An accurate and transferable machine learning potential for carbon.

Area of Science:

• Computational Materials Science
• Machine Learning in Physics
• Atomistic Simulations

Background:

• Accurate atomistic simulations of carbon are crucial for understanding materials properties.
• Existing methods often face trade-offs between accuracy and computational cost.
• Machine learning potentials offer a promising avenue for efficient and accurate simulations.

Purpose of the Study:

• To develop a highly accurate machine learning potential for diverse carbon forms.
• To enable cost-effective atomistic simulations of crystalline, amorphous, and nanostructured carbon.
• To bridge the gap between simulation accuracy and computational efficiency.

Main Methods:

• Construction of a machine learning model using the Gaussian Approximation Potential (GAP) methodology.
• Integration of extensive structural databases for amorphous carbon and graphene, including defect structures.
• Fitting the potential to reference data from optB88-vdW density functional theory (DFT) calculations, including dispersion interactions.

Main Results:

• The developed potential, GAP-20, accurately describes bulk crystalline and amorphous carbon, surfaces, and defects.
• Achieved accuracy approaches that of direct ab initio simulations at a significantly reduced computational cost.
• Rigorous testing confirmed accuracy for lattice parameters, bond lengths, formation energies, and phonon dispersions.

Conclusions:

• GAP-20 successfully combines flexibility for amorphous carbon with high accuracy for crystalline graphene.
• This interatomic potential is suitable for a wide range of applications involving diverse carbon materials.
• Demonstrates the power of machine learning for developing accurate and efficient simulation tools in materials science.