Abstract Accurate prediction of depletion-induced effective stress is essential to reservoir engineering decisions across the field life cycle, influencing well instability, completion and stimulation design, production management, and geomechanical risk assessment. Although fully coupled flow–geomechanics finite-element (FE) simulations provide high-fidelity stress estimates, the associated computational cost and long runtime limit routine operational use. Therefore, this study develops an artificial neural network (ANN) proxy to rapidly predict 3D effective stress distributions driven by production-induced pore-pressure changes. The proposed workflow integrates field-scale coupled simulations with large-scale supervised learning and independent spatial validation to improve both data scale and evaluation rigor relative to many published ANN proxies. A field-scale workflow was implemented using data from 10 wells in the Buzurgan oilfield. High-resolution stress responses were generated using a fully coupled simulator (CMG-GEM 2021) and converted into supervised input–target pairs through an automated Python-based pipeline, yielding 11.26 million training samples. A compact ANN (one hidden layer with three neurons) achieved strong agreement with simulator outputs (avg. R ² = 0.94) and reproduced spatial effective-stress patterns in a structurally independent full-field case after 10 years of production, with most grid cells exhibiting deviations within ± 200 psi. The proxy reduced turnaround time from approximately hours per coupled FE run to minutes per prediction, enabling near real-time stress screening and sensitivity analysis for operational decision-making.
Jubair et al. (Tue,) studied this question.