Intelligent prediction of hydrocarbon situation in shale natural fractures driven by multi-source data and machine learning

Hydrocarbon situation in fractures (HSFs) is a key parameter for shale oil and gas resource evaluation. However, conventional experimental measurements are often limited by discontinuous sampling, high cost and low efficiency. Here, we propose a continuous quantitative prediction method for HSFs in shale reservoirs that integrates multi-source geological and geophysical data with machine learning algorithms. Using experimental HSFs measurements from Jurassic shales in the Sichuan Basin together with multi-source geophysical responses, Pearson correlation analysis reveals significant associations among HSFs and porosity, total organic carbon (TOC), acoustic transit time and density. The results indicate that not all fracture-rich intervals exhibit hydrocarbon enrichment or enhanced productivity; notably high HSFs occur only in shale intervals with both a high TOC and porosity. On this basis, we developed an AI-based predictive framework that couples geological constraints with multi-source data fusion. The optimized eXtreme Gradient Boosting (XGBoost) model achieves excellent predictive performance, with a residual error (RE) of 0.221, root mean square error (RMSE) of 0.311, mean absolute error (MAE) of 0.156 and a coefficient of determination (R2) of 0.959. This approach enables rapid, high-precision, well-scale continuous evaluation of HSFs in shales with prediction accuracy exceeding 95%, providing a powerful tool for shale sweet-spot prediction and resource evaluation, and holds significant potential for intelligent predicting of geological parameters and digital twin model establishment across whole petroleum systems worldwide.