Integrated artificial intelligence approach for well-log fluid identification in dual-medium tight sandstone gas reservoirs

Introduction With the development of complex tight sandstone oil and gas reservoirs, accurately and cost-effectively characterizing these reservoirs have become a critical yet challenging task. To address the limitations of conventional machine learning algorithms, which have low accuracy due to data inhomogeneity and weak fluid logging responses, this study introduces a novel method for fluid logging evaluation in dual-medium tight sandstone gas reservoirs. Methods The method integrates core, thin section, and scanning electron microscope observations, taking into account the effect of fractures. Results Reservoirs are divided into three types: fractured reservoirs (FR), porous reservoirs (PR), and microfracture-pore composite reservoirs (MPCR), highlighting the distinct fluid logging responses of each type. Reservoir classification based on geological genetic mechanism significantly reduces data noise and prediction ambiguity, thereby improving the efficiency of model training. Discussion The final model is constructed by an ensemble method that integrates multiple sub-models, including fuzzy C-means clustering (FCM), gradient boosting decision tree (GBDT), backpropagation neural network (BPNN), random forests (RF), and light gradient boosting machines (LightGBM). Applied to the West Sichuan Depression in the Sichuan Basin, the model validation accuracy reached 91.96%. In summary, this novel and reliable method for log fluid prediction, significantly improved its accuracy and robustness compared with single models and traditional methods, providing a comprehensive perspective across geological and geophysical disciplines for fluid logging evaluation in dual-medium tight sandstone gas reservoirs.