Simulation of physical systems with variants of the Interaction Network

Abstract

[en] The aim of this thesis is to primarily learn to predict trajectories of physical systems by using Machine Learning. We have used the Interaction Network as a base model and introduced tweaks to its structure in the framework of Graph Networks in order to deal with systems without interaction, with force-based interactions and systems governed by the Vicsek model. We have been able to replicate very well systems without interaction and systems governed by a Vicsek model of infinite reach. However the results with force-based systems are mediocre because they need more trainable parameters and training data. The results for the Vicsek model with a finite radius of reach are the worst but we have learned the necessity and the methodology of introducing attention to the base model to deal with this class of problem.