Benchmarking machine learning models against full-scale acoustic absorption measurements in gas turbine liners

Abstract

Accurate prediction of acoustic energy absorption in gas turbine combustors is critical for enhancing performance, mitigating thermoacoustic instabilities, and reducing emissions. Traditional experimental and mathematical modeling approaches often struggle to capture the nonlinear and frequency-dependent behaviour of combustor liners, especially under varying operational conditions. This study presents a comprehensive evaluation of four machine learning (ML) models—Random Forest (RF), Extreme Gradient Boosting (XGB), Convolutional Neural Networks (CNN), and K-Nearest Neighbors (KNN)—against full-scale experimental measurements across six double-liner and three single-liner gas turbine configurations. A total of 615 operational records were used to train and validate model performance. In the double liner configuration, the RF and CNN models achieved the lowest average RMSE of 0.044, closely followed by XGB (0.045). However, XGB demonstrated the highest predictive accuracy (R² = 0.896), outperforming RF (0.891) and CNN (0.895), whose performance was partly constrained by limited dataset size. In the single liner configuration, XGB again delivered superior results (RMSE = 0.077, R² = 0.782) compared with RF (0.081, 0.580) and KNN (0.092, 0.470). The KNN model exhibited poor generalisation in the high-dimensional feature space, with R² dropping to 0.32 in certain tests. Overall, the machine learning framework reduced prediction error by approximately 31–56.2 % compared with conventional semi-empirical models (COS and COT), while requiring significantly lower computational effort to achieve comparable or superior accuracy. This is the first study to benchmark ML models against full-scale acoustic absorption data, introducing a validated data-driven framework with implications for combustor liner design, predictive maintenance, and real-time diagnostics. The integration of ML models into simulation workflows promises substantial reductions in experimental cost and improved design responsiveness. These findings offer a scalable, computationally efficient solution for aeroacoustic optimization—supporting the broader goal of cleaner, quieter, and more efficient gas turbine systems.