Abstract

Excited-state dynamics govern how photoexcited systems transform light into chemical change, but direct simulation remains prohibitively expensive; even DFT can require days per picosecond of dynamics, while higher-accuracy excited-state methods are rarely practical for routine studies. This talk highlights how machine learning can relieve this bottleneck and make excited-state dynamics accessible at scale. We present X-MACE, an excited-state ML framework built on the MACE message-passing architecture with an integrated diagonalization scheme implemented through an autoencoder, designed to improve modelling of excited-state potential energy surfaces. The emphasis is on accelerating nonadiabatic dynamics, enabling surface-hopping workflows through rapid prediction of excited-state energies and related quantities, and moving toward excited state dynamics screening. As a case study, we outline preliminary dynamics guided design of the GFP chromophore, commonly used for bioimaging. We map key relaxation pathways and explore how targeted substitutions at specific positions can increase the fluoresce of the GFP chromophore. Overall, the approach aims to bring excited-state dynamics into high-throughput regimes and support screening for photophysical outcomes such as quantum yields in increasingly complex molecular systems.