Learning the forces in active matter from the trajectories: a Graph-neural-network approach

Active particles exhibit complex collective phenomena that emerges from their local interactions. To model such systems, one would usually propose some inter-particle interactions and active forces, simulate the dynamics of a system with many individual elements and finally compare the results with experiments via, for instance, an order parameter. However, not only choosing one order parameter might introduce a bias, but also it is difficult to assess how well the model describes the experimental system. In our work we suggest a completely different approach. What if we could learn the inter-particle interactions and the active forces directly from the data? We propose a graph-neural-network-based scheme that learns the interactions between particles and the active forces to predict the correct particle dynamics. After training the network, one can extract both passive and active interactions between particles and use them (analytically or numerically) to make new predictions or unravel dynamical features of experiments of active particles.