SFMMoE: A Semi-Empirical Descriptor-Augmented Multi-Expert Graph Neural Network for Accurate Prediction of Singlet Fission Properties

Efficient identification of singlet fission (SF) candidates remains a significant challenge in the development of high-performance organic photovoltaic materials due to the high computational cost of accurately evaluating excited-state energetics. Here, we present SFMMoE, a graph neural network (GNN) that integrates a multiexpert multigating (MMoE) architecture with 2-HOP message passing. By integrating local topological information from molecular graphs with global molecular descriptors derived from semiempirical methods, SFMMoE enables simultaneous prediction of five key excited-state properties, including two thermodynamic criteria critical to SF: Δ<i>E</i>_gap1 = Δ<i>E</i>_S1 - 2Δ<i>E</i>_T1 and Δ<i>E</i>_gap2 = Δ<i>E</i>_T2 - 2Δ<i>E</i>_T1. The model achieves a mean square error below 0.04 eV across all tasks, outperforming traditional machine learning and testing of GNN baselines. This work demonstrates that integrating multitask learning with expert specialization and graph-descriptor feature fusion substantially improves the prediction accuracy of excited-state energetics, enabling large-scale, low-cost virtual screening of SF materials with quantum-chemical accuracy. To facilitate broader access, a freely available and user-friendly online prediction server is provided at http://tech.iawnix.xyz/SFMMoE.