Modeling molecular trajectories using long short-term memory with stacked state pooling

Abstract

Atomistic molecular dynamics can be used for simulating large molecular systems with great accuracy. The downside to this is that simulations are computationally expensive and thus take a long time to perform. Many times, one is only interested in a subset, or a summarizing statistic, of the information available in the results of the simulations. This raises the question whether it is necessary to run full molecular dynamics simulations when one is interested in only partial outputs or if similar results can be achieved with a cleverly designed machine learning algorithm. In this thesis, an attempt to answer this question is made by proposing a deep learning model for solving general many-body problems. The model generates long sequences of particle positions by predicting one time step at a time, using its outputs as new inputs. It is demonstrated that this model is capable of running simple molecular dynamics. Suggestions for further experiments to measure the capability of the model to generalize to complex many-body problems are presented.