Prediction of triplet geometries across chemical space with flow-matching based machine learning

Obtaining accurate triplet excited-state geometries is critical for diverse applications with spin-based properties, including photocatalysis, optoelectronics, and photodynamic therapy where triplet energy and structural relaxation govern process efficiency and device performance. Yet, it remains computationally prohibitive for large-scale screening. In this work, we present GoFlow-T1 by adapting a machine learning framework based on flow matching, GoFlow, for the direct prediction of minimum-energy triplet-state geometries across chemically diverse organic molecules. The model is trained on quantum-chemically optimized triplet structures for approximately 11,000 compounds, encompassing chromophores, pharmaceuticals, natural products, and functional dyes. By learning the structural relaxation patterns associated with intersystem crossing and triplet-state stabilization, GoFlow-T1 generates equilibrium, that is minimum-energy, triplet geometries without the need of explicit excited-state geometry optimization, which in turn enables the rapid computation of triplet energies and singlet-triplet energy gaps. A benchmark against time-dependent density functional theory demonstrates near-quantum-chemical accuracy for both structural and energetic observables, while reducing computational cost by several orders of magnitude. Integrating GoFlow into photophysical screening pipelines enables the scalable identification of molecules with optimal triplet state properties for diverse photochemical applications. This work demonstrates that machine learning-driven prediction of excited-state equilibrium geometries is a practical, efficient strategy for high-throughput modeling and the data-driven discovery of functional photoactive candidates.