Physically Interpretable AI-Driven Prediction of Instability Regimes in Hydrogen-Enriched Premixed Combustion

Hydrogen-enriched combustion is central to decarbonizing high-efficiency energy systems, yet its practical adoption is limited by the onset of thermoacoustic and hydrodynamic instabilities in premixed flames. In this study; a novel, integrated framework that combines high-fidelity computational fluid dynamics (CFD) simulations with interpretable deep learning for the prediction and physical diagnosis of combustion instability was proposed. A parametric suite of 1,500 axisymmetric CFD simulations was carried out, systematically varying hydrogen blending ratios (0–100% by volume), equivalence ratios (ϕ = 0.6–1.4), and turbulence intensities (5–25%). Key instability markers including root-mean-square (RMS) pressure, flame front wrinkling, and radical pool dynamics were extracted from both stable and unstable flame regimes. The data collected was used to train a hybrid convolutional neural network–long short-term memory (CNN–LSTM) model, which achieved a test accuracy of 94.3%, F1-score of 94.4%, and area under the receiver operating characteristic curve (AUC-ROC) of 0.978 in binary regime classification. SHAP-based interpretability analysis demonstrated that the model’s predictions were grounded in physically relevant features, with RMS pressure, OH fluctuations, and dominant acoustic frequencies serving as the principal contributors. AI-predicted instability regime maps showed an 88.6% overlap with CFD-derived instability thresholds, highlighting the physical consistency of the approach. Distinct field visualizations showed that unstable regimes (ϕ = 1.1, H₂ = 80%) exhibit pronounced front wrinkling, broader high-temperature zones, and spatially distributed radical production compared to stable flames. This approach opens a promising path for data-driven, physically interpretable instability diagnostics, which could directly impact for burner design, operational safety, and real-time combustion monitoring in hydrogen-based systems. In future work, it is aimed to extend this approach to multi-fuel configurations and experimental integration for real-world deployment.