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

Area of Science:

• Materials Science
• Computational Chemistry
• Organic Electronics

Background:

• Identifying singlet fission (SF) materials for organic photovoltaics is computationally expensive.
• Accurate excited-state energetics are crucial for SF material development.

Purpose of the Study:

• To develop a cost-effective and accurate method for identifying SF candidates.
• To accelerate the virtual screening of organic photovoltaic materials.

Main Methods:

• A graph neural network (GNN) named SFMMoE was developed.
• SFMMoE integrates a multiexpert multigating (MMoE) architecture with 2-HOP message passing.
• It fuses molecular graph topology with global molecular descriptors.

Main Results:

• SFMMoE simultaneously predicts five key excited-state properties, including SF-critical thermodynamic criteria.
• The model achieved a mean square error below 0.04 eV across all tasks.
• It outperformed traditional machine learning and GNN baselines.

Conclusions:

• Integrating multitask learning and expert specialization improves prediction accuracy for excited-state energetics.
• SFMMoE enables large-scale, low-cost virtual screening of SF materials with quantum-chemical accuracy.
• An online prediction server is available for broader access.