Transferable machine learning interatomic potential for carbon hydrogen systems

Area of Science:

• Computational Materials Science
• Machine Learning in Chemistry
• Artificial Neural Networks for Interatomic Potentials

Background:

• Accurate modeling of carbon-hydrogen (C-H) systems is crucial for materials discovery.
• Traditional methods can be computationally expensive for large-scale atomistic simulations.

Purpose of the Study:

• To develop and validate a machine learning interatomic potential for C-H systems.
• To assess the accuracy and transferability of the developed artificial neural network (ANN) potential.

Main Methods:

• An ANN interatomic potential was trained using data from density functional theory (DFT) calculations.
• The potential was evaluated on various C-H systems (0D-3D), chemical processes, and lattice dynamics.
• Phonon dispersion analysis was used to verify predictions of lattice dynamics.

Main Results:

• The ANN potential demonstrated high accuracy and transferability in predicting geometries and formation energies.
• Accurate prediction of lattice dynamics was confirmed, essential for crystal structure stability.
• Efficient force constant calculations enabled exploration of energy landscapes, leading to the discovery of a novel carbon polymorph.

Conclusions:

• The developed ANN potential offers a robust and versatile tool for precise atomistic simulations of C-H materials.
• This machine learning approach significantly advances computational materials science research.
• The potential facilitates efficient exploration of complex systems and discovery of new materials.