Stability Analysis of a Liquid Oxygen–Methane Rocket Combustor Using Hybrid CFD–Machine Learning Approach

Modern rocket combustors are highly susceptible to thermoacoustic instability (TAI) due to the positive feedback between pressure and heat release oscillations that occurs inside the acoustically confined combustion chamber. TAI in rocket combustors may result in structural damage and subsequent mission failure. To avoid such an outcome, careful mapping of the stable and unstable regimes of the rocket combustors is typically undertaken during the design and development stages. While experiments with industry-scale rocket combustors are prohibitively expensive, large eddy simulation (LES) can greatly aid in understanding the stability characteristics of the combustor during the early stages of design. However, LES of industry-scale rocket combustors is itself computationally very expensive. Therefore, we explore an alternative hybrid approach. LES simulations of an industry-scale LOX-methane rocket combustor are run for a small set of operating points, and the obtained data are used to map the stable and unstable regimes of the combustor using a dynamics-attuned data-driven tool, namely the temporal convolutional network (TCN). TCN sensitivity to the approach of TAI is then bench-marked against the 0–1 test and translational error analysis, two well-known dynamics-based tools that have previously been successfully applied in thermoacoustic instability characterization. Among the methods explored, the TCN tool is seen to be most suitable for stability analysis of the rocket combustor, at least for capturing the dynamical transition observed in the present combustor configuration. As the TCN tool is computationally cheap and inherently parallelizable, a hybrid LES-TCN approach can significantly reduce the computational requirement for high-fidelity simulation-based stability analysis of rocket combustors. Moreover, the results indicate that the TCN tool can also be used for early prediction of impending TAI on a real-time basis. However, the suitability of TCN for real-time early detection remains to be validated experimentally.