E(3)-VERDE: A 3D-Equivariant Neural Network for the Complete Ground and Excited- State Redox Landscape of Organic Photocatalysts

Photoredox catalysis enables mild visible-light-driven chemical transformations; the rational design of organic photocatalysts is difficult because it requires ground- and excited-state redox potentials challenging to obtain by experiment or quantum chemistry. We introduce E(3)-VERDE, an E(3)-equivariant neural network trained on the Virtual Excited State Reference for the Discovery of Electronic materials database (VERDE materials DB) to predict ground- and excitedstate redox potentials together with zero-zero excitation energies for organic photocatalysts from a single low-cost ground-state geometry, without requiring excited-state geometry optimizations at inference. Its three-dimensional representation retains molecular shape and atomic arrangement as predictive information for these properties. On held-out DFT test data, the model reproduces reference values with R 2 of 0.97-0.99 and MAEs of 0.04-0.07 eV across the predicted redox potential and excitation-energy properties. On an experimental benchmark of 40 structurally diverse chromophores including dihydrophenazines, d ihydroacridines, cyanoarenes, N-phenyl carbazoles, N–N donors, acridinium salts, and Eosin Y derivatives, E(3)-VERDE achieves MAEs of 0.09-0.15 eV and lower errors than 2D-fingerprint, distance-only, and partially angular baselines on this benchmark. These results support E(3)-VERDE as a practical and efficient tool for high-throughput screening of organic photocatalysts from low-cost ground-state geometries.