Sparse Gaussian Process Regression-Based Machine Learned First-Principles Force-Fields for Saturated, Olefinic, and...

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Accurate interatomic potentials are crucial for molecular simulations.
• Developing universal potentials for diverse hydrocarbon systems remains challenging.

Purpose of the Study:

• To develop universal machine learning interatomic potentials (ML-IAPs) for saturated, olefinic, and aromatic hydrocarbons.
• To validate the accuracy of these ML-IAPs in predicting molecular structures and energies.

Main Methods:

• Utilized Sparse Gaussian Process Regression for ML-IAP development.
• Trained potentials by combining existing alkane/polyene data with new cyclic/aromatic hydrocarbon data and cross-terms.
• Employed Density Functional Theory (DFT) with PBE + D3 functional for on-the-fly adaptive sampling.

Main Results:

• Successfully generated universal ML-IAPs for a wide range of hydrocarbons.
• ML-IAPs accurately predicted structures and energies for beta-carotene monomer and dimer.
• Simulations of liquid ethylene and toluene crystals showed good agreement with experimental/theoretical data.

Conclusions:

• The developed ML-IAPs offer ab initio accuracy for hydrocarbon systems.
• These potentials show promise for broader applications in organic, polymeric, and biomolecular systems.
• This work represents a significant step towards universal force fields in computational chemistry.