Answers are grounded in this paper and cite their sources.
Thermoacoustic instability poses a major challenge to the reliable operation of combustion systems, yet many existing prediction approaches rely on system-specific training data and therefore require retraining or calibration when applied to new configurations. Here, we present two frameworks for instability prediction, both based on phase-space reconstruction of scalar pressure measurements, and evaluate them on two distinct experimental setups: an annular combustor fueled with a hydrogen–methane blend exhibiting isolated instability episodes, and a small-scale methane–air combustor exhibiting rapidly recurring instability events. A cluster-based anomaly detection framework serves as a data-driven baseline by computing a probabilistic atlas of the system attractor directly from experimental data. A convolutional neural network trained on synthetic data consisting of 766 labeled recurrence matrices generated from a reduced-order model is evaluated zero-shot on both configurations. On the annular combustor, the deep learning approach achieves a mean warning time of 11.1 s, compared to 7.8 s for the cluster-based baseline on the same holdout set. On the small-scale combustor, the deep learning approach yields a mean warning time of 1.692 s, compared to 0.119 s for the baseline. The results indicate that recurrence structure encodes precursor patterns that remain informative across synthetic and experimental combustor dynamics, supporting zero-shot instability prediction without combustor-specific retraining. Novelty and significance statement: This work addresses the challenge of generalization in thermoacoustic instability precursor prediction. A convolutional neural network trained exclusively on synthetic data from a reduced-order model is evaluated zero-shot on two experimentally distinct combustor systems, thereby avoiding the threshold-dependent calibration required by many recurrence-based quantification approaches (Nair et al., 2014; Nair and Sujith 2015). The approach also avoids the need for combustor-specific training data used by many existing deep learning methods (Zhou et al., 2021; Lyu et al. 2021; Cellier et al., 2021). A cluster-based anomaly detection approach based on the same phase-space reconstruction and calibrated using target combustor data is used as a baseline. Evaluation across two configurations differing in geometry, fuel composition, and instability behavior provides a stringent assessment demonstrating cross-configuration transferability.
No takes yet. Share an insight, caveat, or question.
Agarwal et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: