Comparative Study of ANN and 3D-CNN Proxy Models for Predicting Reservoir Geomechanical Stress Distribution

High-resolution geomechanical simulations provide accurate representations of reservoir behaviors but are computationally expensive, with typical runtimes many hours per scenario, rendering optimization under realistic decision time scales impractical. This study develops and evaluates two intelligent proxy models using deep learning—an Artificial Neural Network (ANN) and a 3D Convolutional Neural Network (CNN)—to significantly reduce computational costs while maintaining predictive accuracy. The models were trained on an extensive dataset comprising 381 million data points generated from 3,000 high-resolution finite-element geomechanical simulations. Following an 80/10/10 split at the scenario level to prevent data leakage, the CNN model demonstrated superior performance on the 10% hold-out test set, achieving a mean coefficient of determination (R²) of 0.94 and a mean root mean square error (RMSE) of approximately 96 psi, compared to the ANN model which achieved a mean R² of 0.93 and RMSE of 145.3 psi. Comprehensive field validation was performed using a real-world case study from the Mishrif formation in southern Iraq. The CNN maintained its high accuracy with an R² of 0.92 and RMSE of 108 psi, while the ANN achieved an R² of 0.90 and RMSE of 169 psi. The CNN also exhibited excellent spatial generalization, accurately capturing localized stress concentrations near active production wells. Most significantly, the developed proxy models reduced computational time by over 95%, decreasing simulation runtime from 2–3 hours to less than 10 minutes per scenario. This substantial efficiency gain enables near real-time reservoir monitoring, fracture risk analysis, and wellbore stability assessment throughout the production life of the reservoir.