Max Wilson1,Arghya Bhowmik1,Ole Winther1,Tejs Vegge1
DTU1
Max Wilson1,Arghya Bhowmik1,Ole Winther1,Tejs Vegge1
DTU1
Modelling molecular systems, and more specifically their time dynamics, often requires large computational cost associated with the complicated interactions many body problems. There is a growing field for modelling molecular systems with neural networks motivated by the ability of neural network methods, notably graph neural networks to capture correlations and features of systems that previously would have to be modelled and understood analytically. This field extends in the time domain, by predicting the evolution and behaviour of molecular systems including protein relaxation, catalytic reactions, Markov chains in latent spaces, and recurrent neural networks. State-of-the-art neural network methods for modelling time dynamics include temporal variational autoencoders, neural ordinary differential equations, temporal residual networks, and transformers. Here, we combine state-of-the-art a molecular graph modelling framework with hierarchical generative (variational autoencoder) latent space models to open an alternate path to modelling time dynamics in molecular systems.