Dual-Attention Graph Neural Network Capturing Intramolecular Noncovalent Interactions for Precise Prediction of Bond Dissociation Energies in Organic Optoelectronic Materials

Intramolecular noncovalent interactions (Intra-NIs) have emerged as key determinants of structure–property relationships in molecular systems, exerting profound influences on material performances of organic optoelectronic materials. In the development of organic light-emitting diode (OLED) materials, Intra-NIs have recently been recognized as an effective strategy for enhancing bond dissociation energies (BDEs), thereby improving molecular stability and device lifetimes. Accurate and rapid prediction of BDEs of molecules featuring pronounced Intra-NIs is therefore essential for accelerating the development of robust OLED materials. However, existing machine learning approaches for predicting BDEs primarily focus on short-range local properties, while failing to adequately capture long-range Intra-NI effects. Herein, we report a dual-attention graph neural network model for the accurate and rapid prediction of BDEs affected by Intra-NIs. By integrating through-bond and through-space attention mechanisms and incorporating fresh molecular dispersion matrix as an additional input to encode geometric information, the model achieves high accuracy for C–N BDEs, with a mean absolute error below 0.05 eV, while reducing computational cost by 4 orders of magnitude relative to density functional theory calculations. Furthermore, the model exhibits strong transferability to C–P and C–S bonds via transfer learning, demonstrating its potential applicability to other fragile bonds in diverse organic optoelectronic materials. This work establishes a robust paradigm for incorporating physically meaningful Intra-NI information into data-driven molecular property predictions, providing a powerful tool for the iterative optimization and high-throughput virtual screening of next-generation high-performance optoelectronic materials across a wide range of applications.