Intermolecular Non-Bonded Interactions from Machine Learning Datasets.

Area of Science:

• Computational chemistry
• Materials science
• Polymer science

Background:

• Accurate molecular dynamics (MD) simulations of polymer systems require precise determination of intermolecular non-bonded interactions.
• Balancing accuracy and computational cost in force field development remains a significant challenge.
• Representing calculated energy data as a continuous force function is a key difficulty in force field modeling.

Purpose of the Study:

• To develop a general-purpose intermolecular non-bonded interaction force field for organic polymers using machine learning.
• To address the challenge of accurately and efficiently modeling non-covalent interactions in polymer systems.

Main Methods:

• Utilized a well-documented ab initio dataset (SOFG-31) as the training set.
• Employed the CLIFF kernel type machine learning scheme to predict interaction energies.
• Focused on heterodimers selected from the SOFG-31 dataset for training and testing.

Main Results:

• Achieved overall errors well below the chemical accuracy threshold of approximately 1 kcal/mol.
• Demonstrated the effectiveness of machine learning in predicting intermolecular interaction energies.
• Validated the feasibility of the developed machine learning scheme for force field modeling.

Conclusions:

• Machine learning techniques show significant promise for developing accurate and computationally efficient force fields for polymer simulations.
• The developed force field effectively models intermolecular non-bonded interactions in organic polymers.
• This approach offers a viable solution to the accuracy-cost trade-off in molecular dynamics force field development.