Rock Physics and Machine Learning for Lithology Identification and Estimation of Unconventional Reservoir Properties

Unconventional reservoirs, characterized by strong heterogeneity and complex pore structures, pose challenges for conventional reservoir evaluation using well-logging data. In this study, elastic parameters including V-P/V-S ratio, P- and S-wave impedances, Poisson's ratio (sigma), and Lam & eacute; constants (mu and lambda) are calculated to identify lithologies, revealing three types: shale, sandstone, and muddy sandstone layers. Principal component analysis (PCA) reduces the multidimensional dataset to three principal components (PC1, PC2, and PC3), capturing maximum variance and aiding in lithology classification. K-means clustering applied to these components identifies three optimal clusters corresponding to the lithologies. Borehole logs are analyzed for lithology clustering, classifying distinct reservoir units. Six machine learning algorithms (KNN, SVR, GBR, RFR, Lasso, and ridge regression) are evaluated for predicting porosity, water saturation, and shale volume using two feature sets: the original set (V-P, V-S, and rho) and an integrated set (elastic parameters, PCA components, and clustering labels). Models are assessed using RMSE and R-2, with the integrated set yielding significantly improved accuracy. SHAP analysis reveals that PC2 has the most influence, followed by PC1, as these components effectively aggregate variance from multiple elastic parameters, enabling the integrated model to outperform the basic model that relies directly on individual features. Incorporating subsurface characterization techniques enhances well-log prediction accuracy using machine learning.