提出者：

编者注：本文出现在新的 E&P 时事通讯中。请在此处订阅时事通讯 。

开发地球物理学的圣杯之一是使用测井数据和地震反演结果对不可预测的储层单元（如致密页岩）进行岩石特性估计。横向钻井是在最有效点区间内进行的，该区间的厚度通常不超过 20 英尺，深度可达 10,000 英尺。但是，在此区间内，地质情况可能会迅速变化，并且岩石类型或储层质量也会发生变化通常低于或正好等于地震分辨率。

为了克服这个问题，对地震叠前数据进行反演以预测弹性特性。岩心信息和电子测井等井数据用于得出各种岩石类型和弹性特性之间的关系。然后，这些关系以各种方式应用于地震属性。如果它们很简单，则可以得出线性关系。如果它们更复杂，则使用机器学习和人工智能（AI）。其结果是一个 3D 岩石属性/岩性立方体，可帮助选择井位和加密井，并提供有效的工具来引导井的横向路径。

TGS 和 Integrated Subsurface Technology 正在使用阿纳达科盆地的陆上数据集演示这种方法。

输入数据

South Gloss 地震 3D 数据位于阿纳达科盆地，面积 232.5 平方英里。如图 1 所示，大量具有声波和密度的井可供分析。 

从历史上看，具有测量的声波和密度测井的井数量不是很多。通常，这些日志类型仅在很短的时间间隔内记录，通常针对感兴趣的区域。在地球物理应用中，p 声波和密度以及次要的 s 声波至关重要。这些测井是井与地震信息之间的主要联系。TGS 通过使用独特的分析方法来估计在较大间隔内丢失或根本没有记录的 p 声波和密度测井来解决这个问题。该 ARLAS 产品可提供非常丰富的测井记录，如图 1 所示。

图 2a、2b 和 2c 显示了阿纳达科盆地所示轨迹上测量的声波测井以及测量和估计的声波测井之间的比较。显示并非按比例绘制，孔的间距相等以便于演示。S-声波测井并不那么容易获得。 

方法

首先，对偏移的叠前道集进行反演调节。这种调节包括氡变换到衰减的层间多次波、线性噪声衰减和残余速度分析。沿着可用的偏移范围，生成了三个角度堆栈。井数据用于校准目的并提取用于反演的小波。主要输出是 P 速度、S 速度、Vp/Vs、LambdaRho、MuRho，公司还尝试估计密度。 

如有必要，ARLAS 测井曲线会进行编辑，并为该过程做出有效贡献。 

地震反射率由速度和密度驱动。广泛使用的 Zoeppritz 方程将反射能量分为三个分量：P 速度、S 速度和密度。传统的 Vp - Vs 交会图通常用于证明这两个关键地震分量区分岩性。通常可以使用 Lame 参数 LambdaRho（不可压缩性）和 MuRho（刚性）来进行更优雅和明确的岩性区分。密度通常是不可靠的输出，因为它必须从通常不可用的远偏移中得出。LambdaRho 和 MuRho 的使用排除了在岩性预测中使用密度的要求。图 3 显示了感兴趣区域的 LambdaRho - MuRho 交会图，并表明该区间内的各种岩石单元在该空间中明显分开。 

图 4 显示了顶部的 PSTM 叠层横截面和底部的倒转 LambdaRho — MuRho 岩性属性。可以观察到属性和已知地质之间的良好相关性。

使用源自传统三重组合和声波测井的“电相”模型可以进行更完整的岩性预测。该模型可以通过岩心与电测井的详细关联来进一步完善。在此示例中，地震勘探中的 12 口井被用作神经网络电相分类的输入。测井包括 GR、中子、密度、P-sonic 和 PE。 

在该项目中，根据测井分析，感兴趣的地质部分被分为七个不同的相。该过程的下一步包括使用多个地震体作为输入，对这些导出良好的电相进行神经网络预测。然后从神经网络输出概率函数，显示特定相预测的质量或置信度。这对于产品价值的诠释尤为重要。

图 5a 显示了更传统的 LambdaRho-MuRho 与同一在线勘测的电相预测之间的比较。请注意垂直和横向的详细描述。图 5b 显示了相应的色键。两种方法都能产生优异的结果。然而，电相产品中有更多细节。 

结论

识别储层层段的岩性/岩石类型是可取的，尤其是陆上非常规储层类型。在此分析中，TGS 和 Integrated Subsurface Technology 展示了如何通过集成地震叠前数据和通过机器学习增强的井数据来实现这一目标的示例。

此外，TGS 和集成地下技术研究了两种不同类型的岩性反演。第一个产品包含 LambdaRho-MuRho 岩性属性。这是一种更传统的方法，使用 LambdaRho MuRho 交会图来划分各种岩性。第二个岩性预测属性是利用地震属性、井信息和机器学习得出的。两个结果都很好，但神经网络衍生的电相属性显示了储层岩石内的更多细节。 

尽管地震数据的垂直和横向分辨率有限，但不应忘记，所有地质都存在于叠前数据中，并且可以通过地震波形检测沉积物的变化性。使用适当的小波分析和严格的神经网络方法进行叠前反演，包括测井和地震尺度，可以提供详细识别地层横向变化的结果，并减少对垂直分辨率的要求，从而更现实地理解垂直检测。 

Presented by:

Editor's note: This article appears in the new E&P newsletter. Subscribe to the newsletter here.

