• model order reduction
• graph neural networks
• autoencoders
• deep learning

The thesis focuses on the proposal of method, based on deep learning, for performing compression of physical systems and the subsequent analysis of the characteristics and proerties of the reduced system. The method uses autoencoding neural networks to produce a compressed representation of the physical system which can be used to derive a new compressed system of equations describing the dynamic. The method is experimentally validated on mass-spring models by using the compressed system to perform their simulation. Moreover, the compressed representations are analyzed to have insights on the role of their variables and how the reconstructed system changes when varying their size. Finally, part of the work focus on trying to produce a model that is able to compress multiple and new systems.