Development and validation of a physics-informed neural network-based WSGG model for multi-species gas mixtures

Accurate modeling of radiative heat transfer is critical for Computational Fluid Dynamics (CFD) simulations of high-temperature reactive flows, especially in combustion systems involving complex gas mixtures. The Weighted-Sum-of-Gray-Gases (WSGG) model is a widely used radiative spectral model due to its simplicity and efficiency. However, traditional WSGG models suffer from several inherent limitations: (1) the use of polynomial curve fitting to derive weighting factors and absorption coefficients restricts model accuracy, especially in capturing spectral nonlinear behavior; (2) most existing models are tailored to binary gas mixtures, making extension to multi-species systems inherently difficult; and (3) due to the lack of inherent physical consistency in model construction, parameters are often non-unique and require extensive tuning to achieve acceptable accuracy. To address these limitations, we introduce a novel Physics-Informed Neural Network (PINN)-based WSGG modeling framework. The approach employs two compact neural networks to predict weighting factors and absorption coefficients for the “gray gases”, trained using a dual-loss function that minimizes both emissivity and calculated radiative heat loss residuals, thereby ensuring accuracy and physical consistency. Validation against Line-by-Line (LBL) benchmarks for emissivity and 1D Radiative Transfer Equation (RTE) solutions confirmed the model’s accuracy. Applied to CFD simulation and compared against the Full-Spectrum Correlated-$k$ Distribution (FSCK) benchmark, the PINN-WSGG model demonstrated great agreement in simulations of a scaled Sandia D flame. This adaptable approach simplifies the development of WSGG models and easily accommodates multi-species mixtures. Although demonstrated for H₂O-CO₂-CO mixtures at atmospheric pressure, the methodology can be readily extended to other gas mixtures and operating conditions.