One of the holy grails in development geophysics is the estimation of rock properties using log data and seismic inversion results over unpredictable reservoir units like tight shales. Lateral drilling is performed within a sweet spot interval that’s often not more than 20 ft thick at a depth of up to 10,000 ft. However, within this interval, the geology can vary rapidly, and the variations in the rock type or reservoir quality often are below or just at the seismic resolution.

To overcome this, seismic prestack data are inverted to predict elastic properties. Well data, such as core information and electronic logs, are used to derive a relationship between various rock types and elastic properties. These relationships are then applied to the seismic attributes in various ways. If they are simple, a linear relationship can be derived. If they are more complex, machine learning and artificial intelligence (AI) are utilized. The result is a 3D rock property/lithology cube that assists in choosing well locations and infill wells and provides a valid tool to geosteer the lateral path of the well.

TGS and Integrated Subsurface Technology are demonstrating this method using an onshore dataset from the Anadarko Basin.

Input data

The South Gloss seismic 3D data are located in the Anadarko Basin and 232.5 sq miles. As Figure 1 illustrates, an excellent number of wells with sonic and density are available for analysis. 

Historically, the number of wells that have a measured sonic and density log is not very high. Often, those log types are only recorded over short intervals, usually targeted toward the zone of interest. In geophysical applications, the p-sonic and density, and in a less crucial way, the s-sonic are essential. Those logs are the main link between well and seismic information. TGS solves this problem by using a unique analytical method to estimate p-sonic and density logs that are missing over larger intervals or haven’t been recorded at all. This ARLAS product leads to a very rich well log offering, as shown in Figure 1.

Figures 2a, 2b and 2c display a comparison between measured sonic, and measured and estimated sonic logs over the indicated trajectory in the Anadarko Basin. The display is not at scale, as wells are spaced equally for an easier presentation.  S-sonic logs were not as readily available. 

Method

To begin, the migrated prestack gathers were conditioned for the inversion. This conditioning included a Radon transform to attenuated interbed multiples, linear noise attenuation and a residual velocity analysis. Along the usable offset range, three angle stacks were generated. The well data were used for calibration purposes and to extract wavelets for the inversion. The main outputs are P-velocity, S-velocity, Vp/Vs, LambdaRho, MuRho, and the companies also attempted to estimate density. 

The ARLAS well logs were edited if necessary and made a valid contribution to the process. 

Seismic reflectivity is driven by velocity and density. The Zoeppritz equation widely used partitions reflection energy into three components: P-velocity, S-velocity and density. The traditional Vp – Vs cross plot is commonly used to demonstrate that these two critical seismic components distinguish lithology. A more elegant and definitive lithology distinction can often be made using the Lame parameters LambdaRho (incompressibility) and MuRho (rigidity). Density is often an unreliable output, as it has to be derived from the far offsets that are often not usable. The use of LambdaRho and MuRho precludes the requirement of using density in the lithology prediction. Figure 3 displays the LambdaRho - MuRho cross plot at the zone of interest and demonstrates that the various rock units within this interval are clearly being separated in this space. 

Figure 4 illustrates a PSTM stack cross-section at the top and the inverted LambdaRho – MuRho lithology attribute at the bottom. A good correlation between the attributes and known geology can be observed.

A more complete lithology prediction can be made using an “electrofacies” model derived from a traditional triple combo and sonic logs. The model can be further refined by a detailed correlation of core to electric log. In this example, 12 wells from across the seismic survey were used as input to the neural network electrofacies classification. Logs included GR, Neutron, Density, P-sonic and PE. 

In this project, the geological section of interest was broken out into seven distinct facies from the log analysis. The next step in the process comprised running a neural network prediction of those well-derived electrofacies using multiple seismic volumes as inputs. A probability function was then output from the neural network showing the quality or confidence in the prediction of a particular facies. This is especially important for the interpretation of the value of the product.

Figure 5a displays a comparison between the more traditional LambdaRho-MuRho and the electrofacies prediction on the same survey inline. Notice the detailed description in both the vertical and lateral direction. Figure 5b shows the corresponding color key. Both methods yield excellent results. However, more details are present in the electrofacies product. 

Conclusions

Identifying lithology/rock types at reservoir intervals is desirable, especially onshore in unconventional reservoir types. In this analysis, TGS and Integrated Subsurface Technology showed an example of how they achieved this by integrating seismic prestack data and well data that have been augmented through machine learning.

Furthermore, TGS and Integrated Subsurface Technology investigated two different types of lithology inversions. The first product comprised a LambdaRho-MuRho lithology attribute. This is a more conventional approach that uses a LambdaRho MuRho crossplot to break out various lithologies. The second lithology prediction attribute was derived using seismic attributes, well information and machine learning. Both results are good, but the neural network-derived electrofacies attribute shows more details within the reservoir rocks. 

Although seismic data have a limited vertical and lateral resolution, it should not be forgotten that all geology resides within the prestack data and that it's possible to detect the variability of the deposits by seismic waveforms. Prestack inversion using proper wavelet analysis and a rigorous neural network approach includes both well log and seismic scale and can provide results that identify in great detail the lateral variability of the formation and reduce the demand on vertical resolution to a more realistic understanding of vertical detection.