Chemically Transferable Electronic Coarse Graining for Polythiophenes.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Electronic coarse-graining (ECG) methods show promise for mesoscopic electronic predictions in soft materials.
• A key limitation of current ECG models is their lack of chemical transferability.

Purpose of the Study:

• To develop chemically transferable ECG models for polythiophenes using graph neural networks.
• To assess the impact of coarse-grained representation and training data on model accuracy and transferability.

Main Methods:

• Graph neural networks were employed to train ECG models on diverse polythiophene sequences.
• Models were trained on data encompassing 15 monomer chemistries and varying polymerization degrees.
• The influence of preserving specific atomic coordinates (C-β) on accuracy was investigated.

Main Results:

• The developed ECG models demonstrate chemical transferability across different polythiophene sequences.
• Preserving C-β coordinates in the coarse-grained representation is crucial for accuracy.
• Integrating unique polymer sequences improved performance more than augmenting existing conformational sampling.

Conclusions:

• Chemically transferable ECG models for polythiophenes were successfully developed.
• These models can be efficiently adapted to related properties and higher levels of theory with minimal data.
• The approach provides a foundation for broader chemically transferable ECG predictions across diverse chemical spaces.