With the increasing size of the seismic data volumes, machine learning applications have been found to accelerate the discrimination of seismic facies used in the identification of geologic patterns, defining stratigraphy and the direct indication of hydrocarbons. Many practitioners have demonstrated the application of dimensionality reduction tools such as principal component analysis and independent component analysis (see many of the 2018 installments of Geophysical Corner for examples), or clustering techniques such as kmeans, Gaussian mixture models, self-organizing mapping and generative topographic mapping, among others. Some of these are available in several commercial interpretation software packages.
A comparison of results obtained from such techniques shows their capabilities, with the more sophisticated techniques like GTM exhibiting an edge over others. However, all the above-mentioned techniques depend on the quality of the input seismic attributes and the computation cost involved. Once the decision on the ML technique to be adopted for unsupervised facies classification has been made, the next important aspect to consider is the optimum choice of a subset of these attributes. For example, the choice of input seismic attributes for accurate detection of salt domes will be different from a similar choice for porosity prediction, or facies classification. Two key considerations to make such a choice are to avoid redundant attributes and to find those that best differentiate the target facies from the background geology. An optimal choice results in a geologically consistent interpretation of the facies, which conforms with the reservoir information available from well data.
Figure 1: Two attribute subsets used for comparison of the results obtained from GTM computation forseismic facies classification.
With these considerations in mind, we compare the performance of GTM on two sets of input seismic attributes. Both sets of attributes are shown in figure 1. The first set comprises the P- and S-impedance attributes derived after performing prestack simultaneous impedance inversion on the available seismic data. The other attributes in this set are E-Rho (derived from P- and S-impedance), peak frequency, sweetness and spectral magnitudes at 25, 35 and 45 hertz. In the second set we drop the inverted P- and S- impedance as well as E-Rho and substitute them with relative acoustic impedance and GLCM energy. The peak frequency is also replaced with peak magnitude and the sweetness attribute is replaced with amplitude envelope attribute.
The choice was made for these attributes for the second set in view of the fact that, quite often, the absolute impedance inversion may not form a part of a project. In such cases the relative acoustic impedance could be used as one of the input attributes. We retained the three spectral magnitude attributes at 25, 35 and 45 hertz in attribute set-1, replaced the sweetness attribute with amplitude envelope, the peak frequency with peak magnitude and brought in GLCM energy. This combination is different but has a bit of an overlap.
Brief Description of Geology in Area of Interest
The area of interest is from Golfo San Jorge Basin, located in central Patagonia, Argentina. The formations of interest are the Pozo D-129 Formation, which is the source rock for the basin, and the reservoirs of interest in this work are located in Lower Cretaceous Mina del Carmen and the Upper Cretaceous Comodoro Rivadavia formations. The Mina del Carmen formation consists of pyroclastics, mainly greenish-gray tuffs and shales with scarce interbedded tuffaceous sandstones in the central sections. Regionally, the sands in the Mina del Carmen unit vary in their distribution from being absent to abundant and it is informally subdivided into three subunits of A, B and C.
The Comodoro Rivadavia formation reservoirs consist of lithic feldspathic to lithic quartz sandstones, with medium to large grain size, deposited in river channels with moderate to low sinuosity. The floodplains are composed of clays with increasing tuffaceous content to the base. Reservoir thicknesses range from 1 to 30 meters, with lateral extensions up to 3 kilometers, showing vertical and lateral stacking of multiepisodic channels. The reservoir content is medium to good. In figure 2 we show a segment of a seismic section exhibiting the different formations.
Interestingly, due to the volcanic activity throughout the history of the Golfo San Jorge Basin, the sediments are found to be tuffaceous. The volcanic clastics contain minerals with a wide range of matrix densities. The challenge is to characterize the reservoir sands in the different formation units. Consequently, seismic facies determination formed part of the reservoir characterization exercise which could hopefully prove useful toward cost-effective drilling. In the present case we focus on the seismic facies comparison below the P-marker. It may be pointed out that the name “P-marker” is not associated with P-impedance. It is simply a geological horizon denomination here.
