Enhancing proper orthogonal decomposition with physics-informed neural networks for wet gas compression analysis

Wet gas compression technology is increasingly vital for natural gas and power generation applications, but its development is hindered by an insufficient understanding of nonlinear liquid–gas interactions. Proper orthogonal decomposition (POD), a widely used mathematical tool in turbomachinery analysis, is based on linear kernel assumptions. However, in highly unsteady and nonlinear flow fields—such as those encountered in wet gas compressors—this linearity results in reconstruction distortions and ambiguous modal interpretations, thereby limiting the applicability of the POD method. This paper proposes an enhanced POD method based on physics-informed neural networks, NN POD, which combines prior physical principles with data-driven techniques. Based on the results of a full-channel transient simulation of a centrifugal compressor, this study first evaluates the reconstruction results of two modal decomposition methods and elucidates the advantages of NN POD. Based on this, an in-depth flow field analysis was conducted under wet gas conditions, achieving visualization of key flow mechanisms. This study found that under wet gas conditions, the flow within the channel can be improved by alleviating the tip clearance flow, reducing vortex shedding and circumferential convection in the tip region, thereby enhancing the pressure ratio and efficiency. However, the wake effect of the water droplets in the wet gas leads to high-amplitude pressure fluctuations at the trailing edge of the blades and in the diffuser. This issue should be addressed in industrial design and manufacturing.