Machine-Learning-Driven Permeability Prediction & Improved Reservoir Characterization in Heterogeneous Sandstone Formation of KG Basin

Estimating permeability and reservoir characterization in heterogeneous reservoirs like the Mandapeta Formation in the Krishna Godavari (KG) Basin poses significant challenges due to limited core and NMR data. This study addresses these challenges using two machine-learning approaches to predict permeability and characterize the reservoir into moderate-to-good & poor rock types, providing critical insights into reservoir quality and production performance. The first approach uses a hybrid ensemble model to predict permeability directly from standard well logs, including processed outputs like effective porosity and clay volume. Separate models were trained for high and low permeability ranges, achieving an R² score of over 85%. Data conditioning and augmentation enhanced model accuracy, effectively managing the formation's complex heterogeneity. The second approach integrates NMR and core data to classify the reservoir into five rock types using pore-throat radius distributions derived from Winland plot. A Multi-Resolution Graph-Based Clustering (MRGC) model predicted rock types using basic log curves, enabling permeability estimation through porosity-permeability (Poro-Perm) relationships. Both methods produced permeability predictions strongly aligned with actual NMR permeability in test wells. Production performance correlated with the thickness of high-quality flow units, highlighting their significance in hydrocarbon production. This dual approach not only enables permeability estimation in data-scarce wells but also enhances reservoir characterization. These insights provide a robust framework for optimizing field development and production strategies in heterogeneous formations.