An interpretable ensemble machine-learning workflow for permeability predictions in tight sandstone reservoirs using logging data

Abstract

Tight sandstone reservoirs exhibit strong vertical heterogeneity and complex pore structures, challenging conventional permeability evaluation methods based on well-logging data. Although rising machine-learning (ML) techniques have demonstrated excellent accuracy for industrial applications, the physics and rationality within such a powerful “black box” remain less clear. Hence, reliable permeability prediction would benefit from an interpretable ML-based workflow that could reveal the controlling factors. To compare the models and examine the underlying features, 16 different ML submodels are tested after data preprocessing, feature selection, and hyperparameter optimization. By comparing the fitting accuracy and tuning time, the light gradient boosting machine optimized by the whale optimization algorithm, referred to as LGB-WOA, is determined to be the optimal model with the best fitting accuracy and relatively short tuning time. A field data application demonstrates that even in highly heterogeneous reservoir sections, the LGB-WOA model outperformed conventional petrophysical models by being the most consistent with reservoir permeability directly measured from the core samples ([Formula: see text]). The Shapley additive explanation values are then used to interpret the predictions of our LGB-WOA model. As expected, the porosity curve exhibits the highest feature importance among all input features, significantly contributing to permeability predictions. Conversely, a wellbore diameter and compensated neutron log contribute the least and need not be used for subsequent model improvements. These experiments and workflow provide a powerful method for accurately assessing the permeability in complex reservoirs and contribute to a broader understanding of the application of ML in reservoir characterization, paving the way for establishing more interpretable and reliable prediction models.