CONI-Net: Machine Learning of Separable Intermolecular Force Fields.

Area of Science:

• Computational chemistry
• Molecular modeling
• Soft matter physics

Background:

• Noncovalent interactions (NCIs) are crucial for soft matter and biomolecular systems.
• Ab initio symmetry-adapted perturbation theory precisely analyzes NCIs but is limited to small systems.

Purpose of the Study:

• To develop a scale-bridging approach for deriving intermolecular force fields from ab initio data.
• To preserve the energy decomposition inherent in ab initio methods.
• To enable rapid development of predictive force fields with insights into NCIs.

Main Methods:

• Systematic derivation of an intermolecular force field from ab initio data.
• Application in molecular dynamics simulations of solvents.
• Comparison of predicted thermodynamic observables (mass density, enthalpy of vaporization) with experimental data and existing force fields.
• Investigation of extrapolation capabilities to untrained molecules.

Main Results:

• Promising results were achieved despite using moderately accurate ab initio data.
• The method allows for straightforward data generation and fitting without target data.
• The developed force fields provide insights into the balance of noncovalent interactions.

Conclusions:

• The scale-bridging approach facilitates the rapid development of predictive force fields.
• This method enhances understanding of intermolecular interactions in complex systems.
• The approach offers a valuable tool for molecular simulations in various scientific domains.