Gaussian Process Regression Models for Predicting Atomic Energies and Multipole Moments.

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Quantum Chemistry

Background:

• Developing accurate and efficient force fields for molecular simulations is challenging.
• Traditional methods often face a trade-off between computational speed and predictive accuracy.
• Quantum theory of atoms in molecules (QTAIM) provides accurate atomic properties but is computationally expensive.

Purpose of the Study:

• To introduce FFLUX, a novel computational approach for developing accurate molecular force fields.
• To demonstrate the capability of Gaussian process regression (GPR) in achieving *ab initio* accuracy for atomic properties.
• To validate the FFLUX training pipeline on a diverse set of molecules.

Main Methods:

• Utilized Gaussian process regression (GPR) for predicting atomic energies and multipole moments.
• Employed an in-house FFLUX training pipeline with active learning strategies.
• Trained models on six representative molecules, including peptide-capped amino acids, glucose, paracetamol, aspirin, and ibuprofen.
• Simulated molecular dynamics using AMBER-GAFF2 to generate diverse molecular configurations.

Main Results:

• Generated successful GPR models for all tested molecules, achieving high accuracy.
• Models required approximately 2000 training points due to active learning, enhancing efficiency.
• Achieved prediction errors below 1 kcal mol^-1 for validation sets.
• Demonstrated successful transfer learning between different atomic properties.
• Intermolecular electrostatic interactions were predicted with root-mean-square errors below 0.1 kJ mol^-1.

Conclusions:

• FFLUX successfully balances speed and accuracy in force field development using GPR.
• The FFLUX pipeline provides a robust method for generating accurate atomic property predictions.
• This approach has significant implications for improving the efficiency and reliability of molecular simulations.