<p>In this study, the capability of deep neural network is verified in the aerodynamic design of supersonic through-flow fan blade profile. A highly flexible parameterization method was first proposed, followed by generating a flow field data set using Latin Hypercube Sampling and batch processing. Through flexible deep neural network architectures, positive designs of prediction on performance, static pressure distribution and flow field from design variables, and the inverse design from static pressure distribution to blade profile, were well achieved. The effects of different deep neural network architectures on prediction performance are systematically compared, and the inverse design accuracy is quantitatively evaluated using representative error metrics. Moreover, this study provides comparisons of prediction accuracy among different architectures, sample sizes, physical complexities, and treatment approaches, with detailed assessments for each neural network. It was found that the deep neural network is powerful for establishing the physical relationship between input and output parameters in aerodynamic design, supporting flexible input-output combinations while ensuring a certain level of prediction accuracy. However, caution is advised during the later stage of design, as the prediction error is hard to be totally eliminated. Additionally, regions with drastic changes and boundaries of the design space tend to have larger prediction error.</p>