Exploring the physical mechanisms of complex systems and making effective use of them are the keys to dealing with the complexity of the world. The emergence of big data and the enhancement of computing power, in conjunction with the improvement of optimization algorithms, are leading to the development of artificial intelligence (AI) driven by deep learning. However, deep learning fails to reveal the underlying logic and physical connotations of the problems being solved. Mesoscience provides a concept to understand the mechanism of the spatiotemporal multiscale structure of complex systems, and its capability for analyzing complex problems has been validated in different fields. This paper proposes a research paradigm for AI, which introduces the analytical principles of mesoscience into the design of deep learning models. This is done to address the fundamental problem of deep learning models detaching the physical prototype from the problem being solved; the purpose is to promote the sustainable development of AI.