Automated scientific discovery from complex systems via deep learning

Abstract

With the advent of the large-scale dataset and advanced machine learning, automated algorithms began to play a central role in the progress of science. In this thesis, we address the problem of scientific discovery in an automated manner with deep neural networks. We show that various scientific concepts, such as hidden interaction rules between agents in complex systems and conserved quantities hidden in dynamical systems can be autonomously discovered by employing deep learning and analyzing the trained model. At this point when AI’s performance continues to improve rapidly, our research demonstrates how AI can benefit the domain of natural sciences and scientists.