Rollout-Aware Graph Neural Network-Based Simulator for Hydraulic Fracture

Abstract

Abstract Hydraulic fracturing is a key technology for the efficient development of unconventional oil and gas reservoirs, and numerical simulation is an effective approach for investigating hydraulic fracture propagation. However, conventional hydraulic fracture propagation models struggle to achieve a satisfactory balance between computational efficiency and predictive accuracy, making it difficult to simulate fracture evolution in real time as field operations progress. Incorporating machine learning into next-generation hydraulic fracture simulators offers a promising route toward real-time prediction of fracture propagation. In this study, we develop, for the first time, a hydraulic fracture propagation simulator based on a rollout-aware graph neural network (RAGNN), termed RAGNN-FPS, to enable second-scale prediction of hydraulic fracture evolution. The simulator is designed to facilitate rapid adjustment and optimization of field fracturing parameters and thereby improve reservoir stimulation performance. RAGNN-FPS learns from massive, high-fidelity, mesh-scale simulation data generated using the finite-discrete element method (FDEM), including displacement fields, velocity fields, and hydraulic fracture propagation states. A heterogeneous graph containing both solid-element edges and joint edges is constructed, while static geological parameters and dynamic engineering parameters are encoded as node and edge features. A rollout-aware training strategy is further introduced, whereby errors arising from the model's own predictions are progressively incorporated during training. This strategy mitigates error accumulation during autoregressive inference and improves long-horizon prediction stability. The results show that the proposed simulator can stably reproduce the complete hydraulic fracturing process, including fracture initiation, propagation, and interactions with natural fractures, while achieving second-scale prediction of fracture evolution. Compared with the teacher-forcing baseline, the median relative L1 error of the full-field displacement is reduced to below 0.073, while the F1 score for joint failure-state classification increases from 0.197 to 0.965. All test cases lie outside the parameter ranges used for training, with extrapolation levels ranging from 25% to 150%. Under these extrapolation conditions, the relative error of the displacement field increases from approximately 5% for interpolation to approximately 10%. On the same workstation, prediction of one complete trajectory requires approximately 9 s, whereas an FDEM simulation of the same scale takes approximately 10 min, corresponding to a speedup of about 65 times. Evaluation of 50 design scenarios requires only approximately 8 min. By integrating a heterogeneous graph neural network architecture with a rollout-aware training strategy, RAGNN-FPS provides a pioneering learning paradigm for real-time simulation of hydraulic fracture propagation and offers a potential solution for digital-twin physical modeling of fracture propagation in geotechnical systems.