How well can a multi-task learning model extract collective coordinates for nuclear fission?

Describing nuclear fission with a microscopic theory remains one of the most important and challenging problems in nuclear theory. A conventional approach involves a use of empirical parameters, such as quadrupole moments, for collective coordinates. However, dynamics of a nucleus does not necessarily select these parameters, and thus they do not guarantee an accurate description of fission. Current challenges persist, particularly in accurately reproducing observables such as half-lives with low-order multiple-pole moments. In this presentation, we will propose a novel method for deriving collective coordinates with multi-task learning, that is a kind of deep learning. This method involves extracting collective information from randomly generated nuclear density and energy data using Density Functional Theory (DFT), and it is independent of empirical coordinates. By implementing this method, we successfully extracted a minimal number of degrees of freedom that aptly describe the deformation dynamics with imposed axial symmetry of states near the ground state of 236U. We will show that Q20 and Q30, which have been often employed for fission study, do not contain as much information as one would have thought. This presentation will delve into this innovative approach and the nature of these new coordinates.