A Learning Framework for Atomic-Level Polymer Structure Generation.

Area of Science:

• Materials Science
• Computational Chemistry
• Polymer Science

Background:

• Synthetic polymers are crucial for energy, electronics, consumer goods, and medical applications.
• Current polymer development faces lengthy design cycles and challenges in generating realistic 3D atomic structures.
• Existing generative models do not adequately address synthetic polymers due to representation and data constraints.

Purpose of the Study:

• To introduce polyGen, a novel generative model for the on-demand creation of realistic 3D polymer structures.
• To address the limitations of traditional methods in simulating polymer conformational diversity.
• To enable faster and more accurate design of synthetic polymeric materials.

Main Methods:

• Developed polyGen, a generative model utilizing graph-based encodings and a latent diffusion transformer with positional biased attention.
• Incorporated joint training with small molecule data to augment the limited dataset of DFT-optimized polymer structures.
• Established structure matching criteria for benchmarking the model's performance on polymer structure generation.

Main Results:

• polyGen successfully generates realistic and diverse 3D atomic structures for linear and branched polymers.
• The model demonstrates promising performance even for polymers with large repeat units.
• Achieved a significant advancement in atomic-level polymer structure generation, capturing intrinsic flexibility.

Conclusions:

• polyGen overcomes limitations of existing crystal structure prediction methods for synthetic polymers.
• This model represents a transformative capability for material structure generation, enabling accelerated design.
• The approach facilitates the creation of diverse polymer conformations from minimal input, such as repeat unit chemistry.