The predictions of kinetics and mechanisms of complex chemical reactions are severely limited by the high dimensionality of potential energy surfaces (PES). While most reactive processes are rationalized in terms of low-dimensional structures, such as internal reaction coordinates, these structures are obtained by means of energy minimizations and transition state (TS) searches on the full, high-dimensional PES. It is the strong and nonlinear coupling between reactive and nonreactive degrees of freedom (DOFs) that prevents dimensionality reduction by linear methods, for example, principal component analysis (PCA). Here, we propose a method for constructing a low-dimensional geometric structure within the PES, which locally approximates the reactive DOFs, specifically, bond breaks and bond formations. Due to the nonlinear coupling between reactive and nonreactive DOFs, this structure cannot be simply described as a subspace of the internal coordinate space R3n-6 of an n-atomic system but as its submanifold. Generally, a manifold as a subset of points in the Euclidean space RD that has a local one-to-one mapping to a lower-dimensional Euclidean space Rd. The manifold of distance displacements (DD) is defined algorithmically using a random walk in R3n-6 projected locally onto the set of interatomic distances. Since this construction does not rely on energy evaluations or TS searches, it can be carried out at minimal computational cost. We show that DD manifolds have low local dimension d^loc, about an order of magnitude smaller than the PES dimension d^PES or the number of interatomic distances d^int. Moreover, their dimension increases only slowly with the number of degrees of freedom. DD manifolds are thus convenient low-dimensional approximations to study chemical reaction dynamics using the methods of manifold learning.

Discovering low-dimensional manifolds within potential energy surfaces

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.