Creating Gaussian process regression models for molecular simulations using adaptive sampling.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning

Background:

• Classical force fields lack the accuracy of quantum mechanics for certain chemical problems.
• Quantum Chemical Topology (QCT) provides an atom-focused approach to analyzing molecular properties.

Purpose of the Study:

• To introduce FFLUX, a new force field combining quantum mechanical accuracy with computational speed.
• To develop efficient and accurate predictive models for atomic energies using Gaussian process regression.

Main Methods:

• Utilizing Quantum Chemical Topology (QCT) for parameter-free topological atoms.
• Employing Gaussian process regression (kriging) for atomic energy predictions.
• Implementing an adaptive sampling technique (maximum expected prediction error) for model training.

Main Results:

• Achieved high accuracy in atomic energy predictions (sub-kJ/mol for small molecules, sub-kcal/mol for N-methylacetamide).
• Developed data-compact and efficient kriging models capable of handling large molecular distortions.
• Created an automated Python pipeline (ICHOR) for streamlined model training.

Conclusions:

• FFLUX offers a computationally efficient and accurate alternative to traditional force fields.
• The developed kriging models are suitable for molecular simulations, even with significant molecular distortions.
• The ICHOR pipeline simplifies the application of FFLUX in computational chemistry research.