Enhanced sampling of structural phase transformations using a neural network based path collective variable

In-depth understanding of the kinetics and mechanisms of rare events in complex systems requires robust sampling of the high-dimensional phase space and the exploration of associated free energy surfaces. In this work, we combine enhanced sampling techniques, such as driven adiabatic free energy dynamics and metadynamics, with a path collective variable defined in a global classifier space. The global classifiers are determined based on local structural environments that are identified using a classification neural network. We demonstrate that the proposed scheme can efficiently sample transformation between different crystalline phases in metallic tungsten and reproduce the free energy landscape.