Causal Inference-Based Hybrid Model for Distinguishing Geological Controls in Tight Sandstone Stimulation: A Construction-Parameter-Free Framework with Field Monitoring Data

Introduction: Accurately discerning the influence of geological factors on well productivity remains a significant challenge in tight sandstone reservoirs due to substantial interference from highly variable fracture stimulation parameters. This study aims to develop a reliable methodology to isolate geological controls and predict stimulation sweet spots, independent of inconsistencies in completion design. Methods: An integrated machine learning framework was developed, combining an improved Analytic Hierarchy Process (AHP), Grey Relational Analysis (GRA), and Entropy Weight Method (EWM) to quantify the contribution of geological and engineering factors to productivity. Construction-related influences were normalized using key completion parameters. A Back Propagation Neural Network (BPNN) was subsequently employed to establish a robust correlation model between the dominant geological factors and the normalized productivity, forming the basis for a comprehensive sweet spot identification method. Results: The model demonstrated high accuracy in predicting stimulation sweet spots, showing strong agreement with actual production data. Application in Block X identified reservoir thickness, permeability, porosity, horizontal stress difference, oil saturation, and brittleness index as the principal geological controlling factors. Implementing a multi-cluster differentiated fracturing design in high-sweetness stages resulted in a 25.28% increase in initial daily oil production and a 136% rise in cumulative production over 15 years. Furthermore, optimized well placement and intensive fracturing in high-sweetness areas improved the recovery rate by over 12%. Discussion: The results confirm that the proposed framework effectively eliminates the confounding effects of completion parameters, enabling a reliable, data-driven identification of geological sweet spots. This underscores the critical importance of decoupling geological potential from engineering effectiveness for optimizing development strategies in tight reservoirs. The method provides a transformative approach that moves beyond traditional, completion- biased interpretations. Conclusion: The study successfully establishes a construction-parameter-free machine learning framework for identifying geological sweet spots in tight sandstone reservoirs. This method offers practical and decisive guidance for well placement, stage selection, and differentiated fracturing design, demonstrating significant potential to enhance production and recovery rates. The approach possesses broad applicability in similar unconventional resources.