Neural Network Flame Closure for a Turbulent Combustor with Unsteady Pressure

TL;DR

This paper develops neural network models to replace flamelet tables in large-eddy simulations of a liquid-propellant rocket engine, achieving high accuracy and successful integration into CFD simulations.

Contribution

The paper introduces neural network models that accurately replace flamelet tables for sub-grid modeling in rocket engine simulations, enabling fully NN-based CFD.

Findings

Neural networks estimated flame variables with less than 1% error.

NN-based CFD results closely match table-based CFD results.

Neural network models were successfully integrated into CFD simulations.

Abstract

In this paper, neural network (NN)-based models are generated to replace flamelet tables for sub-grid modeling in large-eddy simulations of a single-injector liquid-propellant rocket engine. In the most accurate case, separate NNs for each of the flame variables are designed and tested by comparing the NN output values with the corresponding values in the table. The gas constant, internal flame energy, and flame heat capacity ratio are estimated with 0.0506%, 0.0852%, and 0.0778% error, respectively. Flame temperature, thermal conductivity, and the coefficient of heat capacity ratio are estimated with 0.63%, 0.68%, and 0.86% error, respectively. The progress variable reaction rate is also estimated with 3.59% error. The errors are calculated based on mean square error over all points in the table. The developed NNs are successfully implemented within the CFD simulation, replacing the flamelet table entirely. The NN-based CFD is validated through comparison of its results with the table-based CFD.