Deep-learning prediction and uncertainty quantification for scramjet intake flowfields

Abstract

Scramjet is a promising propulsion technology that provides efficient and flexible access-to-space and high-speed point-to-point transportation. Since the design process of scramjet (supersonic combustion ramjet) engines requires numerous flowfield evaluations, fast and accurate flow predictions play a key role in promoting the development of knowledge and technologies. Deep learning techniques are increasingly used for flow prediction, and the present study applies them to viscous supersonic flowfields inside scramjet intakes. The capability and limitations of deep learning prediction for such flowfields have been investigated from the viewpoints of both physics and machine learning by means of uncertainty quantification and principal component analysis. The results indicate that the flowfields with complex aerodynamic phenomena such as boundary layer separation and Mach disks are difficult to predict. It has been attributed to lack of similar flowfields as well as the sensitivity of the fluid phenomena to the geometries. Uncertainty quantification effectively allows for the detection of such difficult cases without model verification prior to utilization. It has been further employed to address the issues in the prediction of flowfields with boundary-layer separations by increasing the number of training data effectively.