Algorithms for optimizing gas injection modes to enhance oil recovery with constraints on reservoir water preservation and prevention of induced seismic impacts

Abstract

The study proposes an algorithmic framework for optimizing gas injection modes to enhance oil recovery of mature terrigenous fields while simultaneously satisfying two critical constraints: controlling water cut growth and preventing induced seismicity caused by excessive pore pressure near tectonic faults. Methodologically, the work relies on a full-scale digital twin of the reservoir (3D grid of 250×300×40 cells) integrating 3D seismic data, interpretation of well logs from 87 wells, PVT analysis, and a 15-year production history. Flow simulation is carried out using a compositional black oil model with careful calibration of relative phase permeabilities, and the geomechanical module is calibrated against microseismic data and well tests. The central element is a hybrid optimizer (genetic algorithm for global search combined with conjugate gradients using adjoint modeling for local refinement) that solves a multi-objective problem in the form of unconstrained optimization with two penalty terms for exceeding the water cut and critical pressure thresholds near faults. Ten-year verification showed that compared with the baseline project mode with constant injection volumes, the optimized strategy increases cumulative oil production by 8.19%, reduces total gas injection by 5.36%, decreases water production by 21.33%, and lowers average bottomhole injection pressure by 5.59%, indicating improved energy efficiency and sweep coverage. Geomechanical constraints are met: maximum pressures at control points near the fault remain below the seismic hazard threshold (34.5 MPa), while the baseline scenario exceeds it. Sensitivity analysis revealed a nonlinear trade-off: tightening the water cut limit from 70% to 40% reduces the final recovery factor from 0.418 to 0.375, while the specific gas factor is minimized around 50% due to thinner regimes. In terms of computational efficiency, the hybrid algorithm converges 39.5% faster than a pure genetic algorithm on a small model and provides acceptable computation times on a detailed grid for planning tasks; future developments include parallelization and proxy models to operate in near-real-time conditions.