Does a Machine-Learned Potential Perform Better Than an Optimally Tuned Traditional Force Field? A Case Study on...

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Quantum Chemistry

Background:

• Accurate molecular modeling requires reliable potentials to capture complex interactions.
• Machine learning potentials (MLPs) offer a promising alternative to traditional force fields (FFs).
• Evaluating these models is crucial for advancing computational chemistry methods.

Purpose of the Study:

• To comparatively evaluate the performance of ANI-2x (MLP), GAFF (FF), and a tuned GAFF-like FF.
• To assess their accuracy in predicting energetic, geometric, and conformational properties of γ-fluorohydrins.
• To benchmark their ability to model condensed-phase behavior and nuclear spin-spin coupling constants.

Main Methods:

• Comparative analysis of ANI-2x, GAFF, and a tuned GAFF-like force field.
• Benchmarking against quantum-mechanical data for gas-phase and chloroform simulations.
• Validation using experimental nuclear spin-spin coupling data in chloroform.

Main Results:

• ANI-2x exhibited issues with hydrogen bonding, overstabilization of minima, and dispersion interactions.
• Conventional force fields demonstrated continued relevance, particularly for condensed-phase simulations.
• Both MLPs and FFs showed varying degrees of accuracy across different properties and phases.

Conclusions:

• ANI-2x shows potential but requires refinement for specific interactions and condensed phases.
• Traditional force fields remain essential tools, especially for complex condensed-phase systems.
• This study provides insights for developing improved molecular models and guiding their application.