Conformational free energy surface of cyclooctane from metadynamics in the collective variable space of autoencoder neural network

For rare event problems in which the important free energy basins are separated by large barriers, enhanced sampling methods provide the means to perform simulations within tractable timescales. Metadynamics simulation, which is one of the widely used enhanced sampling methods, requires the definition of a set of collective variables for accumulating the bias potentials. An important aspect of the collective variables is their dimensionality because the efficiency of the method decreases exponentially with the dimensionality. We present here a methodology of incorporating the codes from an autoencoder neural network as the collective variables for metadynamics simulations. This dimensionality reduction of an eight-dimensional space of dihedral angles into a three-dimensional space of features enables the computation of the conformational free energy surface of cyclooctane.