A highly accurate and interpretable gradient boosting machine learning model for predicting permeability in strongly heterogeneous reservoirs using the Optuna framework

Abstract Permeability, a key parameter for evaluating the capability of fluid flow within reservoir pores, significantly influences the calculation of effective reservoir thickness, reserves assessment, and production evaluation. While logging techniques offer detailed characterization of the physical properties of reservoir rocks, constructing an accurate permeability prediction model remains a technical challenge. Conventional logging methods can predict permeability for homogeneous layers but face significant limitations when dealing with more heterogeneous layers. Therefore, a novel hybrid method combining Optuna and a light gradient boosting machine (LightGBM) has been proposed for predicting permeability in heterogeneous reservoirs. During the prediction process, Spearman's rank correlation coefficient was used to select 11 well logging curves from a strongly heterogeneous sandstone reservoir as training features. Permeability was used as the target variable for training. The effectiveness of the trained model was evaluated using the root mean square error (RMSE), mean absolute error (MAE), and explained variance score (EVS). The RMSE, MAE, and EVS of the train set were 0.0158, 0.0631, and 0.9863, respectively. The RMSE and MAE were a similar, and EVS was 0.8698 of the test set. To further verify the model's reliability, permeability predictions were made using blind well data from the area, and the results were analyzed. The analysis showed that the predicted permeability values closely matched the actual core permeability measurements. Additionally, the Shapley additive explanations method was used to interpret the model, and the resulting explanations were consistent with geological understanding. This indicates that the Optuna-LightGBM model possesses high precision and good applicability in the area.