RECONSTRUCTION OF TWO-DIMENSIONAL TO THREE-DIMENSIONAL FLOW TRANSITION FIELDS USING NEURAL NETWORK-BASED GENERATIVE ADVERSARIAL NETWORKS

Flow transition is a common phenomenon in fluid dynamics, widely observed, particularly in scientific and technological fields such as turbomachine. However, accurately measuring and predicting flow transition poses a significant challenge due to its inherent complexity, and acquiring comprehensive data for the entire fluid domain is often a daunting task. Artificial intelligence (AI) emerges as a promising avenue for conveniently and efficiently obtaining typical information about flow transition, thereby assisting conventional experimental and computational methods. In this study, the Generative Adversarial Network (GAN) was extensively used in the fundamental neural network training model for reconstructing the three-dimensional flow fields in a compressor cascade. Through the joint training of the generator (G) and discriminator (D) based on GAN, the velocity distribution was constructed. The training dataset comprised 200 consecutive snapshots, with an additional 200 consecutive snapshots reserved for testing. To ensure a high degree of accuracy, the data for the compressor cascade flow was obtained through Direct Numerical Simulation (DNS). The trained generator model in this study successfully reconstructed three-dimensional flow field data, capturing the flow transition in smooth blade passages as well as the bypass transition induced by roughness elements. In-depth analysis of the data resulted in accurate predictions and reconstructions of the transition location, range, and structure. This study has achieved the reconstruction of information from two-dimensional to three-dimensional flow fields for time instances not included in the training data using the 2D3DGAN model. The implementation of this aspect implies that the 2D3DGAN model can be utilized to infer the flow field status at unknown time instances by leveraging partial information. This capability opens up greater potential applications in the engineering domain, thereby reducing experimental costs and enhancing engineering efficiency.