Graph dynamical networks for forecasting collective behavior of active matter

Abstract

After decades of theoretical studies, the rich phase states of active matter and cluster kinetic processes are still of research interest. And how to efficiently calculate the dynamical processes under their complex conditions becomes an open problem. Recently, machine learning methods have been proposed to predict the degree of coherence of active matter systems. In this way, the phase transition process of the system is quantified and studied. In this paper, we use Graph Network as a powerful model to determine the evolution of active matter with variable individual velocities solely based on the initial position and state of the particles. The Graph Network accurately predicts the order parameters of the system in different scale models with different individual velocities, noise and density to effectively evaluate the effect of diverse condition. Compared with the classical physical deduction method, we demonstrate that Graph Network prediction is excellent, which could save significantly computing resources and time. In addition to active matter, our method can be applied widely to other large-scale physical systems.

Physical information-enhanced graph neural network for predicting phase separation

“…[20][21][22][23][24][25] Since the past decade, machine learning has been used to solve classic problems in complex physical systems. [26][27][28][29][30][31][32][33][34] In particular, the graph network-based simulators show extraordinary predictive capabilities for physical systems such as fluids and rigid solids, thanks to the ability of graph networks to effectively capture relational information and structural properties in the system. [35] At the same time, researchers have successfully used the features of GNNs to predict the longterm dynamics of glass systems.…”