Molecular force fields with gradient-domain machine learning: Construction and application to dynamics of small...

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Molecular Dynamics

Background:

• Accurate molecular force fields are crucial for simulating molecular behavior.
• Traditional methods struggle with complex electronic interactions and high-dimensional potential energy surfaces.
• Quantum mechanical calculations are accurate but computationally expensive for large-scale simulations.

Purpose of the Study:

• To construct accurate molecular force fields for small molecules using the symmetrized gradient-domain machine learning (sGDML) approach.
• To demonstrate the ability of sGDML to reconstruct complex potential-energy surfaces from limited data.
• To explore the insights gained from sGDML molecular dynamics simulations.

Main Methods:

• Utilized the symmetrized gradient-domain machine learning (sGDML) approach for force field construction.
• Generated molecular conformations from ab initio molecular dynamics trajectories.
• Computed atomic forces using high-level coupled-cluster theory with single, double and perturbative triple excitations [CCSD(T)].

Main Results:

• Successfully reconstructed high-dimensional potential-energy surfaces with high accuracy using sGDML.
• The sGDML model captured diverse electronic interactions (e.g., H-bonding, proton transfer, lone pairs, hybridization changes, steric repulsion, n → π* interactions) without prior assumptions.
• Analysis of sGDML molecular dynamics trajectories provided new insights into molecular dynamics and spectroscopy.

Conclusions:

• The sGDML approach offers an efficient and accurate method for developing molecular force fields for small molecules.
• This data-efficient machine learning model enables the use of high-accuracy quantum mechanical data for simulations.
• sGDML simulations provide near-spectroscopic accuracy, advancing the understanding of molecular dynamics and properties.