Machine Learning‐Driven Classification and Production Capacity Prediction of Tight Sandstone Reservoirs: A Case Study of the Taiyuan Formation, Ordos Basin

ABSTRACT Tight sandstone gas (TSG), an unconventional oil–gas resource, has heterogeneous reservoirs. Traditional classification and evaluation methods fail to fully reveal reservoir characteristics and complex reservoir‐productivity links. As a key source of tight sandstone, the Taiyuan Formation in the Ordos Basin exhibits significant variations in reservoir quality due to its marine–terrestrial transitional sedimentary setting, necessitating the establishment of an effective classification and evaluation system. This study focuses on Block A's Taiyuan Formation tight sandstone reservoirs in the Ordos Basin, using core observations, thin‐section petrography, scanning electron microscopy, and high‐pressure mercury intrusion experiments to build a machine learning (ML) model for reservoir evaluation and improve the accuracy of reservoir evaluation. Results demonstrate that the K ‐means algorithm has the optimal classification results based on a variety of internal evaluation metrics. The tight sandstone reservoirs in the study area are classified into three distinct types: Type I reservoirs are mainly lithic quartz sandstone/lithic feldspathic sandstone, showing significant compaction, with large pore throats and low displacement pressure. Type II, mostly feldspathic lithic sandstone, acts as a transitional reservoir. Type III, predominantly lithic feldspathic sandstone, is most affected by cementation, featuring fine‐grained pore throats and high displacement pressure. Furthermore, evaluation results based on confusion matrices demonstrate that the Bayesian‐optimized Random Forest model achieves a classification accuracy of 92.76% in log interpretation, significantly outperforming other models. Through gray relational analysis, the nonlinear correlations between productivity and reservoir parameters are clarified. The prediction results demonstrate over 80% spatial consistency with actual gas reservoir distributions, effectively reflecting the production capacity of the tight sandstone reservoirs in the study area. The ML‐driven evaluation approach proposed in this study not only provides robust technical support for the exploration and development of TSG reservoirs and effectively delineates the distribution of favorable zones, but also offers practical guidance for the efficient exploration of hydrocarbon resources in geologically analogous settings.