Learning Latent Representations to Bridge Coarse-Grained and Atomistic Resolutions in Polymer Simulations

Area of Science:

• Computational Materials Science
• Polymer Physics
• Machine Learning Applications

Background:

• Simulating polymer chain conformations at high fidelity (e.g., united-atom) is computationally expensive.
• Coarse-grained (CG) models simplify polymers but lose structural detail.
• Bridging the gap between CG and high-fidelity models is crucial for efficient simulations.

Purpose of the Study:

• To develop a machine learning framework for reduced-order representations of polymer conformations.
• To enable accurate reconstruction of high-fidelity united-atom (UA) configurations from coarse-grained (CG) simulations.
• To establish prerequisites for accelerated polymer dynamics simulations.

Main Methods:

• Utilized linear singular value decomposition and nonlinear autoencoders to create latent space representations.
• Developed a linear mapping between CG and UA latent spaces for back-mapping.
• Employed molecular dynamics relaxation to correct minor structural inaccuracies.

Main Results:

• Achieved compressed, low-dimensional representations of polymer conformations with minimal structural accuracy loss.
• Demonstrated a near-perfect linear mapping between CG and UA latent spaces.
• Successfully reconstructed high-fidelity UA configurations from CG simulations, with inaccuracies corrected by MD relaxation.

Conclusions:

• The framework provides a compact and accurate latent encoding of polymer conformations.
• A validated multi-fidelity mapping allows reconstruction of UA structures from CG configurations.
• This hybrid machine learning-physics approach accelerates polymer simulations and enables efficient super-resolution back-mapping.