Hybrid Neural-Kernel Approach for Flamelet Modelling of Turbulent Combustion

Conventional tabulated chemistry approaches are hindered by high memory consumption, with resource demands scaling exponentially as the number of control variables increases. This restricts their practicality to high-dimensional flamelet tables. To address these limitations, we propose a hybrid surrogate model that combines neural networks with kernel-based methods to directly predict the progress variable source term and species mass fractions. By employing this hybrid approach we can replace the extensive multidimensional UFPV tables as well as the need for complex interpolation schemes.

The model is first validated through a priori zero-dimensional analysis, demonstrating its ability to closely replicate results from conventional tabulated chemistry. The model is subsequently incorporated into three-dimensional computational fluid dynamics (CFD) simulations of widely recognized reacting flow benchmarks. Its performance is evaluated using key combustion metrics, including ignition delay and lift-off length, along with qualitative assessments of species concentration and temperature fields. Results indicate that this hybrid machine learning approach accurately reproduces the interpolated solution while significantly lowering memory demands and maintaining high computational efficiency.