Machine learning-based sweet spot prediction for lacuscrine shale oil in the Weixinan Sag, Beibu Gulf Basin, China

Abstract

This study aims to overcome challenges related to sparse well distribution, pronounced lithological heterogeneity, and limited core data in offshore lacustrine shale oil reservoirs through the development of an integrated evaluation framework that enables quantitative prediction and spatial mapping of sweet spots. Conventional well logs, nuclear magnetic resonance (NMR) data, geochemical analyses, and laboratory measurements are integrated to construct a grading system for four key properties—reservoir quality, oil-bearing property, mobility, and fracability—with threshold values determined by statistical distributions. To address sample scarcity, a Generative Adversarial Network (GAN) is applied for data augmentation, while Recursive Feature Elimination with Cross-Validation (RFECV) is employed for optimal feature selection. Considering the complex, nonlinear, and anisotropic nature of shale reservoirs, a Generalized Regression Neural Network (GRNN) is constructed, with its smoothing parameter globally optimized using the Sparrow Search Algorithm (SSA), resulting in high predictive accuracy (R² > 0.8). Furthermore, the study establishes a coupling relationship between lithofacies assemblages and the four properties. This integrated approach supports spatial prediction of both individual parameters and the comprehensive sweet spot evaluation index. The results indicate that laminated (LSA) and interbedded (ISA) shales exhibit distinct sweet spot characteristics (porosity >6 %, clay content <25 %, and OSI >100) with high-potential zones primarily located in the central W1, southeastern W2, and central W3 regions. This methodology offers a reliable technical framework for quantitative assessment and development planning of offshore lacustrine shale oil resources under data-constrained conditions.