AI-Driven Well-Spacing Optimization in Depleted Unconventional Reservoirs: A Data-Driven Approach to Maximizing Late-Stage Recovery

Abstract Depleted unconventional reservoirs present distinct challenges for well spacing due to pressure depletion, stress redistribution, and fracture-driven interference. Current well spacing strategies are limited in their ability to dynamically adapt to evolving reservoir conditions, leading to inefficient recovery. This study develops a machine learning (ML)-assisted optimization framework to determine the ideal parent-child well spacing in depleted unconventional plays. The workflow integrates reservoir properties, depletion trends, and hydraulic fracturing strategies, aiming to enhance late-stage hydrocarbon recovery while minimizing well interference and pressure losses. We integrate fully coupled 3D reservoir-geomechanics simulations with ML algorithms, including Extra Trees, Random Forest, and XGBoost, to predict cumulative oil production and pressure loss under a wide range of geological and operational scenarios. Latin Hypercube Sampling (LHS) is employed to systematically generate a high-resolution training dataset that spans variations in rock properties, completion designs, and depletion conditions. The resulting ML models achieve high predictive accuracy (R2 > 0.95), enabling rapid evaluation of production outcomes for different reservoir properties and hydraulic fracturing parameters. The optimization framework simultaneously maximizes cumulative oil recovery while minimizing pressure depletion, with recommendations for infill parent-child well spacing. Our results reveal that optimal spacing is non-monotonic and strongly influenced by local stress anisotropy and reservoir permeability. Production peaks occur at intermediate spacing (770–790 ft), where fracture communication is optimized without inducing excessive interference or pressure sinks. Completion parameters, such as proppant mass and cluster density, also exhibit nonlinear effects, with fewer, high-intensity clusters outperforming larger numbers of clusters. Pareto frontier and sensitivity analyses show that slight deviations from optimal spacing have a limited impact on production, affording flexibility in field applications. This study presents an ML-driven well spacing optimization framework designed specifically for depleted unconventional reservoirs. This AI-driven framework advances reservoir development planning by moving beyond fixed-spacing rules toward adaptive, data-informed decision-making. It enables operators to tailor well spacing to specific reservoir and completion conditions, improving late-stage hydrocarbon recovery while managing depletion-related risks. The methodology also supports future integration of real-time data and reinforcement learning for continuous optimization.