Abstract
<jats:p>The concept of triplet energy has long served as the operative predictor of energy transfer reactivity in photochemistry, yet its reliable computational determination remains challenging for many organic systems. The recently introduced dynamic vertical triplet energy (DvTE) framework recasts triplet energy as a statistical ensemble of instantaneous vertical gaps sampled from molecular dynamics trajectories, achieving mean absolute errors of 1.7 kcal/mol against experimental data, but at a computational cost that limits its broader application. Here, DvTE-ML is introduced, which extends machine-learned interatomic potentials to dynamic, ensemble-based prediction of excited-state properties. Built by fine-tuning the MACE-OMOL foundation model, DvTE-ML incorporates path integral molecular dynamics to capture nuclear quantum effects explicitly, alongside an iterative filtering workflow that addresses a state-error contamination risk intrinsic to open-shell training data. Applied to a diverse set of 52 organic molecules, including biologically relevant systems previously intractable at the reference level of theory, DvTE-ML matches the accuracy of the original DvTE framework. Beyond this validation, higher-resolution sampling reveals that gap distributions can deviate from the normal form previously assumed, with vibrational anharmonicity producing skewed distributions, providing additional insights into the mechanisms of triplet energy transfer.</jats:p>