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

Conjugated organic polymers are the cornerstone of modern organic electronics, yet accurate prediction of their properties remains a challenging task due to their synthetic complexity and high computational cost of quantum-chemical methods. Here, we develop a graph neural network based on the DimeNet++ direct message passing architecture to predict HOMO, LUMO and energy gaps of conjugated polymers directly from their 3D monomer structure. The model was pre-trained on TD-DFT-extrapolated data and trained on a limited dataset of experimentally measured properties. As a result, pre-training had significantly improved model's accuracy compared to direct training (MAEs ~0.3 eV vs. 0.074 eV, 0.141 and 0.172 for HOMO/LUMO and energy gap, respectively). Pre-training on monomer DFT data did not provide comparable gains. The results demonstrate that polymer-relevant pre-training is critical for capturing structure-property relationships and enables accurate predictions without delta-learning or prior quantum-chemical calculations, facilitating efficient screening and rational design of conjugated polymers for organic optoelectronics.