Long-Range Interactions in High-Dimensional Neural Network Potentials: A Benchmark Study for Small Organic Molecules.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning potentials (MLPs) commonly use local atomic environments, limiting their accuracy for long-range intermolecular forces.
• Accurate modeling of long-range electrostatic and dispersion interactions is crucial for predicting molecular behavior.

Purpose of the Study:

• To investigate the impact of incorporating electrostatic and dispersion corrections into high-dimensional neural network potentials (HDNNPs).
• To develop and evaluate a novel model, CombineNet, for predicting gas-phase intermolecular interactions.

Main Methods:

• Augmenting HDNNPs with a machine learning-based charge equilibration (QEq) scheme for electrostatics.
• Utilizing the Machine-Learning eXchange-hole Dipole-Moment (MLXDM) model for dispersion corrections.
• Training the model on density functional theory (DFT) data and comparing against CCSD(T)/CBS benchmarks.

Main Results:

• CombineNet achieved a low mean absolute error (MAE) of 0.59 kcal/mol and root-mean-square error (RMSE) of 3.38 meV/atom on the DES370K dataset.
• Minimal basis iterative stockholder (MBIS) charges provided more accurate long-range interaction trends compared to Hirshfeld charges.
• The training set composition is critical, requiring data that covers both dissociation limits and near-cutoff regions.

Conclusions:

• Explicitly including long-range electrostatic and dispersion corrections enhances the accuracy of MLPs for intermolecular interactions.
• The choice of charge model significantly impacts the prediction of electrostatic contributions.
• Careful consideration of training data is essential for developing reliable MLPs for molecular dimers.