In this talk, I explained the usage of deep learning paradigm into inverse problems solving in high energy nuclear physics, focusing on studies about QCD matter in extreme conditions. To allow for efficient inverse problem solving, well-developed physical priors would be helpful in the solving procedure. Specifically, I introduced two examples with two different strategies involved: one is about QCD transition type identification in heavy ion collisions using supervised learning, where the physics prior is embedded in the training data generated by state-of-the-art model simulations; the other is about effective in-medium heavy quark potential reconstruction based upon lattice QCD data for Bottomonium mass and width, here the prior is manifested inside our devised approach to couple deep neural network represented potential with the Schr\"odinger equations.