Benchmarking JoyAI: Performance Evaluation of the JOYSpectra AI Module

Modeling lanthanide spectroscopy relies on the accurate determination of the ligand singlet (S1) and triplet (T1) excited-state energies, which are essential parameters for calculating non-radiative energy transfer rates. Although Density Functional Theory (DFT) provides reliable excitation energies, its high computational cost limits its application in large-scale screening studies. To address this challenge, JoyAI, a graph-based Artificial Intelligence model integrated into the JOYSpectra platform, was developed to rapidly predict excited-state energies. In this work, JoyAI performance was systematically evaluated using molecular geometries reoptimized with four fast semiempirical methods: GFN2-xTB, Ln-xTB, PM6-Sparkle, and RM1-Sparkle. The predicted S1 and T1 energies obtained from these geometries were compared with reference values calculated using the M06/def2-TZVP DFT methodology. The results showed errors ranging from 5% to 10%, demonstrating that semiempirical geometries provide sufficient accuracy while significantly reducing computational cost for large-scale lanthanide spectroscopy applications.