Evaluating generalizability of a deep learning model for the instantaneous conditional dissipation rate profiles in multi-modal turbulent combustion

More accurate and computationally efficient models applicable to multi-modal turbulent combustion are critical to expedite the development of improved energy conversion devices. Manifold-based modeling approaches offer computational efficiency by projecting the high-dimensional thermochemical state onto a low-dimensional manifold. Traditionally, a single manifold coordinate is chosen, limiting model applicability to a single ‘mode’ of combustion. This limitation may be overcome by projecting onto both a mixture fraction and a generalized progress variable, resulting in a multi-modal manifold model with additional unclosed dissipation rates that appear as parameters in the manifold equations. Previous work sought to model the instantaneous spatiotemporal variation of these unclosed dissipation rate profiles with a deep neural network (DNN), leveraging Direct Numerical Simulation (DNS) data of an auto-igniting n-dodecane jet flame for training. This work builds upon previous efforts, extending the training data to include multiple time snapshots that exhibit early-stage ignition as well as premixed and nonpremixed behavior. The generalizability of the trained DNN model is evaluated a priori through testing with a separate DNS database of a gas turbine combustor simulation exhibiting localized regions of extinction/re-ignition and flame stratification due to back-support provided from recirculation of hot product gases.