Adaptive control of thermoacoustic instabilities using a reinforcement learning approach: an experimental demonstration on a methane/hydrogen turbulent flame

Combustion instability poses a significant challenge in propulsion systems, leading to undesirable pressure oscillations that can compromise the system performance and safety. In this work, we implement an adaptive control policy based on reinforcement learning (RL) to suppress combustion instability in a lab-scale bluff body burner powered with a mixture of hydrogen and methane. We first characterize the combustion instability in the burner by examining its response to changes in equivalence ratio, hydrogen enrichment, and combustor length. In addition, we study the flame shape evolution under self-excited instability and stable conditions. Second, we implement a control system using a microphone, a loudspeaker, and a FPGA chip. We train a RL model to perform active control over a wide range of operating condition and benchmark its performance against a conventional phase-shift controller. Lastly, we discuss RL as an adaptive control scheme to mitigate combustion oscillation across diverse scenarios encountered in real-world applications.