Machine learning is a typical optimisation problem, and its learning process is an iterative optimisation process in the parameter space. It is a natural way of thinking to consider the iterative motion process of this algorithm as a kinetic process. The theory of dynamics has been developed over a long period of time and is very complete, with quantum dynamics, Newtonian dynamics, thermodynamics, electrodynamics and molecular dynamics, which theoretically describes the laws of motion by establishing a set of kinetic equations. The establishment of the dynamics theory of machine learning is expected to address the lack of theoretical models in this field, thus advancing the development of machine learning theory and applications. The theoretical modeling of optimization problems using a dynamics approach was pioneered by Metropolis in 1953, drawing from thermodynamics [1].