Learn the Dynamics

The laws of nature are described as dynamical systems. Learning dynamics, as a new AI technique, faces the following challenges: (1) the existing pure deep learning models are not that reliable in the scientific field, i.e., they can fit the data but whether it can generalize is still unknown; (2) its training is slow and unstable due to the intrinsic slowly solvable and unstable properties of large-scale dynamical systems; (3) with the different final goals, the expectation for the dynamic learning is different.  We develop new deep learning-based methods for physical dynamics identification by incorporating physical priors as conservation law, group symmetry, and symbolic elements.   We also focus on new requirements when learning the dynamic given the goal is to control the system.