The energy landscape of folding in n-C14H30 described by a machine-learned potential.

Folding and unfolding in molecules as simple as short hydrocarbons and as complicated as large proteins continue to be an active research field. Here, we investigate folding in n-C14H30 using both density functional theory (DFT)/B3LYP calculations of 27 772 local minima and a kinetic transition network calculated for a previously reported potential energy surface (PES) obtained by fitting roughly 250 000 B3LYP energies. In addition to generating a database of minima and the transition states that connect them, these calculations and the PES based on them have been used to develop a simple and accurate model for the energy landscape. The model for the local minima is based on the number of gauche torsions as well as their interactions with neighbors and next-nearest neighbors, resulting in three parameters, which are fitted using the direct DFT results. The transition states are governed by 13 parameters based on differences between the two connected local minima. The model predicts that there are 44 530 local minima (not counting permutation-inversion isomers) connected by 525 028 transition states, including degenerate rearrangements. When compared to the actual stationary points, it achieves a minimum absolute energy error of 43 cm-1 for the local minima and 47 cm-1 for the transition state barriers connecting levels of different energy. The model also provides accurate predictions of the most kinetically relevant folding and unfolding pathways, for example, from very highly excited configurations to the global minimum. In addition, it facilitates the determination of a disconnectivity graph using the Cambridge Energy Landscapes programs.