基于图论聚类和最小临近算法的岩性识别方法——以四川盆地西部雷口坡组碳酸盐岩储层为例

碳酸盐岩具有非均质性强、岩性变化快和岩石类型复杂的特征，岩性精细识别难度大，严重制约了储层参数的计算及后续油气开发。以四川盆地西部雷口坡组碳酸盐岩储层为例，结合岩心和薄片等分析测试资料将储层发育的岩性分为8类：藻粘结白云岩、粉晶白云岩、泥晶白云岩、灰质白云岩、白云质灰岩、灰岩、膏质白云岩和石膏，明确了不同岩性的测井响应特征。采用机器学习的思想，将已知岩性定名样本作为训练数据，利用图论聚类分析方法建立岩性识别训练模型，在此基础上结合最小临近算法对未取心井岩性进行预测，实现了不同岩性的精细识别。区块应用结果表明：该方法岩性识别整体符合率高达91.3%，有效提高了岩性识别精度。

A lithology recognition method based on multi-resolution graph-based clustering and K-Nearest Neighbor:A case study from the Leikoupo Formation carbonate reservoirs in western Sichuan Basin

Carbonate rocks have the characteristics of strong heterogeneity, changing lithology and various rock types, which make it difficult to recognize their fine lithologic features and seriously restrict the calculation of reservoir parameters as well as subsequent oil/gas development. The carbonate reservoirs in the Leikoupo Formation in western Sichuan Basin were studied to deal with the problem. Core and thin slice observation and other analysis results revealed eight distinctive lithologic facies in the reservoirs: the algal bonded dolomite, crystal powder dolomite, dolomicrite, calcite dolomite, dolomitic limestone, limestone, gypsum dolomite and gypsum. Their log responses were also identified. In addition, machine learning was combined with multi-resolution graph-based clustering to establish a lithology identification training model by using the known and named lithologic samples as training data. Subsequently, the lithology of reservoirs in other wells was predicted with the K-Nearest Neighbor, thus realizing a fine identification of different lithologic facies. Field application of the method shows a 91.3% of overall coincidence rate of lithology recognition, indicating an improved accuracy in lithology identification.