With the rapid development of big data and the Internet of Things, data-driven technology, especially deep learning (DL), is becoming increasing important in intelligent maintenance. However, the “black box” nature of DL-based intelligent maintenance still seriously hinders wide applications in industry, especially safety-critical applications. In fact, before the rise of DL, the physics-driven approach, as a white box model that relies on the causality to establish physics law from first principles, is also a popular way, but it is not accurate enough. As two ways of observing the laws of the physical world, data-driven and physics-driven models are not opposite, but two sides of one coin, and they have consistent insight. Therefore, integrating physics model into DL, namely physics-informed deep learning (PIDL), is a nature and promising pathway towards scientific intelligent maintenance. This talk mainly aims to emphasize the importance of PIDL in scientific intelligent maintenance, where users can understand the operation mechanism inside the model and realize human-in-the-loop. At last, some applications of PIDL are discussed to illustrate the merit of scientific intelligent maintenance.