Deep convolutional neural networks for generating atomistic configurations of multi-component macromolecules from...

Area of Science:

• Computational chemistry
• Materials science
• Polymer physics

Background:

• Molecular simulations are limited by computational resources for large systems.
• Coarse-grained (CG) models reduce complexity but lose atomistic detail.
• Reconstructing atomistic detail from CG models is a challenging inverse problem.

Purpose of the Study:

• To introduce an efficient and versatile deep learning method for backmapping multi-component coarse-grained (CG) macromolecules.
• To enable the prediction of atomistic polymer structures from CG configurations.

Main Methods:

• Utilized deep learning, specifically a convolutional neural network (CNN).
• Trained the CNN on structural correlations between atomistic and CG polymer configurations.
• Applied the trained model to predict atomistic structures from CG inputs.

Main Results:

• Demonstrated accurate backmapping for polybutadiene copolymers with varying microstructures (cis-1,4, trans-1,4, vinyl-1,2).
• Showed the methodology's transferability across molecular weights and microstructures.
• Successfully generated diverse, equilibrated polymer configurations by modifying CG chemistry.

Conclusions:

• The deep learning approach provides an efficient and versatile solution for CG model backmapping.
• This method overcomes limitations in simulating large-scale polymer systems at the atomistic level.
• The technique allows for the generation of various polymer microstructures from a single CG input.