Establishment and validation of a relationship model between nozzle experiments and CFD results based on convolutional neural network

Abstract

The acquisition of experimental data in a supersonic wind tunnel often faces challenges of complexity and high costs. Furthermore, there are limitations in the control of experimental conditions and measurement techniques. In this study, experiments were conducted on the start-up and shutdown processes of a single expansion ramp nozzle (SERN), and a dataset was established in conjunction with computational fluid dynamics (CFD). By utilizing experimental data and CFD simulation results, we employed a deep learning-based approach, utilizing one-dimensional convolutional neural networks (CNN), to establish a relationship model between the experimental and CFD results, with the aim of using CFD results to predict experimental outcomes. We provide detailed descriptions of the experimental methods, numerical simulation methods, and CNN model, and conduct training and testing of the model. The results demonstrate that the CNN model is capable of accurately predicting the wall pressure distribution of the SERN, with low prediction errors. In interpolation and extrapolation cases, the model exhibits good accuracy and generalization ability, with coefficients of determination above 96%. Compared to CFD calculations, the CNN model proves to be more reliable and accurate in predicting SERN performance. Additionally, we establish a comparative model that does not utilize CFD data to verify the superiority of the proposed CNN model. The results show that the CNN model with CFD data improves performance on average by 14%. This approach reduces reliance on experiments to a certain extent, lowering experimental costs and complexity. It not only saves the resources and time required for experiments but also compensates for the limitations in controlling experimental conditions and measurement techniques. Therefore, this research provides a viable solution to overcome the challenges and limitations of acquiring experimental data.