Kinematic physics-informed neural network for continuous hydraulic fracture network reconstruction

Real time estimation of fracture network characterization during injection remains one of the biggest open problems in unconventional reservoir stimulation. Physics based simulators offer deterministic forecasts but do not incorporate downhole measurements collected during pumping, while data driven methods cannot maintain physical consistency under distribution shifts. This paper proposes cvPINN RTK, a kinematic physics informed neural network architecture that continuously reconstructs hydraulic fracture geometry using field DAS records. The term kinematic refers to the time evolving nature of the fracture network, capturing fracture initiation, propagation, branching, and coalescence as functions of injection time from recorded DAS strain data, rather than performing a static inversion. The model incorporates offline multi case physics simulator pretraining, DAS signal decomposition using Hankel singular value decomposition and adaptive poroelastic wavelet denoising, an innovative Model Confidence Index that dynamically balances data and physics constraints, and a two-source ensemble Kalman filter that assimilates DAS strain records with simulated values using adaptive weights. Validation on a hybrid triplet library from shale and hot dry rock reservoirs demonstrates 90.1 percent out of distribution detection, 93.2 percent fracture half-length accuracy, 89.7 percent fracture height accuracy against independent microseismic benchmarks, and an average SRV overlap ratio of 0.786 with calibrated uncertainty bounds.