Dual Machine-Learning Prediction of Excited-State Energetics from Ground-State Descriptors and Its Implication on...

Accurately describing excited-state (ES) properties in large condensed-phase systems remains challenging due to structural flexibility and the limited scalability of quantum methods, both crucial for understanding photo-functional processes such as singlet fission (SF). Traditional approaches often struggle to link macroscopic observations with transient, configuration-specific ES dynamics and ensemble behavior in fluctuating environments. To bridge this gap, we introduce a dual machine learning (ML) framework: first, an ML potential trained on AIMD data enables efficient molecular dynamics sampling (MLMD); second, a novel physics-informed ML model (GD-ML) predicts vertical excitation energies with high accuracy using only ground-state descriptors (GDs) related to electronic properties. This framework is computationally far more efficient than performing large-scale TDDFT calculations directly, achieving approximately 56% acceleration. Applied to a tetracene dimer in solution, the framework reveals that toluene and DMF exhibit similar probabilistic distributions of the thermodynamic SF driving force (ΔESF). However, toluene favors conformations with localized molecular orbitals (MOs), opening a kinetic channel that enhances LE/CT state coupling and facilitates SF. In contrast, DMF stabilizes conformations with delocalized MOs, suppressing this channel. These insights offer a mechanistic explanation for the experimentally observed solvent dependence of SF efficiency. Key results underscore how symmetric homodimer heterogenization promotes SF by modulating MO delocalization and charge-transfer contributions to Coulomb/exchange energies, thereby influencing excitation character. Ensemble analysis further informs solvent selection for tailored photoexcitation performance. This work establishes a transformative paradigm for high-throughput exploration of photo-functional processes in complex molecular systems by correlating ES properties with ground-state electronic descriptors.