Classification of regular and chaotic motions in Hamiltonian systems with deep learning
Alessandra Celletti, Cătălin Galeş, Víctor Rodríguez-Fernández, Massimiliano Vasile
Abstract
This paper demonstrates the capabilities of convolutional neural networks (CNNs) at classifying types of motion starting from time series, without any prior knowledge of the underlying dynamics. The paper applies different forms of deep learning to problems of increasing complexity with the goal of testing the ability of different deep learning architectures at predicting the character of the dynamics by simply observing a time-ordered set of data. We will demonstrate that a properly trained CNN can correctly classify the types of motion on a given data set. We also demonstrate effective generalisation capabilities by using a CNN trained on one dynamic model to predict the character of the motion governed by another dynamic model. The ability to predict types of motion from observations is then verified on a model problem known as the forced pendulum and on a relevant problem in Celestial Mechanics where observational data can be used to predict the long-term evolution of the system.