Physics-Simulation-Informed Conditional Surrogate for Reconstruction of Multi-Cluster Hydraulic-Fracture States
Abstract
Multicluster hydraulic-fracture growth is governed by stress interference and time-varying fluid partitioning. High-fidelity simulation is therefore costly to update frequently during a treatment. This study develops PGStateNet, a conditional surrogate that reconstructs clusterlocal fracture states from physics-generated sequences synchronized with recorded pumping schedules. The reference calculation retains displacement-discontinuity elastic coupling, fracture flow, mass conservation, perforation-entry loss, and dynamic cluster-rate allocation. Three-dimensional fracture elements are projected onto fixed maximum-aperture grids. PGStateNet carries the current grid forward as explicit spatial memory and combines it with a 17-component treatment, completion, and geomechanical condition vector. Perforation- centred, distance, and local-support maps provide spatial context. A FiLM-modulated residual encoder-decoder then predicts a bounded aperture increment. Footprint and projected length–height losses complement the aperture-field loss. On 2,110 transitions from nine complete stages excluded from fitting and validation, PGStateNet achieved an active-field MAE of 0.237 mm and a footprint IoU of 0.902. The projected length and height MAEs were 4.68 mand 1.04 m, respectively. Under the same stage-held-out protocol, PGStateNet outperformed the original CNN-MLP and the AR-DenseNet-style baseline. Autonomous rollout began from the initial physical-model field and subsequently updated each cluster recursively. In these tests, PGStateNet preserved projected height more consistently than the two comparators. Across 13 microseismic-matched stages, PGStateNet-derived envelopes had MAEs of 5.1 m for length and 1.1 m for height, corresponding to mean absolute percentage errors of 1.2% and 4.3%. The framework supports one-step reconstruction, spatial diagnosis, and bounded autonomous state updating within the sampled treatment domain.