Coordinated and efficient prediction of multiple parameters for complex shale reservoirs based on deep learning: A case study of Qingshankou Formation in Songliao Basin

The Upper Cretaceous Qingshankou Formation of the Songliao Basin represents a critical exploration target for continental shale oil in China. Nevertheless, this shale reservoir is characterized by strong reservoir heterogeneity and complex lithological assemblages. Practical exploration is further constrained by limited core test samples and scarce nuclear magnetic resonance (NMR) logging data in the study area. Conventional reservoir evaluation approaches suffer from limited prediction accuracy and poor regional adaptability. Meanwhile, purely data-driven machine learning models frequently suffer from deteriorated generalization ability when training datasets are insufficient. To resolve these bottlenecks, this study develops a multi-parameter coordinated prediction method integrating geological prior knowledge and machine learning algorithms for quantitative well-log evaluation of key shale reservoir parameters.

An integrated prediction framework was constructed by combining an improved ΔlogR method, a lightweight fully connected neural network (FCNN), and core-calibrated optimized empirical formulas to achieve high-precision coordinated inversion of total organic carbon mass fraction w (TOC), three major mineral components, and total porosity. The improved ΔlogR algorithm adopted segmented stratigraphic baseline calibration and dynamically adjustable weighting coefficient D, which mitigated the limitations of fixed-baseline linear assumptions and adapted to nonlinear logging responses within high-maturity shale intervals. The lightweight FCNN took seven conventional well-logging curves as input variables: acoustic travel time (AC), gamma ray (GR), caliper (CAL), compensated neutron log (CNL), logarithmic deep lateral resistivity (logLLD), spontaneous potential (SP), and compensated density (DEN). It established a nonlinear inversion model for felsic, clay, and carbonate mineral contents. For porosity estimation, an acoustic-density-neutron synergistic calculation formula was optimized using core-measured porosity data. It substantially enhanced the prediction accuracy for ultra-low-porosity shale reservoirs.

Blind-well validation demonstrated that the improved ΔlogR method increased the coefficient of determination R² of w (TOC) prediction from 0.613 to 0.786. The lightweight FCNN mineral prediction model yielded an average R² of 0.867 and maintained favorable generalization ability under small-sample conditions. After formula optimization, the R² of porosity prediction increased from 0.68 to 0.91.

The proposed prediction workflow balances geological rationality and physical interpretability and effectively alleviates the problem of insufficient model generalization caused by limited training samples. It remarkably improves prediction accuracy and stability for multi-parameter inversion in complex shale reservoirs and forms an integrated well-log interpretation system applicable to small-sample unconventional reservoirs. This study can provide reliable technical support for “sweet spot” identification, fracability assessment, and efficient development of shale oil reservoirs in the Songliao Basin and other similar continental basins worldwide.