Numerical sensitivity analysis of HTPB counterflow combustion using neural networks

Solid fuel combustion requires pyrolysis gases to burn near its surface to provide enough heat feedback to decompose the solid and continue to provide the volatile gases required to sustain combustion. This coupled process defines the difficulty in sustaining solid fuel combustion in a variety of propulsion environments and necessitates a fundamental understanding of the physical processes in order to drive system design. This study explores the combustion of hydroxyl-terminated polybutadiene (HTPB) in a counterflow diffusion flame burner with 50% and 100% oxygen content and compares the regression rate and flame standoff to experimental data. A sensitivity analysis is pursued to identify the model parameters that need improvement and to help guide the next campaign of experiments. Neural networks are developed in a compact way as a means of providing quantitative results on the sensitivity of input parameters. Then a fully connected, deeper neural network is trained on the input parameters – oxidizer mole fraction, solid fuel heat of formation, pre-exponential factor of pyrolysis Arrhenius rate, molecular weight of pyrolysis species, oxidizer mass flux, separation distance, and the oxidizer temperature, – and shown to predict output variables – regression rate and flame standoff – within 90% and 95% accuracy respectively. This network is then used to create millions of data points with an overlapping parameter space for further statistical analysis and improvement of model parameters. In all, the data analysis presented using a neural network approach will help drive the design of experiments and is shown to increase the accuracy of the model in comparison to experimental measurements.