Data-driven estimation for permeability of simplex pore-throat reservoirs via an improved light gradient boosting machine: A demonstration of sand-mud profile, Ordos Basin, northern China

Abstract

Based on the theory of logging interpretation, permeability of the sand-mud reservoirs routinely can be viewed as a comprehensive response of some reservoir characters including porosity, clay content, median grain diameter, etc., and since each of these characters can be directly revealed by different well log, there then exists a multivariable fitting relationship between permeability and logging sequences. Light gradient boosting machine (LightGBM), a kind of ensemble learning, is demonstrated as one state-of-the-art solver for a variety of regression issue, and therefore can be regarded as a potential candidate to address the fitting of permeability via some logs. However, for this model, as the embedded exclusive feature bundling (EFB) algorithm appears incompatibly to logging data and meanwhile a great deal of hyper-parameters are employed during the training, a raw usage of LightGBM seems to be ineffective for a logging-based permeability regression. Thus, continuous restricted Boltzmann machine (CRBM) and artificial fish swarm algorithm (AFSA) are adopted as assistants to enhance the predicting capability of LightGBM, which will respectively deal with the transition of input logs and the optimization of hyper-parameters. Accordingly, a new hybrid predictor for the permeability fitting is proposed and named as CRBM-AFSA-LightGBM. To further verify the working performance, robustness, and generalization of the new predictor, the dataset collected from the tight sandstone reservoirs of member of Chang 4 + 5, Ordos Basin, Northern China, is applied to implement a validation. Simultaneously, to highlight the validating effect, two sophisticated fitting models, support vector regression (SVR) and extreme gradient boosting (XGBoost), are introduced as competitors. To fairly conduct the designed competitions, the integration of CRBM and AFSA also will be embedded in SVR and XGBoost, and then the real forms of two competitors are CRBM-AFSA-SVR and CRBM-AFSA-XGBoost. Given a comprehensive analysis of the gained experimental results, four points are summarized as the critical conclusion: 1) compared to SVR-cored predictor, the closer permeability results fitted by XGBoost-cored and LightGBM-cored predictors are proved more reliable, but the latter hybrid model only averagely spends 1/23 shorter times on the regression, hence argued as a higher-efficient fitting solver for the permeability of sand-mud reservoirs; 2) training more learning samples is testified effectively to raise an enhancement on the predicting capability of LightGBM-cored predictor, and then the application of a larger-volumetric learning dataset is encouraged during the training; 3) LightGBM-cored predictor will perform worse under the employment of a learning dataset derived from different study zone, thereby presenting an undesirable generalization, while such unsatisfactory predicting circumstance can be improved if more samples are learned during the establishment of input-output mapping; 4) since LightGBM-cored predictor has the capability to figure out the reasonable permeability outcomes even utilizing a sparse learning dataset, a strong robust nature is demonstrated for CRBM-AFSA-LightGBM.