Speaker
Description
Rare events such as chemical reaction are transitions across an energy barrier. To rationalize the mechanism of such rare events and calculating their reaction rates, the search for an optimal reaction coordinates has long been very active field of research.[1] Using the minimum energy path (MEP) as reaction coordinate simply neglects entropic contributions and dynamic effects. Alternatively, the committor function offers a complementary perspective in which the transition state is defined as an isocommittor surface at which trajectories reach the reactant or product with equal probability.
I will present our findings on the reaction mechanism of the cis-trans isomerization of retinal in gas phase and various solvation environments.[2] We apply the recently developed AIMMD algorithm,[3] where the committor of isomerization reaction is learned by a Neural Network (NN) function from short, dynamically unbiased trajectories that form a transition path ensemble (TPE) and are generated by two-way shooting molecular dynamics (MD) simulations. MD simulations are performed using self-developed machine-learned interaction potentials (MLIPs) and adaptively trained universal MLIPs to provide ab initio accuracy for comparatively low computational costs.[4] The low-dimensional approximation of committor function is obtained through symbolic regression, which yields a human-interpretable reaction mechanism in terms of a small set of internal degrees of freedom.
We observe a significant deviation of the TPE from underdamped Langevin dynamics with time scales in the range of hundreds of femtoseconds compared to the MEP obtained from overdamped simulations in the nanoseconds regime. Overdamped simulations do not show the asymmetric distribution of kinetic energy along different internal coordinates at various stages of the isomerization reaction, as seen in the TPE. Such differences must be taken into account when calculating accurate reaction rates with respect to experimental results, e.g., transition state theory.[5]
[1] Chen, H.; Roux, B.; Chipot, C., J. Chem. Theory Comput. 2023, 19, 4414–4426
[2] Ghysbrecht, S.; Donati, L.; Keller, B. G., J. Comput. Chem. 2025, 46, e27529
[3] Jung, H.; Covino, R.; Arjun, A.; Leitold, C.; Dellago, C.; Bolhuis, P. G.; Hummer, G., Nat. Comput. Sci. 2023, 3, 334–345
[4] Käser, S.; Vazquez-Salazar, L. I.; Meuwly, M.; Töpfer, K., Digit. Discovery 2023, 2, 28–58
[5] Ghysbrecht, S.; Keller, B. G., J. Comput. Chem. 2024, 45, 1390–1403