Improving permeability prediction in carbonate reservoirs through gradient boosting hyperparameter tuning

For efficient permeability modeling and prediction in oil and gas reservoirs, the recorded well-log data and the discrete lithofacies are calibrated with the limited core data available and then used for the prediction in unobserved and unsampled well-bore sections. Like other machine learning models, the gradient boosting model (GBM) requires hyperparameter tuning to optimize their prediction performance. Two GBM configurations are evaluated that involve grid search (GS-GBM-HPO) and random search (RS-GBM-HPO) to optimize (tune) the model’s hyperparameters settings and compare their prediction performances with default settings (GBM-DHP). These three GBM configurations are first trained and tested through cross-validation, in classification mode, to predict lithofacies, based on six well-log distributions, from a partially cored well drilled into an extensive carbonate oil and gas reservoir (Mishrif Formation, Iraq). The predicted discrete lithofacies distribution is then incorporated with the six well-log distributions to tune another set of GS-GBM-HPO, RS-GBM-HPO, and GBM-DHP models is trained and tested, in regression mode, to predict Mishrif reservoir permeability throughout the wellbore. Total-correct percentage (TCP) values assess the lithofacies classification accuracy of the GBM models. The highest TCP of 96.6% was achieved by the RS-GBM-HPO model applied to the lithofacies testing subset. Mean absolute error (MAE) and root mean square error (RMSE) assess permeability prediction accuracy. The lowest permeability prediction errors were achieved by the RS-GBM-HPO model, configured in regression mode, and applied to the testing subset (MAE = 6.85 mD; RMSE = 15.34 mD). The trained RS-GBM-HPO model was used to predict Mishrif permeability in a separate non-cored well (Well X) and these predictions closely matched permeability values derived from nuclear magnetic resonance (NMR) logging of Well X. The presented RS-GBM-HPO model can provide dependable permeability predictions in non-cored sections of the well used for GMB model training. Moreover, the trained GMB models can be generalized for application to other field wells lacking core data.