Abstract: In today’s 5G or future wireless communications, smart new materials and deep learning algorithms are widely used, and they have begun to show an prosperous development trend in improving spectrum utilization and reducing energy consumption. In this paper, the millimeter wave and massive multiple-input multiple-output (MIMO) wireless network datasets are used as mobile communication scenarios to study channel model estimation. Furthermore, a large-scale smart surface with a small number of reflective units that can be flexibly configured is introduced, and a multi-layer perceptron network is designed with the channel characteristic parameters as the input, and the achievable rate as the output label. The smart metasurface composition such as reflective units, the number of active units, the transmit power, and the size of the training data set on the achievable rate of mobile users and their impacts are evaluated by using simulation benchmarks, smart metasurfaces, deep learning. The computational results show that through the proper design of these options, the effectiveness of deep learning in the perception of wireless network environmental characteristics can be increased.