Fine-Tuning Directional Message Passing Neural Networks: Predicting Properties of Conjugated Organic Polymers with...

Area of Science:

• Materials Science
• Computational Chemistry
• Organic Electronics

Background:

• Conjugated organic polymers are essential for organic electronics.
• Predicting their properties is difficult due to complexity and high computational costs.
• Accurate property prediction is crucial for efficient material design.

Purpose of the Study:

• To develop a graph neural network (GNN) model for predicting HOMO, LUMO, and energy gaps of conjugated polymers.
• To investigate the impact of pre-training strategies on model accuracy.
• To enable faster screening and design of novel conjugated polymers.

Main Methods:

• Utilized a graph neural network based on the DimeNet++ architecture.
• Developed a model to predict properties directly from 3D monomer structures.
• Employed pre-training on TD-DFT-extrapolated data and training on experimental data.

Main Results:

• Pre-training on polymer data significantly improved prediction accuracy (MAEs of ~0.074 eV, 0.141 eV, 0.172 eV for HOMO, LUMO, and energy gap, respectively).
• Direct training without pre-training resulted in lower accuracy (MAE ~0.3 eV).
• Pre-training on monomer DFT data did not yield comparable improvements.

Conclusions:

• Polymer-relevant pre-training is critical for accurate structure-property relationship prediction.
• The developed GNN model enables efficient property prediction without prior quantum-chemical calculations.
• This facilitates the rational design and screening of conjugated polymers for organic optoelectronics.