Application of machine learning for permeability prediction in heterogeneous carbonate reservoirs

Accurate prediction of reservoir permeability based on geostatistical modeling and history matching is often limited by spatial resolution and computational efficiency. To address this limitation, we developed a novel supervised machine learning (ML) approach employing feedforward neural networks (FFNNs) to predict spatial permeability distribution in heterogeneous carbonate reservoirs from production well rates. The ML model was trained on 25 black oil reservoir simulation cases derived from a geologically realistic representation of the Upper Kharaib Member in the United Arab Emirates. Input features for training included cell spatial coordinates ( x i , y i , z i ) , distances between cells and the n closest wells, and corresponding time-weighted oil production rates extracted from simulation outputs for each well. The target output was the permeability at each cell. The grid consisted of 22,739 structured cells, and training scenarios considered different closest well counts ( n = 1, 5, 10, and 20). The prediction performance of the trained model was evaluated across 12 unseen test cases. The model achieved higher accuracy with increased well input ( n ), demonstrating the potential of ML for efficient permeability estimation. This study highlights the effectiveness of integrating physical simulation outputs and spatial production patterns within a neural network structure for robust reservoir characterization. • FFNN approach predicts carbonate reservoir permeability from production data. • Trained on 25 COSTA model realizations with 22,739 cells and 248 production wells. • Novel feature engineering using distance-weighted well production for permeability.