Prediction of Gas-Liquid Flow Parameters in Pipes Based on Physics-Informed Neural Network

Gas-liquid two-phase flow is a prevalent phenomenon in unmanned industrial settings, and understanding the vortex behavior in pipes is crucial for its analysis and prediction. The occurrence of vortex is intricately linked to fluid velocity, pressure distribution, and pipe geometry. Accordingly, the presence of vortex induces friction and vibration on the pipe wall, thereby impacting the mechanical properties of the pipe. Presently, many investigations on gas-liquid two-phase flow vortex in pipes rely on conventional computational fluid dynamics (CFD) methods, which suffer from lengthy computational cycles. Conversely, some studies employ deep learning techniques for flow field prediction, albeit requiring extensive data. To address these challenges, this paper simplifies the model to a homogeneous flow, exploits the physical information neural network for spatial and temporal prediction of parameters distribution in gas-liquid two-phase flow within pipes, as well as proposes the utilization of the weak-constrain 4DVar approach to refine the data. This framework resolves the computational inefficiency and data-intensive nature of CFD and NN method. Moreover, the impact of different neural network structures on prediction accuracy is investigated. By comparing with the validation data, it is observed that the method proposed in this study achieves an accuracy with an error less than 2%, which has a SOTA performance, and the prediction time is shortened to less than twenty minutes compared with the dozens of hours required by traditional CFD.