Domain-decomposition physics-informed neural network approach for the acoustic boundary impedance estimation of a combustion chamber model

Acoustic simulations of combustion chambers often exhibit significant uncertainties owing to our limited understanding of impedance boundary conditions. Traditional impedance inversion methods based on optimization or sampling often produce substantial prediction errors and entail considerable computational costs in complex acoustic simulations. Recently, the impedance inversion methods based on physics-informed neural networks (PINNs) have emerged as promising solutions to overcome these challenges. The PINN approach replaces forward models with a fully data-driven modeling framework, thereby eliminating the iterative optimization process of traditional inversion methods. However, when applied to combustion chamber acoustics, conventional PINN-based impedance inversion methods struggle to resolve these challenges due to localized source-term distributions, and abrupt variations in flow parameters induced by the flame region. Furthermore, obtaining global noise training data poses challenges for practical applications. Accordingly, this paper proposes a domain-decomposed physics-informed neural network (DD-PINN) approach for acoustic boundary impedance inversion in combustion chambers. Case studies were performed on two numerical examples featuring complex geometric shapes and heterogeneous acoustic fields. The results demonstrate that the DD-PINN-based impedance inversion method effectively captures the acoustic characteristics of combustion chambers and accurately inverts inversion of unknown boundary impedances, offering a viable and efficient solution for complex combustion chamber acoustic impedance inversion problems.