Real-Time Fracture Geometry Inversion, Geomechanical Updating, and Hydraulic Fracturing Design Optimization Through AI Surrogate Modeling and Data Assimilation

Abstract

Efficient multistage hydraulic fracturing requires reliable characterization of subsurface fracturing conditions and adaptive adjustment of treatment parameters as stimulation proceeds. This paper presents an intelligent stage-by-stage framework that integrates pump-schedule optimization, real-time fracture-geometry inversion, and geomechanical updating for multistage hydraulic fracturing. The framework builds upon data-driven fracture inversion and further connects pressure-conditioned subsurface characterization with the optimization of subsequent fracturing stages, enabling continuous and adaptive improvement throughout the multistage operation. A calibrated numerical hydraulic-fracturing model was first established using the open-source MSEEL MIP-3H well dataset. Systematic variation of geomechanical properties and treatment parameters generated simulation databases for training AI surrogate models. During pumping, a variational autoencoder and a dynamic treatment-response surrogate are coupled with an Ensemble Kalman Filter to assimilate field treatment pressure, dynamically estimate fracture propagation geometry, and predict fracture width and area. After each stage, the complete treatment-pressure history is assimilated through a static pressure surrogate and the Ensemble Kalman Filter to update the minimum horizontal stress, horizontal stress difference, and Young's modulus. These pressure-conditioned estimates are then transferred to the subsequent stage. Before the next treatment, Tabular Prior- data Fitted Network (TabPFN) surrogate models combined with constrained Non-dominated Sorting Genetic Algorithm III (NSGA-III) optimize pumping rate, fluid volume, and proppant mass by balancing cumulative gas production and treatment cost under operational constraints. The framework is demonstrated using Stages 23-25 of the MIP-3H well. Real-time pressure assimilation achieves posterior R2 values of 0.9967-0.9999 and mean absolute errors of 4.6-13.2 psi. The inverted fracture networks reproduce the dominant propagation directions and cover 82.7%-90.4% of the corresponding microseismic extents along the wellbore-normal direction. The optimized Stage 24 and Stage 25 schedules increase simulator-verified cumulative gas production by 9.97% and 13.44%, respectively, while the Stage 25 treatment cost is reduced by 11.96%. These results demonstrate that information obtained from a completed fracturing stage can be effectively transferred to the characterization and optimization of subsequent stages. The proposed coupled inversion-optimization framework provides field engineers with a practical, flexible, computationally efficient, and reliable tool to dynamically characterize subsurface fracturing conditions and adaptively optimize treatment parameters throughout multistage hydraulic fracturing operations.