Adaptive chemistry via pre-partitioning of composition space and mechanism reduction

Numerical simulations of multi-dimensional laminar flames with complex kinetic mechanisms are computationally very demanding, because of the large number of species and the strong non-linearity and stiffness of governing equations. In this work, we present and apply a novel adaptive chemistry methodology for mitigating the computational cost of such simulations, based on machine-learning algorithms which automatically classify the composition space via a priori defined classifiers. The methodology, called SPARC (Sample-Partitioning Adaptive Reduced Chemistry), is based on four steps: generation of data sets covering the temperature and composition space which is expected to be visited by the multi-dimensional flame; partitioning of the composition space in a prescribed number of clusters with similar composition via Local Principal Component Analysis (LPCA); generation of reduced kinetic mechanisms for each cluster via Directed Relation Graph with Error Propagation (DRGEP); CFD simulation of a multi-dimensional flame based on locally reduced mechanisms. The approach has been firstly demonstrated for the CFD simulations of steady and transient laminar coflow diffusion flames fed with a mixture of CH4 and N2 burning in air. The transient behaviour was artificially induced by different sinusoidal perturbations in the velocity profiles. Several numerical tests were carried out to explore the impact of partitioning parameters and the degree of mechanism reduction on the flame simulation, and very satisfactory results were observed, both in terms of accuracy and computational efficiency. In particular, even if a relatively small mechanism (84 species) was adopted, speed-up factors of ∼ 4 were observed. Larger speed-up factors can be achieved if more complex fuels are considered, since also larger and stiffer kinetic mechanisms are required and a higher level of reduction via DRGEP can be reached.