Machine learning photodynamics reveals the role of solvent and pressure on the [2+2]- cycloadditions toward cubanes

Photochemical [2+2]-cycloaddition of [3]-ladderdienes enables promising access to highly strained molecules, cubanes. Gas phase machine learning (ML) photodynamics showed that non-covalent interactions increase the chemoselectivity towards cubanes and decrease competing electrocyclic ring-opening pathways. However, these simulations omit the solvent and pressure, which could be a controlling factor for the [2+2]-photocycloaddition. We now extend our ML-photodynamics approach with a two-layer ONIOM scheme combining the neural network potential and GFN-xTB to understand effects of the solute-solvent interactions and external pressure on the [2+2]-photocycloadditions of octamethyl (1), octatrifluoromethyl (2), and octacyclopropyl (3) [3]-ladderdienes. We simulate the experiments reported in n-pentane and n-perfluoropentane with 2000 trajectories. The 1ps ML photodynamics featuring explicit solvation indicate an increasing non-radiative S1→S0 transition rate. The S1 half-life decreases from 232 to 111 fs in 1, 477 to 283 fs in 2, and 92 to 74 fs in 3. The trajectories of solvent models 1–3 show little differences in the quantum yields of cubanes (1%, 12%, and 17%) compared to the previously reported gas-phase results (1%, 14%, and 15%). Then, we simulate the high-pressure droplets by applying external forces to the solvents. By reducing 50% of the volumes, the S1 half-life is further shortened to 86 fs, 188 fs, and 68 fs in models 1–3, respectively. Meanwhile, the predicted yields of cubanes increase to 2%, 18%, and 21%. Thus, our findings suggest a promising control to the [2+2]-photocycloaddition of [3]- ladderdienes toward the cubane by harnessing high pressure.