Quantum-Trained AI Enables Inverse Design of Organic Frontier Orbitals at Billion-Scale.

Frontier orbital energies govern charge transfer, energy-level alignment, and optoelectronic performance in organic molecules, yet their prediction across vast chemical spaces remains challenging because they arise from coupled effects of functionality, conjugation, and topology. In this study, we develop an interpretable AI framework for frontier orbital prediction using QM9 quantum-chemical dataset and validate it across conventional and deep learning models. A direct message passing neural network delivers the best performance for frontier orbital prediction, while interpretability analyses consistently recover chemically meaningful substructures linked to orbital modulation. Extension of the optimized DMPNN model to 908 545 821 molecules in GDB-13 database enables AI-guided ultra large-scale donor screening framework development, referencing frontier orbital alignment to a representative ITIC acceptor for solar cells. This identifies only 37 visible region candidates, revealing the rarity of simultaneously satisfying donor-like level alignment and narrow optical gaps. Scaffold-resolved analysis further unravels diamino-triketone class as a privileged donor building-unit motif, combining small positive frontier orbital offsets with narrow gaps and favorable energy alignment. These show that quantum-trained AI can unify predictive accuracy, chemical interpretability, and billion-scale screening, providing a practical route for inverse design of organic optoelectronic molecules.