Subgrid-scale Models with Interpretable Machine Learning in LES of Transcritical Reacting Flows

Many practical combustion systems operate under high pressures that surpass the thermodynamic critical limit of fuel-oxidizer mixtures. One challenge, which arises from the resulting real fluid behavior, includes the validity of existing subgrid-scale (SGS) models in large-eddy simulations of these systems. Data-driven methods can provide accurate closure in simulations of turbulent flames, but often lack interpretability, wherein they provide answers but no insight into their underlying rationale. The objective of this study is to investigate the accuracy of SGS models from conventional physics-driven approaches and an interpretable machine learning algorithm, i.e., the random forest, in a turbulent transcritical non-premixed flame. To this end, a priori analysis is performed on direct numerical simulation data of transcritical liquid oxygen/gaseous-methane (LOX/GCH4) inert and reacting flows. Results demonstrate that random forests can model SGS stresses as accurately as algebraic models, when trained on a sufficiently representative database. The random forest feature importance score is shown to provide insight that can be applied towards discovering SGS models through sparse regression.