Optimization of SAGD well elevation utilizing characteristics of vertical permeability distribution in multi-realized 3-D reservoir models

• Developed a vertical permeability-based framework for SAGD well elevation optimization. • Introduced a steam-chamber-shaped filter to quantify continuous pay thickness in 3-D heterogeneous reservoir model. • Achieved strong agreement with simulation results (R 2 ≈ 0.83–0.95 for calibration sets and 0.67–0.97 for blind test sets). • Performed probabilistic optimization across 50 sub-models under geological uncertainty. • Delivered comparable results at total 12 days runtime vs 10 months of numerical simulation. Optimizing the vertical placement of steam-assisted gravity drainage (SAGD) well-pairs is challenging in oil sands reservoirs due to geological uncertainty associated with low-permeability zone (LPZ). While numerical simulation can capture LPZ effects, it is computationally impractical for probabilistic well placement optimization using multi-realized reservoir models. This study proposes a computationally efficient framework for SAGD well elevation optimization that explicitly incorporates LPZ distribution using vertical permeability. The framework employs a steam-chamber-shaped filter to parameterize LPZ characteristics and estimates continuous pay thickness as a proxy for bitumen accessibility. Two field-scale, geostatistics-based reservoir models with contrasting LPZ lateral continuity were constructed, and representative sub-models were selected using a newly developed static ranking method. Simulation-based optimal well elevations were used to calibrate framework parameters under different LPZ scenarios and bitumen price conditions. Optimization results indicate that optimal well elevation is controlled primarily by LPZ distribution rather than bitumen price, with vertical permeability between the injector and producer playing a dominant role. Using calibrated parameters, the framework reproduced simulation-based optimal well elevations with high accuracy, yielding R-squared values of 0.83–0.95 for calibration sets and 0.67–0.97 for blind test sets. Probabilistic optimization showed that final optimal elevations varied by only 0–1 m across LPZ scenarios. The proposed framework provides a practical alternative to exhaustive simulation-based optimization for SAGD well placement under geological uncertainty.