Periodic Reporting for period 2 - MAGISTER (Machine learning for Advanced Gas turbine Injection SysTems to Enhance combustoR performance.)
Periodo di rendicontazione: 2019-09-01 al 2021-11-30
Clean combustion technology for aircraft engines can reduce the impact of air transportation on ecosystems and humans’ health. The vision for European aviation defines stringent regulations on pollutant emissions. To meet these goals, the major engine manufacturers develope lean premixed combustors. This development introduces a large risk for reduced reliability of engines due to pressure oscillations in the combustor (ThermoAcoustics). Industrial experience shows that pressure oscillations often only surface when the full engine has been built. Traditional engineering methods fail for the design of the engines due to a high sensitivity of the oscillations with respect to input parameters. Aviation industry encounters currently the 4th industrial revolution, cyber-physical systems analyse and monitor technical systems and take automated decisions with Machine Learning. The ITN [MAGISTER] has utilized ML to predict ThermoAcoustics in aircraft engine combustors. The participation of the aircraft engine OEMs GE, Rolls Royce, Safran ensured industrial relevance and outreach of the results. The project has shaped 15 Early Staged Researchers in a network of scientists and industry to work on these design issues in aviation technology using ML.
Objectives
Develop methods that can predict and control thermoacoustics from TRL 2 to 9.
Apply machine learning algorithms to improve models to predict thermoacoustics in aircraft engines and derive combustor hardware design implications from the predictions.
Devise and adapt machine learning algorithms to thermoacoustic experiments at the laboratory scale and to industrial scale for aircraft engines.
Advance acoustic and combustion models to capture the interaction of acoustics with liquid fuel sprays with high accuracy.
Generate a sophisticated experimental data base for thermoacoustics of liquid fuel combustion for validation.
Objectives
Develop methods that can predict and control thermoacoustics from TRL 2 to 9.
Apply machine learning algorithms to improve models to predict thermoacoustics in aircraft engines and derive combustor hardware design implications from the predictions.
Devise and adapt machine learning algorithms to thermoacoustic experiments at the laboratory scale and to industrial scale for aircraft engines.
Advance acoustic and combustion models to capture the interaction of acoustics with liquid fuel sprays with high accuracy.
Generate a sophisticated experimental data base for thermoacoustics of liquid fuel combustion for validation.
6 work packages: Machine Learning, Combustion models, Acoustic models, Validation data, Aircraft engines and Training.
ESR 1 Cambridge trained Bayesian ensembles of Neural Networks on a laboratory rig and a set of full-scale engines from an industrial partner. His neural network model was able to input the noise from the system and output a.o. the distance from instability.
ESR 2 Armines has designed estimation algorithms for systems with thermoacoustic instabilities, based on physical models of this phenomenon. The algorithms successfully estimated a.o. acoustic impedances of an electrically heated Rijke tube.
ESR 3 GE: The dynamic response to inlet velocity fluctuations of flames can be modeled through neural networks which have been trained on selected CFD runs. The limit cycle amplitudes of a combustion system were predicted by coupling the trained neural network model with an acoustic solver. The trained neural network model can predict the burner system time evolution.
ESR 4 ANSYS: Pure control theory and machine learning techniques were combined. An observer algorithm was designed using artificial neural network methods for low error estimation nonlinear systems states.
ESR 5 TUM: CFD simulations were combined with System Identification (SI) to estimate flame dynamics. The uncertainties in flame models obtained were quantified through a Gaussian Process ML algorithm.
ESR 6 CERFACS: Implemented in AVBP a multicomponent evaporating spray model with detailed chemistry. Simulated were spray flames of two multicomponent liquid fuels in the large scale LOTAR configuration at ONERA to determine the transfer function.
ESR 7 UT: Unsupervised machine learning algorithms of Proper Orthogonal Decomposition are used to deconstruct the complex flow fields into their basic modes to identify acoustic characteristics.
ESR 8 Cambridge: obtained experimental data from a thermoacoustic rig and assimilated this into a ThermoAcoustic model. The physics-based model was tuned to match experimental data to be accurate with known error bounds.
ESR 9 TUM: improved models for acoustically absorbing combustor liners consisting of a perforated wall backed by a cavity. Impedance measurements of perforation patterns were performed, and an analytical method was developed with ML.
ESR 10 UT: In high-order unsteady CFD solvers with LES, non-reflective boundary conditions (NRBCs) are crucial for preserving the numerical accuracy of the solution.NRBCs were investigated for a high-order discontinuous Galerkin solver.
ESR 11 KIT: At KIT a facility was built to generate an oscillating airflow for the MAGISTER designed burner. The spray response was found to fluctuate with the same frequency as the air velocity at constant liquid flow rate.
ESR 12 UT: Self- excited high amplitude oscillations were studied, and post processed by chaotic analysis to obtain the geometric order of the generated phase portraits . A test rig wwas built ith the MAGISTER burner for liquid fuel combustion at high pressure.
ESR 13 GEDE: Machine learning algorithms were used for the interpolation and extrapolation of Flame Describing Functions aiming to allow a minimal amount of data points to be collected while still providing enough information to model the FDF accurately. A method of optimally selecting data points at test time using Gaussian Processes was demonstrated. Next it was demonstrated that the prediction of combustion instabilities is improved by supervised machine learning without any need for manual calibration and the quality of the warning is improved.
ESR 14 Safran tech: Investigated were Lattice Boltzmann methods to the problem of primary atomization using the colour gradient method to model immiscible two-phase flows. This method has been generalized to an arbitrary Equation of State, relieving the limitation between density and sound speed ratios.
ESR 15 SHE: performed CFD simulation of hydrogen combustion at high pressure. CFD simulations were done of the Safran spinning combustion concept.
