisfying the electron counting model.
• Preservation of essential physical symmetries: permutational, translational, and rotational invariance.
• An unbiased perturbation-based dataset construction scheme.

Main Results:

• Achieved low root mean squared errors (RMSE) for potential energy (0.57 meV/atom for BN, 0.43 meV/atom for Si) and forces (59 meV/Å for BN, 36 meV/Å for Si).
• Demonstrated applicability to 3D systems, validated with bulk silicon.
• Developed a continuous population distribution using a training dataset 10x smaller than traditional methods.

Conclusions:

• The proposed method offers accurate and efficient machine learning potentials.
• The parameter-free approach and data generation scheme significantly reduce computational expenses.
• The model's general validity is confirmed through structure optimization, molecular dynamics, and extrapolation tests.