Neural networks, though called as black-box uniform approximator and difficult to interpret, have an unreasonable effectiveness in learning unknown mechanisms with bless of dimensionality, and have lots of applications. In this talk, I will introduce a recent state-of-the-art universal differential equation method that embeds neural networks into differential equations. Three applications will be shown. (1) Using deep learning techniques to estimate effective reproduction number and compared with EpiEstim and EpiNow2 method. (2) Discovering unknown human behavior change mechanisms in transmission dynamics. (3) Using deep learning techniques to solve optimal control problems by representing optimal control function as neural networks, and compared with traditional direct, indirect, and dynamic programming methods.