Artificial neural networks for combustion modelling

Abstract

Combustion plays an important role on the energy production network throughout the entire world, from internal combustion engines to gas turbines. Frequently, the chemical reactions that characterize the combustion phenomena occur under a turbulent flow field, in which the mixture among the reactants is enhanced, improving the efficiency of the thermal systems. The compre- hensive modelling of turbulent combustion systems incorporating large chemical mechanisms is a mandatory step towards the desing of highly efficient low polutant thermal engines. These simulations, however, are highly time demanding, mainly due to the direct numerical integration (DI) of the chemical species time evolution, which prevent their use in large scale industrial applications. In this current work, a methodology to reduce the CPU time consumption of turbulent combustion simulations via machine learning techniques is developed and applied in a variety of test cases comprising different fuels and combustion regimes. The direct numerical simulation (DNS) of high Reynolds number flow fields requires high spatial resolution due to the presence of very small time and lenght scale of turbulent structures. This issue is aggravated when the incorporation of finite rate chemistry effects is performed, since that at every point in time and space the chemical species dynamics have to be computed. In order to reduce the computation burden, a large eddy simulation (LES) framework, where a spatial filter is introduced, resolving the large scales of motion while modelling the small scales is employed. Since that such model presents the chemical source term in an unclosed form, an extension combining LES and joint-scalar probability-density-function (pdf) for the transport equation, accouting for the turblence-chemistry interaction, is adopted. The solution of such transported LES-pdf equation, however, is still unfeasbile due to the high dimensionality of the pdf. Our efforts, nonetheless, are mainly focused on resolving the first and second moments of the reactive scalars evolution, hence a solution scheme via Eulerian stochastic fields, solving the equivalent stochastic differential equation via the Monte Carlo estimator is performed. The LES-pdf model, although significantly less time demanding when compared to the DNS case, is still unaffordable for large scale applications. We then focus our efforts in the reduction of the chemical reaction integration computational burden via a tabulation through machine learning techniques. First, we employ a reduced mechanism generated via rate-constrained controlled equilibrium (RCCE), in order to reduce the chemical space dimensionality and allow a feasible tabulation workflow. Subsequently, a framework involving clustering and regression techniques via self-organizing-maps (SOM) / Multilayer Perceptron (MLP) is developed. This methodology is comprehensively built from the data sampling, SOM-MLP training, testing, validation and application to turbulent flame simulations. One of the main challenges is the data acquisition, which is performed here by the simulation of various abstract combustion models, such as flamelets and one-dimensional premixed flames, to span the chemical composition space accessed by a series of turbulent flames. Following, the SOM-MLP framework is trained via standard optmitzation tech- niques and validated against unseen data and with one-dimensional laminar flamelets and premixed flames. Finally, the application of the SOM-MLP synergy alongside the LES-pdf formulation is employed to turbulent non-premixed flames, namely Sydney L (CH4) and Sandia D/G (DME) and turbulent partially premixed flame namely Cambridge SWB6. Results of the developed LES-pdf-ANN methodology simulations demonstrate good agreement with LES-pdf-RCCE simulations at the same time that the overall approach shows good agreement with experimental data, while great savings in CPU time consumption are observed (approaximately three orders of magnitude O(3)). Finally, the disk space storage requirement is very low compared with other tabulation techniques, since the space is only required to store the training parameters. These results suggest that our methodology offers a feasible solution for the comprehensive simulation of turbulent combustion systems within a moderate amount of time and computational resources.