Physics Informed Neural Networks for Real Time 3D Flow Field and Emission Prediction in Lean Hydrogen Swirl Combustors

Abstract

Real-time monitoring and control of hydrogen-fueled gas turbines demand high-fidelity reconstruction of turbulent reacting flows and emissions. While conventional computational fluid dynamics (CFD) yields accurate velocity and thermal fields, the computational overhead precludes online digital twin integration. Purely statistical surrogates offer speed but violate physical conservation laws. To address these limits, we deploy a hybrid, data-assisted Physics-Informed Neural Network (PINN) mapping three-dimensional coordinates to velocity, pressure, temperature, and species fields within a co-swirling lean hydrogen-air combustor ( ). Chemical kinetics stiffness is bypassed by formulating the network to solve transport equations for conserved scalars (mixture fraction and enthalpy) while retrieving local properties and species from non-adiabatic Equilibrium PDF tables. The PINN is trained on a validated RANS dataset resolved via Menter’s SST -  model with curvature correction. The surrogate preserves mass conservation within 0.5%, limits relative errors to under 3%, and executes in 3.0 ms on a CPU.