New AI Model Fast-Tracks Search for Better Solar Materials

The quest for more efficient organic solar cells faces a major hurdle: the slow and costly process of finding molecules capable of a performance-boosting behaviour called singlet fission (SF). Accurately calculating the necessary energy properties of countless candidates is a significant computational bottleneck.

Now, researchers have developed SFMMoE, a specialised graph neural network (GNN) that dramatically accelerates this discovery process. The AI uses a clever 'multi-expert' organisation, where different parts of the network specialise in different tasks. It uniquely fuses two kinds of data: detailed local information from a molecule’s structure and broader 'global' properties derived from faster, semi-empirical methods.

This integrated approach allows SFMMoE to predict five key energy-state properties simultaneously with remarkable precision, achieving an error rate below 0.04 electron-volts. By outperforming previous machine learning models, it enables vast, low-cost virtual screening of new materials with an accuracy that rivals quantum-chemical calculations. A free online tool has also been released to facilitate wider research.