A Data-Efficient Machine-Learning Approach for Modeling the Photodynamics of All-trans Hexatriene based on Multireference Configuration Interaction Calculations

Abstract

Accurate simulations of excited-state photodynamics of polyenes remain challenging due to the high computational cost of multireference electronic-structure methods required to describe the coexistence of ionic and covalent states. In this work, active learning together with machine learning interatomic potentials (MLIP) trained on multireference configuration interaction with single and double excitations (MR-CISD) level energies and gradients is used to enable efficient and accurate nonadiabatic molecular dynamics for all-trans-hexatriene. A compact yet representative training set is constructed by reducing an initial pool of ~850k configurations to ~3k geometries while preserving the regions of configurational space relevant for the dynamics. The resulting model reproduces the topology of the excited-state potential energy surfaces and yields a physically consistent relaxation mechanism, characterized by a sequential S₂ → S₁ → S₀ pathway and time constants in good agreement with experimental data. Analysis of representative trajectories provides mechanistic insight into the structural factors governing nonadiabatic transitions, including the key role of the terminal C=C bond elongation and identifies a notably broader and more diffuse structural distribution associated with the S₁ → S₀ transition in comparison to the more localized S₂ → S₁ decay pathway. Analysis of the torsional angles around the C=C double bonds reveals preferential isomerization at the central C₃=C₄ bond, providing semi-quantitative evidence for the initiation of cis-trans photoisomerization in the hot ground state. The MLIP simulations trained at high-level MR-CISD level are extremely efficient, enabling 200 trajectories running up to 800 fs in short time, whereas direct on-the-fly multireference dynamics simulations would be practically impossible at this MR level.