Predicting trajectory behaviour via machine-learned invariant manifolds

  • 2021-07-21 15:30:38
  • Vladimír Krajňák, Shibabrat Naik, Stephen Wiggins
  • 1


In this paper we use support vector machines (SVM) to develop a machinelearning framework to discover the phase space structure that can distinguishbetween distinct reaction pathways. The machine learning model is trained usingdata from trajectories of Hamilton's equations but lends itself for use inmolecular dynamics simulation. The framework is specifically designed torequire minimal a priori knowledge of the dynamics in a system. We benchmarkour approach with a model Hamiltonian for the reaction of an ion and a moleculedue to Chesnavich consisting of two parts: a rigid, symmetric top representingthe $\text{CH}_3^{+}$ ion, and a mobile $\text{H}$ atom. We begin withtrajectories and use support vector machines to determine the boundariesbetween initial conditions corresponding to different classes of trajectories.We then show that these boundaries between different classes of trajectoriesapproximate invariant phase space structures of the same type observed inearlier analyses of Chesnavich's model. Our approach is designed withextensions to higher-dimensional applications in mind. SVM is known to workwell even with small amounts of data, therefore our approach is computationallybetter suited than existing methods for high-dimensional systems and systemswhere integrating trajectories is expensive.


Quick Read (beta)

loading the full paper ...