Electronic edition ISSN 1574-0579

Plasma-assisted combustion in a gas turbine engine: CFD and neural-network surrogate modeling

A. Chepurnyi - A. Jakovics

Institute of Numerical Modelling, University of Latvia, 3 Jelgavas str., Riga, Latvia

Abstract
This study combines two complementary approaches to improving gas-turbine combustors: plasma-chemical enhancement of methane combustion using a plasma-chemical element (PCE) and neural-network surrogate modeling to accelerate multi-case simulations. The PCE is driven by an arc discharge - an inherently magnetohydrodynamic (MHD) process in which Lorentz forces govern the plasma-jet acceleration. A CFD model of the combustor was developed using the Reynolds-averaged Navier-Stokes equations, the RNG k − ε turbulence model, and an equilibrium non-premixed combustion model with a presumed probability density function (PDF) for the mixture fraction. The PCE effect is represented through two physically distinct channels: a high-enthalpy plasma-jet boundary condition and a reduction in the effective activation energy according to the empirical correlation ∆E(β) (R² = 0.996) in the finite-rate CO-oxidation submodel; NOx is calculated using the extended Zeldovich mechanism. The model was verified and compared with published experimental data, chemical-reactor-network calculations using the GRI-Mech 3.0 mechanism, and the original combustor model with a PCE. An ensemble of 10 multilayer perceptrons was trained on 1000 CFD solutions using an extended input set consisting of five physical parameters and three engineering-derived variables - the overall and primary-zone excess-air ratios and the equivalence ratio. The ensemble simultaneously predicts NOx, CO2, and CO with coefficients of determination of 0.941, 0.9997, and 0.925, respectively, with a single evaluation time below 0.04 s; the dataset was split by complete baseline operating conditions. The results show that adding plasma-chemical products at β > 0.005 (the volume fraction of plasma-chemical products) reduces CO emissions from 25-28 to 3.9-4.6 ppm with virtually unchanged NOx, whereas combining lean operation with the PCE creates, within the investigated range, an ultra-low-emission region with NOx ≈ 10 ppm and CO < 8 ppm. The effective computational speedup is 1-2 orders of magnitude for typical optimization tasks and reaches 3-4 orders of magnitude for large-scale parametric studies and real-time supervisory control. Keywords: plasma-chemical element, plasma-assisted combustion, gas turbine engine, combustor, computational fluid dynamics, neural network, surrogate model, NOx emissions, carbon monoxide. Tables 3, Figs 7, Refs 30.