Deep learning for scientific computing has been an active research field recently, while reinforcement learning-based scientific computing is relatively less developed. This talk introduces two examples of applying reinforcement learning to solve forward and inverse problems with state-of-the-art results. First, reinforcement learning is introduced to solve a certain class of high-dimensional PDEs to high and even machine accuracy. The accuracy would not increase when the problem dimension grows. Second, reinforcement learning is introduced to optimize sensing strategies in inverse scattering problems. Optimized sensing leads to significantly better data for inverse estimation with limited imaging resources. Optimized sensing produces clear image reconstruction while existing methods fail to provide meaningful reconstruction.