Neural network acceleration of numerical simulation of methane combustion in a gas turbine engine

Abstract

Gas turbines are essential for high-power energy generation, but growing demands to reduce NOₓ and CO₂ emissions make traditional combustion chamber design increasingly complex and costly. This work proposes a new modeling paradigm that combines high-fidelity Computational Fluid Dynamics using neural network learning to accelerate emission prediction. A Computational Fluid Dynamics model was developed using the Reynolds-averaged Navier-Stokes equations with the k–ε turbulence model and a non-premixed Probability Density Function approach to simulate turbulent methane combustion. NOₓ emissions were calculated postsimulation using the Zeldovich mechanism. Model validation included varying fuel flow, excess air ratio, and wall heat loss. To speed up evaluations, a multilayer perceptron neural network was trained on Computational Fluid Dynamics results to predict NOₓ and CO₂ emissions based on key inputs (fuel rate, air excess, temperature, pressure, cooling). The model achieved high accuracy with a coefficient of determination (R^2) of 0.998 for NOₓ and 0.956 for CO₂ on an independent test set. Results showed good agreement with both experimental data and a Network of ideal reactors model using detailed kinetic scheme of methane combustion - Mech 3.0. This neural network serves as a fast surrogate model for emissions assessment, enabling rapid optimization of low-emission combustor designs. The approach is suitable for digital twins and combustion control systems and is adaptable to alternative fuels like hydrogen and ammonia.