Prediction of hydraulic fracture parameters in tight gas reservoir using physics-constrained neural network

Abstract

The prediction of hydraulic fracture parameters is crucial for hydraulic fracturing evaluation and design. For field scale problems, the geometries of hydraulic fractures are mostly evaluated with numerical simulations, assisted by monitoring methods such as micro-seismic, tracer, and optical fiber techniques. However, such simulations are often time-consuming and difficult to meet the computational efficiency requirement for treatment parameter optimizations. In this paper, a physics-constrained neural network (PCNN) model, with the modified PKN model as its loss function, is proposed to predict the fracture parameters in tight gas reservoirs. Random search hyperparameter optimization, 10-fold cross validation, and ensemble learning are further used to increase the model accuracy. This model is systematically validated through hydraulic fracturing numerical simulations and field monitoring data. The results indicate that compared with the modified PKN model and the deep neural network (DNN), the PCNN shows the best generalization ability and prediction accuracy, while also avoiding predictions that violate physical laws. For heterogeneous reservoirs, the PCNN model can still provide a fast and reasonable prediction of fracture parameters.