Physics-Informed Neural Networks for the Prediction of Critical Sand Transport Velocity in Oil and Gas Pipelines

Abstract

Many reservoir formations are poorly consolidated therefore generation of sand with hydrocarbons from them cannot be avoided. The petroleum pipeline's condition may be seriously jeopardized by the creation of sand at very high velocities or by the deposition of that sand at low velocities, which can cause degradation and reduced capacity, respectively. Thus, it's critical to look into the minimal transport velocity that keeps sand from building up in the pipelines. In this study, two machine predictive tools, Physics informed neural network (PINN) and Multilayer perceptron (MLP) regressor were used for the estimation of the minimum transport condition in a solid-liquid-gas pipeline. The models were developed using 182 experimental data points containing superficial velocity, pipe diameter, particle diameter, sand density, sand concentration, liquid density, viscosity, pipe angle of inclination, and critical velocity parameters. The results showed that the physics-informed neural networks performed better than the MLP Regressor with an R2 coefficient of 0.9999 and RMSE of 0.00465. The MLP Regressor achieved an R2 coefficient of 0.9992 and RMSE of 0.0295. The developed models were compared with existing empirical models and data-driven models and from the outcome, the two models performed better in predicting the minimum transport velocity with PINN performing best. A sensitivity analysis was conducted and from the results, the most influential parameter for the minimum transport velocity prediction is the superficial velocity, followed by pipe diameter, sand concentration, and pipe angle. This research showed the potential of incorporating physical laws in the machine learning training process for a more accurate estimation of the minimum transport velocity in multiphase pipelines.