Deep Learning for Collective Variables (DeepCV)
What is DeepCV?
DeepCV is a deep learning framework for training an autoencoder model on molecular representations to learn collective variables (CVs). DeepCV implements the deep learning autoencoder neural network (DAENN) which makes use of the sophisticated symmetric feedforward multilayer perception to extract the latent space of the data and smartly generates non-linear CVs. Our algorithm also adopts a newly developed nuclear-based descriptor called eXtended Social PeRmutation INvarianT (xSPRINT) and applies the customization of loss function with the min-max game for smartly learning unexplored regions in the configurational space.
The recent works show that the CVs generated by DAENN/DeepCV can successfully capture slow-mode rare event of chemical reaction corresponding to unexplored metastable states on the configurational space. Moreover, combinding DAENN/DeepCV CVs with metadynamics give accurate free energy surface (FES) of a set of chemical reactions, yielding accurate thermodynamic properties, such as free energy of activations and heat of reaction.
