Fast Prediction of Two-Dimensional Flowfields with Fuel Injection into Supersonic Crossflow via Deep Learning

Abstract

Fuel injection is one of the most crucial components for scramjet engines, a promising hypersonic airbreathing technology for economical and flexible space transportation systems. While surrogate modeling based on machine learning has been employed to replace computational simulations for performance evaluation in design optimization of such components, it can inherently predict performance parameters only as scalar quantities. This study investigates the capability of deep learning to predict the fuel injection flowfields, aiming to assist with data-driven approaches for data mining and optimization. Two-dimensional flowfields with sonic fuel injection into a Mach 3.8 crossflow have been trained using the multilayer perceptron. The resultant model has been found to be able to predict the flowfields instantaneously with reasonable accuracy. Local sensitivity analysis has been performed to examine the influence of the design variables on flow properties to gain insights into the effects of their variations on local flow phenomena.