Generative Modeling of Entangled Polymers with a Distance-Based Variational Autoencoder.

Area of Science:

• Computational chemistry
• Polymer science
• Machine learning

Background:

• Generating accurate polymer configurations is crucial for understanding material properties.
• Traditional molecular dynamics (MD) simulations can be computationally expensive for sampling diverse polymer structures.

Purpose of the Study:

• To develop a deep learning framework for learning and generating polymer configurations.
• To accelerate the sampling of uncorrelated polymer structures.

Main Methods:

• Utilized a variational autoencoder (VAE) with convolution and attention layers.
• Trained the VAE on coarse-grained molecular dynamics data of polyethylene.
• Employed multidimensional scaling and short MD simulations for postprocessing.

Main Results:

• The VAE effectively encodes polymer structural patterns into a low-dimensional latent space.
• Generated polymer configurations reproduce key observables like energy, size, and entanglement.
• The VAE-based approach significantly reduces the computational time for sampling structures.

Conclusions:

• The proposed VAE framework offers a computationally efficient alternative for generating polymer configurations.
• This method holds potential for accelerating materials discovery and design in polymer science.