Petrophysical Regression regarding Porosity, Permeability, and Water Saturation Driven by Logging-Based Ensemble and Transfer Learnings: A Case Study of Sandy-Mud Reservoirs

Abstract

From a general review, most petrophysical models applied for the conventional logging interpretation imply that porosity, permeability, or water saturation mathematically have a linear or nonlinear relationship with well logs, and then arguing the prediction of these three parameters actually is accessible under a regression of logging sequences. Based on this knowledge, ensemble learning technique, partially developed for fitting problems, can be regarded as a solution. Light gradient boosting machine (LightGBM) is proved as one representative of the state-of-the-art ensemble learning, thus adopted as a potential solver to predict three target reservoir characters. To guarantee the predicting quality of LightGBM, continuous restricted Boltzmann machine (CRBM) and Bayesian optimization (Bayes) are introduced as assistants to enhance the significance of input logs and the setting of employed hyperparameters. Thereby, a new hybrid predictor, named CRBM-Bayes-LightGBM, is proposed for the prediction task. To validate the working performance of the proposed predictor, the basic data derived from the member of Chang 8, Jiyuan Oilfield, Ordos Basin, Northern China, is collected to launch the corresponding experiments. Additionally, to highlight the validating effect, three sophisticated predictors, including k-nearest neighbors (KNN), support vector regression (SVR), and random forest (RF), are introduced as competitors to implement a contrast. Since ensemble learning models universally will cause an underfitting issue when dealing with a small-volumetric dataset, transfer learning in this circumstance will be employed as an aided technique for the core predictor to achieve a satisfactory prediction. Then, three experiments are purposefully designed for four validated predictors, and given a comprehensive analysis of the gained experimented results, two critical points are concluded: (1) compared to three competitors, LightGBM-cored predictor has capability to produce more reliable predicted results, and the reliability can be further improved under a usage of more learning samples; (2) transfer learning is really functional in completing a satisfactory prediction for a small-volumetric dataset and furthermore has access to perform better when serving for the proposed predictor. Consequently, CRBM-Bayes-LightGBM combined with transfer learning is solidly demonstrated by a stronger capability and an expected robustness on the prediction of porosity, permeability, and water saturation, which then clarify that the proposed predictor can be viewed as a preferential selection when geologists, geophysicists, or petrophysicists need to finalize a characterization of sandy-mud reservoirs.