Sensitivity Analysis in Photodynamics: How Does the Electronic Structure Control <i>cis</i>-Stilbene Photodynamics?

The techniques of computational photodynamics are increasingly employed to unravel reaction mechanisms and interpret experiments. However, misinterpretations in nonadiabatic dynamics caused by inaccurate underlying potentials are often difficult to foresee. This work focuses on revealing the systematic errors in the nonadiabatic simulations due to the underlying potentials and suggests a thrifty approach to evaluate the sensitivity of the simulations to the potential. This issue is exemplified in the photochemistry of <i>cis</i>-stilbene, where similar experimental outcomes have been differently interpreted based on the electronic structure methods supporting nonadiabatic dynamics. We examine the predictions of <i>cis</i>-stilbene photochemistry using trajectory surface hopping methods coupled with various electronic structure methods (OM3-MRCISD, SA2-CASSCF, XMS-SA2-CASPT2, and XMS-SA3-CASPT2) and assess their ability to interpret experimental observations. While the excited-state lifetimes and calculated photoelectron spectra show consistency with experiments, the reaction quantum yields vary significantly: either completely suppressing cyclization or isomerization. Intriguingly, analyzing stationary points on the potential energy surface does not hint at any major discrepancy, making the electronic structure methods seemingly reliable when treated separately. We show that performing an ensemble of simulations with different potentials provides an estimate of the electronic structure sensitivity. However, this ensemble approach is costly. Thus, we propose running nonadiabatic simulations with an external bias at a resource-efficient underlying potential (semiempirical or machine-learned) for the sensitivity analysis. We demonstrate this approach using a semiempirical OM3-MRCISD method with a harmonic bias toward cyclization.