Toward Gaussian Process Regression Modeling of a Urea Force Field

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Physical Chemistry

Background:

• The development of accurate and efficient molecular dynamics (MD) force fields is crucial for simulating complex chemical systems.
• Standard ab initio MD methods offer high accuracy but are computationally expensive.
• FFLUX (Force Field based on Quantum chemical topology and Unsupervised learning) is a novel machine-learned force field designed to bridge this gap.

Purpose of the Study:

• To evaluate the performance of the FFLUX force field in modeling urea, a larger and more flexible molecule than previously studied systems.
• To assess the accuracy of FFLUX in predicting molecular geometries and relative energies through geometry optimizations and energy ranking.
• To investigate the impact of geometry-dependent multipole moments on the accuracy of electrostatic energy calculations.

Main Methods:

• Training FFLUX models for urea using B3LYP/aug-cc-pVTZ level of theory.
• Performing FFLUX geometry optimizations on 5 energy minima dimers and 75 random dimers.
• Comparing FFLUX-optimized geometries and relative energies against ab initio references.
• Analyzing the effect of flexible versus fixed multipole moments on electrostatic energy.

Main Results:

• Urea models trained with FFLUX achieved a mean absolute error of 0.4 kJ mol^-1 and a maximum prediction error below 7.0 kJ mol^-1.
• FFLUX successfully recovered the 5 energy minima dimers with root-mean-square deviation below 0.1 Å.
• 68% of random dimers converged to the same qualitative structure as ab initio calculations.
• Energy rankings of FFLUX-optimized dimers closely matched ab initio results, with only one minor crossover.
• The use of flexible multipole moments significantly improved accuracy, preventing errors exceeding two orders of magnitude in electrostatic energy.

Conclusions:

• FFLUX demonstrates high accuracy and reliability in modeling urea, a more complex system than previously tested.
• The force field's ability to perform molecular dynamics with near-quantum accuracy at a lower computational cost is confirmed.
• Geometry-dependent multipole moments are essential for accurate FFLUX simulations, particularly for electrostatic interactions.