Characterizing the Dynamics of Active Matter using Graph Neural Networks

Abstract

Biological systems often have self-organizing properties. On the microscopic scale, the self-organization may be driven by constituents that generate their own motility by expending energy. Systems made up of such constituents, that self-propel, are termed active matter. Via this definition, it is clear that it is important to un derstand the rules that govern the microscopic constituents of an active matter in order to understand the active matter itself. Even though active matter is a topic of interest today, finding good methods of qualitatively and quantitatively charac terizing the constituents of active matter remains an issue. Coincidentally, different types of artificial neural networks (ANN) have, in recent years, been used increas ingly in research settings with great success. One such network is called the graph neural network (GNN). As the name suggests, this network is specifically designed to work with graphs as input data. Graphs can act as a useful representation of a system of particles, including active matter systems. Therefore, this project aims to characterize the forces that underpin an active matter system consisting of inter acting particles that also have an active component, using a special type of GNN called message passing network (MPN). This was done by creating a rudimentary simulation of this type of active matter, and training an MPN on it using standard machine learning algorithms. In the end, the simulations were found to give rise to characteristic active matter phenomena, and the MPN was able to correctly predict the force dynamics of a particle in the given active matter system.