Machine learning (ML) has advanced reservoir engineering by enabling data-driven optimization for improved injection performance and cost efficiency. This study develops three supervised ML regression models trained on high-fidelity numerical data from an industry-standard black oil simulator to predict key parameters governing injectivity and fracture propagation in horizontal water injectors under a constant-pressure boundary scenario: bottom-hole pressure (BHP), fracture-tip pressure, and the percentage of injection rate entering the fracture. The pressure quantities are expressed in terms of pressure drops relative to the reservoir boundary, defined as delpres (well-to-boundary pressure drop) and delptip (fracture-tip-to-boundary pressure drop), from which absolute pressures can be readily obtained. Conventional analytical models rely on oversimplified assumptions, while fully coupled numerical simulations incur prohibitive computational cost. A comprehensive dataset of 30 000 numerical simulations was generated, capturing near-wellbore and fracture-tip dynamics across diverse reservoir, fracture, and operational conditions. Key input features include reservoir thickness, porosity, permeability anisotropy (permeabilities in the x, y, and z directions), rock and water compressibility, water viscosity, horizontal wellbore radius and length, fracture half-length and width, fracture conductivity, skin factor, and near-fracture damage. Eleven ML algorithms—CatBoost, LightGBM, XGBoost, Random Forest, Gradient Boosting, Support Vector Regression, Linear, Ridge, Lasso, Elastic Net, and K-Nearest Neighbours—were rigorously trained and evaluated. CatBoost achieved MAE ≈ 30 psi and R 2 ≈ 98%–99% across targets. Although trained on static configurations under constant-pressure lateral boundaries and no-flow top and bottom boundaries, these surrogates enable rapid evaluation of horizontal well injectivity and fracture propagation potential. The models can be coupled with time-dependent skin evolution and dynamic fracture growth mechanics to simulate real-time system response in field-scale injection operations.
Singh et al. (Mon,) studied this question.