ESR 1 Cambridge trained Bayesian ensembles of Neural Networks on a laboratory rig and a set of full-scale engines from an industrial partner. His neural network model was able to input the noise from the system and output a.o. the distance from instability.
ESR 2 Armines has designed estimation algorithms for systems with thermoacoustic instabilities, based on physical models of this phenomenon. The algorithms successfully estimated a.o. acoustic impedances of an electrically heated Rijke tube.
ESR 3 GE: The dynamic response to inlet velocity fluctuations of flames can be modeled through neural networks which have been trained on selected CFD runs. The limit cycle amplitudes of a combustion system were predicted by coupling the trained neural network model with an acoustic solver. The trained neural network model can predict the burner system time evolution.
ESR 4 ANSYS: Pure control theory and machine learning techniques were combined. An observer algorithm was designed using artificial neural network methods for low error estimation nonlinear systems states.
ESR 5 TUM: CFD simulations were combined with System Identification (SI) to estimate flame dynamics. The uncertainties in flame models obtained were quantified through a Gaussian Process ML algorithm.
ESR 6 CERFACS: Implemented in AVBP a multicomponent evaporating spray model with detailed chemistry. Simulated were spray flames of two multicomponent liquid fuels in the large scale LOTAR configuration at ONERA to determine the transfer function.
ESR 7 UT: Unsupervised machine learning algorithms of Proper Orthogonal Decomposition are used to deconstruct the complex flow fields into their basic modes to identify acoustic characteristics.
ESR 8 Cambridge: obtained experimental data from a thermoacoustic rig and assimilated this into a ThermoAcoustic model. The physics-based model was tuned to match experimental data to be accurate with known error bounds.
ESR 9 TUM: improved models for acoustically absorbing combustor liners consisting of a perforated wall backed by a cavity. Impedance measurements of perforation patterns were performed, and an analytical method was developed with ML.
ESR 10 UT: In high-order unsteady CFD solvers with LES, non-reflective boundary conditions (NRBCs) are crucial for preserving the numerical accuracy of the solution.NRBCs were investigated for a high-order discontinuous Galerkin solver.
ESR 11 KIT: At KIT a facility was built to generate an oscillating airflow for the MAGISTER designed burner. The spray response was found to fluctuate with the same frequency as the air velocity at constant liquid flow rate.
ESR 12 UT: Self- excited high amplitude oscillations were studied, and post processed by chaotic analysis to obtain the geometric order of the generated phase portraits . A test rig wwas built ith the MAGISTER burner for liquid fuel combustion at high pressure.
ESR 13 GEDE: Machine learning algorithms were used for the interpolation and extrapolation of Flame Describing Functions aiming to allow a minimal amount of data points to be collected while still providing enough information to model the FDF accurately. A method of optimally selecting data points at test time using Gaussian Processes was demonstrated. Next it was demonstrated that the prediction of combustion instabilities is improved by supervised machine learning without any need for manual calibration and the quality of the warning is improved.
ESR 14 Safran tech: Investigated were Lattice Boltzmann methods to the problem of primary atomization using the colour gradient method to model immiscible two-phase flows. This method has been generalized to an arbitrary Equation of State, relieving the limitation between density and sound speed ratios.
ESR 15 SHE: performed CFD simulation of hydrogen combustion at high pressure. CFD simulations were done of the Safran spinning combustion concept.
The tools developed guide design changes that will eliminate pressure oscillations over the desired operational window.
MAGISTER contributed significantly to the improvement of GE Aviation’s thermoacoustic tool processes for combustion dynamics across the operating range of aero engines. MAGISTER increased the ability of GE Aviation to develop advanced, low emission combustors of its aero engines. For the exploitation of the new tools, the ESR's were supervised by GE Aviation engineers involved in thermoacoustic tools. MAGISTER helped SAFRAN Aviation in the application of LBM to multiphase flows characterized by a high density ratio. A new colour gradient algorithm was proposed that can be employed with an arbitrary equation of state and leads to an improved numerical stability. MAGISTER improved the investigation of atomization phenomena, especially in presence of acoustic oscillations, which may represent a real technological problem in the design of less-pollutant aeronautical chambers. The innovation capacity of SAFRAN HE was improved . The novel methods and test beds developed within the project have already allowed a greater understanding of hydrogen and spinning flame lean combustion devices to design innovative ultra low-NOx combustors and hydrogen technologies . MAGISTER provided new light to ANSYS on methods that is driving the future of their software by bringing it closer to measurements. A hybrid reduced order model is created which combines simulation software and experimental data.
MAGISTER contributed significantly to the improvement of GE Aviation’s thermoacoustic tool processes for combustion dynamics across the operating range of aero engines. MAGISTER increased the ability of GE Aviation to develop advanced, low emission combustors of its aero engines. For the exploitation of the new tools, the ESR's were supervised by GE Aviation engineers involved in thermoacoustic tools. MAGISTER helped SAFRAN Aviation in the application of LBM to multiphase flows characterized by a high density ratio. A new colour gradient algorithm was proposed that can be employed with an arbitrary equation of state and leads to an improved numerical stability. MAGISTER improved the investigation of atomization phenomena, especially in presence of acoustic oscillations, which may represent a real technological problem in the design of less-pollutant aeronautical chambers. The innovation capacity of SAFRAN HE was improved . The novel methods and test beds developed within the project have already allowed a greater understanding of hydrogen and spinning flame lean combustion devices to design innovative ultra low-NOx combustors and hydrogen technologies . MAGISTER provided new light to ANSYS on methods that is driving the future of their software by bringing it closer to measurements. A hybrid reduced order model is created which combines simulation software and experimental data.