Ensuring accurate characterization of petroleum-bearing reservoirs is fundamental for successful exploration of hydrocarbon resources. Petrophysical models have been employed as solvers for quantitative evaluations of porosity, permeability, water saturation, and brittleness index. However, due to the high costs of well drilling, the core data obtained from limited wells are generally insufficient. Thus, most petrophysical models are generally impractical because of their unreliable employing coefficients determined statistically from insufficient core data. From the mathematical perspective, the popular petrophysical models reveal that porosity, permeability, water saturation, and brittleness index exhibit nonlinear fitting relationships with well logs. Thereby, the specific characteristics can be determined using a logging-based fitting. Light gradient boosting machine (LightGBM), a state-of-the-art regression technique, exhibits good performance. However, its performance depends on the support of favorable input and parametric optimization. To address this, continuous restricted Boltzmann machine (CRBM) and Bayesian optimization (Bayes) were integrated in this study, resulting in the proposed CRBM–Bayes–LightGBM framework. CRBM enhanced the input value by extracting significant features, while Bayes carried out the optimal initialization of hyperparameters of LightGBM. To comprehensively evaluate the proposed predictor, four experiments utilizing a dataset obtained from the tight sandstone reservoirs of the Ordos Basin were conducted. For further validation, five well-known regression models, including three-layer neural network, k-nearest neighbors, support vector regression, random forest, and extreme gradient boosting, were employed as competing models. An in-depth and comprehensive analysis of the experimental results yielded five key points: (1) The integration of CRBM and Bayes significantly enhanced the prediction performance of LightGBM; importantly, with the addition of geological regularization to mitigate the negative impact of mudstone during modeling, the fitting precision was further enhanced. (2) Compared to the five competing models, LightGBM-cored predictor yielded smaller fitting errors, establishing it as the best choice for petrophysical characterization. (3) LightGBM-cored predictor exhibited better generalization when trained using a larger dataset. (4) The application of transfer learning addressed the issue of under fitting in petrophysical characterization. (5) The proposed predictor demonstrated more robustness, as missing values in the training dataset do not hinder its feasibility.

确保含油储层的准确表征是油气资源成功勘探的基础。岩石物理模型已被用作定量评估孔隙度、渗透率、含水饱和度和脆性指数的求解器。但由于钻井成本较高，有限井获取的岩心数据普遍不足。因此，大多数岩石物理模型通常是不切实际的，因为它们使用根据不充分的岩心数据统计确定的系数不可靠。从数学角度来看，流行的岩石物理模型揭示了孔隙度、渗透率、含水饱和度和脆性指数与测井曲线呈现非线性拟合关系。因此，可以使用基于日志记录的拟合来确定具体特征。光梯度增强机（LightGBM）是一种最先进的回归技术，表现出良好的性能。然而，其性能取决于有利输入和参数优化的支持。为了解决这个问题，本研究集成了连续受限玻尔兹曼机（CRBM）和贝叶斯优化（Bayes），从而提出了 CRBM-Bayes-LightGBM 框架。 CRBM通过提取显着特征来增强输入值，而Bayes则对LightGBM的超参数进行优化初始化。为了全面评估所提出的预测器，利用从鄂尔多斯盆地致密砂岩储层获得的数据集进行了四次实验。为了进一步验证，采用五种著名的回归模型作为竞争模型，包括三层神经网络、k-近邻、支持向量回归、随机森林和极限梯度提升。对实验结果进行深入、全面的分析，得出五个关键点：（1）CRBM和贝叶斯的集成显着增强了LightGBM的预测性能；重要的是，通过在建模过程中加入地质正则化来减轻泥岩的负面影响，拟合精度进一步提高。 (2) 与五个竞争模型相比，LightGBM 核心预测器产生的拟合误差更小，使其成为岩石物理表征的最佳选择。 (3) 以 LightGBM 为核心的预测器在使用更大的数据集进行训练时表现出更好的泛化能力。 (4)迁移学习的应用解决了岩石物理表征中拟合不足的问题。 (5) 所提出的预测器表现出更强的鲁棒性，因为训练数据集中的缺失值并不妨碍其可行性。