Gas-liquid two-phase flow rates measurement using physics-guided deep learning

Machine learning has been widely used in the gas-liquid two-phase flow field. However, data-driven machine learning methods are often considered black-box methods which may lead to insufficient interpretation and extrapolation abilities. Combining machine learning and physical principles is a promising method to improve the generalization abilities and interpretation of the deep neural network model, which is still a challenge in gas-liquid two-phase flowrates measurement. In this work, we propose a physics-guided deep learning model for gas-liquid two-phase flow rate measurement through the differential pressure meter. The simplified physical constraint term of gas-liquid two-phase flow is established based on the momentum equation and continuity equation of gas-liquid two-phase flow, and then, the physical constraint term is added to the loss function of the deep neural network. Consequently, the predictions of the model not only satisfy the target values but also conform to the physical term. The performance of the physics-guided neural network (PGNN) is tested and compared with an artificial neural network (ANN) model that is trained without the physical constraint term. Results show that the gas and liquid flow rates could be measured simultaneously with satisfactory accuracy using the proposed PGNN model, more importantly, the model also performs well on a wider range of testing conditions (Gas Volume Fraction, GVF>95%). It indicates that this method is possible to improve the generalization and interpretation of the purely data-driven neural network model.