Porosity and permeability prediction from petrographic point-counting data using machine learning: Applications to Rotliegendes and Buntsandstein reservoirs

Machine learning approaches are widely used in geosciences. However, one widely available dataset in reservoir geology remains underrepresented in published works: petrographic data from classical point-counting analyses. Such data are widely available for reservoir lithology characterization, often in combination with routine core analysis data (porosity and permeability). Since porosity and permeability in siliciclastic rocks are controlled by the detrital and authigenic composition and samples record effects of compaction during diagenesis, these datasets are often linked to assess reservoir quality controls.
Datasets from six wells, covering four regions and two large reservoir lithologies in central Europe, the Permian Rotliegendes and Triassic Buntsandstein, were used to apply machine learning to the petrographic and reservoir quality data to predict porosity and permeability. Predictions are based on point-counting data including detrital and authigenic phases, optical porosity, grain-to-IGV (GTI) and grain-to-grain (GTG) coating coverages, and granulometry. For both regression tasks, a Random Forest and a Support Vector Regression machine learning model were implemented, with performance compared and the best model selected based on coefficient of determination (R$^2$) and error metrics. ... mehrPorosity predictions using a Random Forest algorithm yielded an R$^2$ of 0.92, a mean average error (MAE) of 1.25%, and a root mean square error (RMSE) of 1.56%. Permeability predictions of real-scale permeability using Support Vector Regression gave an R$^2$ of 0.85, MAE of 29.4 mD, RMSE of 68.3 mD, and a range-based normalized RMSE of 8.76% (real-scale). Log-transformation of measured and predicted permeability resulted in a more representative R$^2$ of 0.83, MAE of 0.21, and RMSE of 0.24, reflecting its log-normal distribution. Predictions are acceptable despite the limited dataset, which reduces operator bias by using curated data. This machine learning approach may simultaneously unlock another understanding of reservoir quality controls based on SHapley Additive exPlanations (SHAP) value plots.
Further training of such models on cored reservoir sections can improve understanding of which detrital and authigenic mineral phases influence reservoir properties. Trained models could also potentially evaluate reservoir properties from cuttings, which, like well logs, are more continuous than cores while allowing diagenetic interpretation based on petrographic analysis.