Interpretable Graph Neural Network for Predicting Transient Electronic Structures of Semiconductor Quantum Dots

Understanding atomic-scale fluctuations in semiconductor quantum dots (QDs) is crucial for optoelectronic material design. We combine ab initio methods and Atomistic Line Graph Neural Networks (ALIGNN) to predict femtosecond time-resolved electronic properties in technologically relevant Cd28Se17X22 QDs (X = Cl, OH). These models reveal weaker vibronic coupling in Cl-passivated QDs, highlighting ligand-dependent electron–phonon interactions. ALIGNN models with ensemble learning trained on only ∼10–17% of available data accurately predict bandgap and gap above the conduction band edge (ΔEgap) (MAE < 2.8 meV) across long MD trajectories. Transfer learning extends accurate electronic structure predictions to new trajectory segments with minimal retraining. The Feature Nullification Analysis framework uniquely links transient electronic properties, especially trap state formation, to atomic environments. While bandgap dynamics depend on localized atomic sites, ΔEgap stems from distributed ones. Such a scalable, atom-resolved methodology efficiently probes long-time-scale quantum dynamics, offering atom-resolved insights for designing optoelectronic nanomaterials.