A top level parallelization and data fusion approach for identification of flame transfer functions with increased reliability, accuracy and efficiency

The present study investigates a novel design of numerical experiment and data analysis procedure for Flame Transfer Function (FTF) identification to reduce return time and computational cost in thermoacoustic stability analysis. A machine learning approach called Multi-Fidelity Gaussian Process (MFGP) fuses mono-frequent and broad-band excitation data in order to retain the strengths of each method and avoid their weaknesses. The FTF is determined as the most likely function considering all data sets through a maximum likelihood approximator. Therein the global trend of the function is provided by the broad-band part and its accuracy and certainty is greatly increased near the frequencies of harmonic excitation. Only a few mono-frequent training samples are required for the resulting function to satisfy the accuracy and uncertainty bounds provided by harmonic test samples over the whole frequency range. This greatly reduces the effort compared to pure mono-frequent excitation and increases the accuracy compared to broad-band excitation. This is the first time this method is applied to data originating from an LES of an applied burner, which allows the use as a predictive tool in burner design. Experimental FTF measurements serve for validation purpose. The uncertainties of the mono-frequent excitation are calculated from the LES data and not assumed. As the single simulation already runs in parallel scalability limit, a top-level parallelization approach is introduced to decrease the return time significantly. Overall the MFGP greatly increases the accuracy of the FTF identification and reduces return time and computational effort by a wide margin.