Early detection of thermoacoustic instability via deep learning of recurrence plots

We present a deep learning framework for the early detection of thermoacoustic instabilities in laminar and turbulent flow systems. These instabilities, typically arising from Hopf bifurcations or via intermittency, manifest as large-amplitude pressure oscillations at a limit cycle. Predicting their onset is challenging owing to the nonlinearity and sensitivity of the thermoacoustic feedback loop, which involves complex multiscale interactions among combustion, hydrodynamics, and acoustics. Our approach leverages the pattern recognition capabilities of convolutional neural networks to forecast the onset of such instabilities using recurrence plots as input. These plots are two-dimensional topological representations of high-dimensional phase space trajectories. We train a ResNet-18 deep learning model on unbinarized recurrence plots generated from experimentally measured pressure data, producing a scalar output indicative of the proximity to the instability boundaries. We find that this hybrid framework is sensitive enough to extract evolving topological features from recurrence plots, generating reliable early warning indicators of impending thermoacoustic instabilities in various flow systems.