Development of scalable and generalizable machine learned force field for polymers.

Area of Science:

• Computational Chemistry
• Materials Science
• Polymer Science

Background:

• Predicting polymer properties is crucial for designing new materials.
• Classical force fields often struggle with specific polymer systems like polyethylene glycol.
• Accurate modeling is needed for advanced polymer applications.

Purpose of the Study:

• To develop a more accurate potential energy surface for polymer simulations.
• To create a Charge Recursive Neural Network (QRNN) model for polymer property prediction.
• To enable large-scale molecular dynamics simulations of polymeric systems.

Main Methods:

• Training a QRNN model on DFT calculations of smaller atomic clusters.
• Utilizing active learning to iteratively improve the model.
• Implementing a novel training approach with semi-empirical partial charges.
• Validating the model by simulating ethylene glycol oligomers.

Main Results:

• The QRNN model generalizes well to larger atomic clusters and longer polymer chains.
• The model produces stable molecular dynamics simulation trajectories.
• The simulated polymer chain dynamics show excellent agreement with experimental data.
• The QRNN approach allows simulations of significantly larger systems than DFT.

Conclusions:

• The developed QRNN model offers a more accurate force field than classical methods.
• This work provides a promising method for large-scale molecular simulations of polymers.
• The QRNN model enhances the ability to tailor polymer molecules for specific applications.