Optimization of oil production rates under geomechanical stress constraints: a coupled reservoir–geomechanics approach using cascaded artificial neural network surrogates

Geomechanical constraints can limit oil production from deep carbonate reservoirs, because depletion raises the effective stress and, with it, the risk of compaction and rock failure. Enforcing such a limit inside a production optimizer is costly, since every candidate rate schedule produces its own pore-pressure field and one coupled reservoir–geomechanical simulation takes about 3 h. This study develops a cascaded artificial neural network (ANN) surrogate to maximize oil production from the MB21 unit of the Mishrif Formation, Buzurgan Oil Field, while keeping the effective mean stress below 11,200 psi. A Pressure ANN predicts the pore-pressure field from the well rates and geomechanical inputs, and a Stress ANN converts the predicted field into effective mean stress. Unlike single-network stress surrogates, which need pore pressure as a known input, the cascade evaluates the stress constraint for the depletion that each proposed rate schedule would itself cause. The networks were trained on 2,446 high-fidelity CMG-GEM coupled simulations and coupled with a Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) to allocate the rates of 18 producers. On independent test scenarios, the Pressure and Stress ANNs reached coefficients of determination of 0.95 and 0.94. PSO produced 42,799.02 bbl/day against 42,592.96 bbl/day for GA (+ 206.06 bbl/day, 0.48%) and reached a near-optimal solution within about 25 iterations; both algorithms identified Well 8 as the most geomechanically sensitive producer. Re-simulation of the PSO schedule with the original CMG-GEM model over the seven-year horizon delivered the allocated 42,799.02 bbl/day with a maximum effective mean stress of 11,198 psi, and no violations of the 11,200 psi constraint were observed over the 155,540 evaluated spatio-temporal points. With one candidate evaluated in about 0.13 s, a complete optimization run of about 3 × 10 4 evaluations took about 1 h. The cascaded surrogate therefore makes geomechanically constrained rate optimization tractable at field scale, provided the selected schedule is confirmed by a final high-fidelity simulation.