Abstract
The efficiency of atomic simulations of materials and molecules can rapidly deteriorate when large free energy barriers exist between local minima. We propose smooth basin classification, a universal method to define reaction coordinates based on the internal feature representation of a graph neural network. We achieve high data efficiency by exploiting their built-in symmetry and adopting a transfer learning strategy. We benchmark our approach on challenging chemical and physical transformations, and show that it matches and even outperforms reaction coordinates defined based on human intuition.