Physics-Informed Neural Networks for Gas Hydrate Plugging Risk Assessment Using Intrinsic Kinetics and Flowloop Data

Abstract

Abstract A reliable means of hydrate plugging risk assessment in pipelines is critical to the modern practice of production in the hydrate management regime. Flow assurance engineers utilize computationally expensive multiphase flow simulations to characterize hydrate formation at desired conditions, however, there is no numerical method to assess the risk of a plug occurring from these results. Traditional machine learning models have shown reasonably accurate plugging risk classification and require just milliseconds to return an assessment. Despite this, there has been limited industry use due to concerns about the statistical nature of predictions and the sparsity of available training data. Deep neural networks (DNNs) are a purely data-driven machine learning model that require large quantities of labeled data to make accurate statistical predictions in their trained domain. Physics-informed neural networks (PINNs) are a variation of DNNs in which training additionally considers embedded domain physics, in the form of partial differential equations, to increase accuracy, lessen reliance on training data, and ground predictions. This work presents a PINN that has been trained to predict hydrate plugging risk. Training was directed by the mean squared error of the model's prediction against flowloop data and, critically, the residual of the hydrate intrinsic kinetics equation. The trained model showed improved accuracy over reference DNNs. A PINN of novel architecture embedded with the hydrate intrinsic kinetics equation was built in TensorFlow. Flowloop data from pilot-scale flowloops was used for the training and evaluation of the presented PINN. Performance was compared to two DNNs for plugging risk assessment. DNN1 was an earlier model presented at OTC 2019. DNN2 features identical architecture to the subject PINN but absent of the embedded physics. DNN1 was employed as a baseline for plugging risk assessment performance, whereas DNN2 was used to isolate the contribution of the embedded domain knowledge on inference accuracy. The PINN showed a plugging risk assessment accuracy of 98.7%, which is a meaningful improvement over the 95.0% accuracy offered by DNN1. Moreover, case studies show improved confidence in plug prediction. The effect of the embedded physics on model accuracy is quantified by a reduction in mean squared error of 13.3% in inference of hydrate volume fraction when compared to DNN2. These findings indicate that the increased accuracy is the result of the embedding of the hydrate intrinsic kinetics equation as well as the novel network architecture. Two additional PINNs were presented, further establish the superior behavior of PINNs in learning the solution to PDEs and under data-sparse conditions. This work provides a new approach for machine learning in hydrates by demonstrating a technique to accurately train neural networks through a combination of empirical data and domain knowledge. This line of research could ultimately lead to more informed quantification of hydrate plugging risk.