Coaxial-Injector Surrogate Modeling Based on Reynolds-Averaged Navier–Stokes Simulations Using Deep Learning

Abstract

Facing the need to increase the accuracy of rocket engine design tools, the present work introduces an innovative methodology for the design and optimization of rocket engine combustion chambers using numerical simulations and deep learning. An experimental test case of a single coaxial injector is taken as a reference point, and a design of experiments is generated by varying nine parameters (geometrical and operative conditions). Reynolds-averaged Navier–Stokes simulations are carried out to generate the data set. The data are used to train surrogate models of different fidelities, from low-dimensional outputs (zero-dimensional and one-dimensional) toward the two-dimensional temperature field. Attention is given on the selection of the proper machine learning technique. For low-dimensional outputs, results show that deep neural networks outperform other standard machine learning tools, namely, radial basis function and kriging. Regarding high-dimensional outputs, convolutional neural networks with gradient-based loss functions are found effective to capture the large and smooth temperature variations, as well as the thin and sharp temperature gradients at the flame front. Eventually, the models are used in the framework of an optimization problem. Results highlight the benefits of new design and optimization tools based on deep learning, capable of real-time predictions of complex flowfields.