<p>Flow field computation plays a critical role in both scientific research and engineering applications. For decades, computational fluid dynamics (CFD) has served as the cornerstone of flow field analysis; however, high-resolution simulations are often hindered by considerable computational costs and lengthy processing times. In recent years, deep learning (DL), with its exceptional capability to handle high-dimensional nonlinear problems, has achieved remarkable progress in the field of fluid mechanics. This paper provides a comprehensive review of recent advances in applying DL methods to accelerate flow field computation, with particular emphasis on complex indoor environments characterized by multi-physics coupling. We begin by outlining the fundamental frameworks of deep learning and, on this basis, summarize four representative neural network architectures for flow prediction: end-to-end mapping networks, reduced-order mapping networks, physics-informed neural networks (PINNs), and operator neural networks (ONNs). We then systematically review the specific applications of these DL algorithms in indoor flow field prediction. In addition, we discuss key challenges faced by current research, including the lack of large, high-quality databases, limited interpretability and generalization capability of existing models, and the difficulty of accurately representing real indoor environments. Finally, we propose several promising research directions, such as exploring advanced algorithms, enhancing self-supervised learning techniques, and developing geometry-aware network models and multi-task hybrid frameworks. Advancing these frontiers is expected not only to significantly improve the efficiency and accuracy of flow field computations but also to provide a solid theoretical foundation and technical support for the optimization of intelligent indoor environments.</p>