Flamelet modeling of spray flames with mixture of experts based learning of combustion manifolds

Opeoluwa Owoyele (1), Austin Cody Nunno (1), Pinaki Pal (1), Prithwish Kundu (1)
(1) Argonne National Laboratory, IL, United States

This work presents an a posteriori assessment of a novel Mixture of Experts (MoE) approach for learning tabulated combustion manifolds. The goal is motivated by the poor scaling of flamelet tables with increasing dimensionality, wherein the size of the table increases exponentially as more independent variables are included in the table. To resolve this issue, we present MoE, a divide-and-conquer machine learning (ML) approach to learn flamelet tables. In this approach, a system of neural networks, consisting of a gating network classifier and multiple regression expert models, are trained simultaneously to learn the flamelet table outputs as functions of the control variables. As a result, the flamelet table is partitioned by the gate, with the regression models being “experts” at making predictions within different portions of the manifold. The proposed approach is demonstrated and validated in the context of the unsteady flamelet/progress variable (UFPV) model applied to Reynolds-averaged Navier-Stokes (RANS) simulation of Engine Combustion Network (ECN) Spray A.