Diagenetic effects in carbonate rocks can enhance or occlude depositional pore space. Reliable identification of porosity-enhancing diagenetic features (e.g., vugs and fractures) is essential for petrophysical characterization of reservoir properties (e.g., porosity and permeability), construction of geological and reservoir models, reserve estimation, and production forecasting. Challenges remain in characterizing these diagenetic features from well logs as they are often mixed with variations in mineral and fluid concentrations. Herein, we explore a data-driven approach that is based on a comprehensive well log data set from the Arbuckle Formation in Kansas to classify vuggy facies in carbonate rocks. The available well log data include conventional logs (gamma ray (GRTC), resistivity (RT), neutron/density porosity (NPHI/RHOB), photoelectric factor (PE), and acoustic slowness) and nuclear magnetic resonance (NMR) transverse relaxation time (T₂) logs. We parameterized the measured T₂ distribution using a multimodal lognormal Gaussian density function and combined the resulting Gaussian parameters with conventional logs as inputs into three supervised machine learning (ML) algorithms; namely, support vector machine (SVM), random forest (RF), and artificial neural network (ANN). The facies labeling data used in this study were based on visual examination of vug sizes from core samples, which include five classes; namely, nonvuggy, pinpoint-size, centimeter-size, fist-size, and super-vuggy. In total, 80% of the data set was used as the training set, and a fivefold cross validation was used for hyperparameter tuning. We conducted a detailed comparison of the above three ML algorithms on the basis of different combinations of features. The highest classification accuracy achieved on the holdout testing set is 84% using SVM on a combination of conventional logs and selected NMR Gaussian parameters as inputs. In general, inclusion of conventional log data improves the prediction accuracy compared with using NMR data alone. Feature selection improves the performance for SVM and ANN but is not recommended for RF.

碳酸盐岩中的成岩作用可以增强或封闭沉积孔隙空间。可靠地识别提高孔隙度的成岩特征（例如，孔洞和裂缝）对于油藏性质（例如，孔隙度和渗透率）的岩石物理表征，地质和储层模型的构造，储量估算和产量预测至关重要。在从测井曲线表征这些成岩特征方面仍然存在挑战，因为它们经常与矿物质和流体浓度的变化混合在一起。本文中，我们探索了一种数据驱动的方法，该方法基于堪萨斯州阿巴克尔组的全面测井数据集来对碳酸盐岩中的孔隙相进行分类。可用的测井数据包括常规测井（伽马射线（GRTC），电阻率（RT），中子/密度孔隙率（NPHI / RHOB），T ₂）日志。我们参数化了测得的T ₂使用多峰对数正态高斯密度函数进行分布，并将得到的高斯参数与常规对数组合为三种监督机器学习（ML）算法的输入；即支持向量机（SVM），随机森林（RF）和人工神经网络（ANN）。本研究中使用的相标记数据是基于对核心样品中的孔洞大小进行目视检查的，其中包括五类。即无洞，精确大小，厘米大小，拳头大小和超洞。总计，将80％的数据集用作训练集，并使用五重交叉验证进行超参数调整。我们根据特征的不同组合对以上三种机器学习算法进行了详细的比较。使用SVM结合常规测井和选择的NMR高斯参数作为输入，在保留测试集上实现的最高分类精度为84％。通常，与单独使用NMR数据相比，包含常规测井数据可提高预测准确性。功能选择提高了SVM和ANN的性能，但不建议用于RF。