Several methods for the computation of reacting flows involve real-time integration of chemical kinetics. These methods include transported probability density function (PDF) methods, direct numerical simulation (DNS), conditional moment closure (CMC), unsteady flamelet, multiple mapping closure (MMC), thickened flame model, linear eddy model (LEM), partially stirred reactor (PaSR, as in OpenFOAM) and laminar flame computation. The need for direct integration of kinetics often renders large-scale simulations with these methods prohibitively expensive. This issue can be overcome by tabulation, which is the pre-computing and storage of computationally expensive operations. Tabulation raises further issues, however, such as the anticipation and coverage of the states to be encountered during simulation and the avoidance of excessive memory requirements.
The present seminar will describe a tabulation approach with artificial neural networks (ANNs), which are a class of machine learning models that can be employed for function approximation. After a review of the first-generation ANN tabulation methods, we will discuss recent developments that enable the generation of data-driven models able of accommodating a wide range of reacting flows, including laminar flamelets and premixed flames, turbulent non-premixed and premixed flames and flames with fuel blends and energy losses. The main elements of the methodology will be discussed, together with an explanation of how they lead to enhanced generalisation, accuracy and error control.