Fluid type identification of low-permeability reservoirs based on logging response and GWO–XGBoost model: A case from the 8th Member of Triassic Yanchang Formation in Hongde area, Ordos basin

Accurately identifying low-permeability sandstone reservoir fluid types is essential for unlocking their economic value. However, the complex lithology, strong vertical and lateral heterogeneity, diverse pore structures, and intricate oil–water distributions in these reservoirs result in weak and ambiguous logging responses, which limit the identification accuracy of traditional Archie–based petrophysical interpretations. By contrast, artificial intelligence–based methods provide a new pathway by enabling quantitative extraction of fluid information from complex logging data, effectively overcoming the limitations of conventional petrophysical models. This study develops an intelligent fluid type identification method for low-permeability sandstone reservoirs using the heterogeneous clastic strata of the Chang 8 Member of the Triassic Yanchang Formation in the Hongde area of the Ordos Basin as a case study. Using the lithological results of logging interpretation, the method directly incorporates logging curves sensitive to both fluid type and saturation as indicator variables and incorporates them into an eXtreme Gradient Boosting (XGBoost) ensemble model. The key methodological advancement involves the application of the Grey Wolf Optimizer (GWO) to automatically calibrate three critical XGBoost hyperparameters: maximum tree depth, learning rate, and number of estimators. In contrast to manual tuning or grid search methods commonly employed in previous applications, the global optimization capability of GWO facilitates the systematic exploration of the parameter space. It avoids local optima, and adaptively balances model complexity with generalization, thereby reducing overfitting and enhancing predictive stability. Validated against oil-testing results from 90 perforated intervals, the GWO–XGBoost model achieved a classification accuracy of 94.7%, with a precision of 0.963 and recall exceeding 90% across all fluid types. These results represent improvements of 4.2%, 5.0%, 20.6%, and 25.9% over the standalone XGBoost, random forest, support vector machine models, and the Archie method, respectively. This study introduces a hybrid approach that integrate an advanced optimizer with an existing ensemble learner to substantially enhance the predictive performance for fluid type identification in low-permeability reservoirs, without the need to develop a new algorithm. This study demonstrates that the GWO–XGBoost framework overcomes long-standing barriers in identifying fluid type identification within low-permeability sandstone reservoirs, offering a robust, cost-efficient solution for exploration and development.