We analyzed a 130-square-kilometer subvolume of a 3-D survey acquired in 1998. The sample increment for the data was 2 milliseconds and bin size 25-by-25 meters. The volume corresponds to an AVO compliant reprocessing including 5-D interpolation followed by prestack depth imaging that was completed in 2019. The final depth migrated volume converted to time was made available for reservoir characterization. The quality of the seismic data was good. Well log data for a few wells was available, including acoustic and lithology interpretation curves and three of them included the complete elastic set which were employed for the prestack simultaneous impedance inversion.
Generative Topographic Mapping
Figure 3: A stratal slice from the GTM crossplot volume generated using attribute set-1 at 20 milliseconds below P-marker when (a) the input attributeswere used as such, and (b) the input attributes were filtered using a gaussian filter. Apart from the east-west running fault (indicated with gray arrows)seen more prominently, the colored patches representing different seismic facies appear to be smoother in (b). The location of the seismic profile shownin figure 2 is indicated by the vertical dashed line in white. Data courtesy of Pan American Energy S.L., Argentina.
Among the different methods of unsupervised machine learning facies classification mentioned above, the PCA/ICA as well as the SOM/GTM applications were discussed in the January 2022 installment of Geophysical Corner. The PCA/ICA methods reduce the dimensionality of the input seismic attributes, whereas the latter two project attributes used from a higher dimensional space (i.e., 7-D when seven attributes are used as input) to a 2-D deformed surface called a “manifold” that best represents the data. Such data can be corendered with RGB or crossplotted using a 2-D color bar. We show the results of GTM machine learning application in this article for comparison of the two input sets of seismic attributes.
Even though all seismic attributes were computed on preconditioned input seismic data, parts of the GTM displays appear as noisy (as seen within the white and brown polygons in figure 3a and red and green polygons in figure 4a).
Figure 3 shows a stratal slice 20 milliseconds below the P-marker level extracted from the GTM-1 and GTM-2 crossplot volume that resulted from input attribute set 1. The display in figure 3a depicts a noisier look. Such an undesirable effect can be addressed by running a smoothing filter on the input attributes such as a structure-oriented filter, or a Gaussian filter, both commonly available in commercial interpretation software packages. In this case, a 3-by-3-by-3 samples Gaussian filter was run on each of the input seismic attributes for GTM computation, and an equivalent display to figure 3a is shown in figure 3b. Notice the smoother look of the display in figure 3b. The gray arrows exhibit better definition of the faults.
Figure 4: A stratal slice from the GTM crossplot volume generated using attribute set-2 at 20 milliseconds below the P-marker when (a) the input attributeswere used as such, and (b) the input attributes were filtered using a gaussian filter. Apart from the east-west running fault (indicated with gray arrows)seen more prominently, the colored patches representing different seismic facies not only appear to be smoother, but more distinct and resolved in (b).Data courtesy of Pan American Energy S.L., Argentina.
An equivalent display to figure 3 is shown in figure 4, extracted from the GTM crossplot volume, where the GTM computation was carried out using attribute set 2. Comparing the images in figure 4a and 4b, we notice that while the overall seismic facies pattern seems similar, the areas highlighted by the different colored ellipses indicate higher resolution. A comparison of the images in figures 3b and 4b indicates that the latter image is more meaningful as it would allow a more accurate interpretation. Each of the main colors seen on the display in figure 4b represents seismic facies and can be quantified with a geologic nomenclature after corroboration with lithology information derived from the well data.
In view of the lower frequency content of the seismic data, we considered spectral balancing the seismic data at hand before the generation of input attribute sets. The method of choice was the peak frequency method described in the May 2014 installment of Geophysical Corner. In this method, the seismic data are first preconditioned using structure-oriented filtering which would suppress the background noise in the signal, and then the data are decomposed into time-frequency spectral components. Next, a smoothed average spectrum is computed. If the survey has sufficient geologic variability within the smoothing window (that is, no perfect “railroad track reflections”), this spectrum will represent the time-varying source wavelet. This single average spectrum is used to design a single time-varying spectral scaling factor that is applied to each and every trace. Geologic tuning features and amplitudes are thus preserved.
The seismic attribute set-2 was computed on the frequency-balanced data and taken forward to GTM computation as before. An equivalent stratal slice generated from the GTM crossplot volume 20 milliseconds below the P-marker is shown in figure 5.
Figure 5: Stratal slice at 20 millisecondsbelow the P-marker extracted from the GTMcrossplot volume generated using attributeset-2 after frequency enhancement. Thedashed outlines in black overlaid on thedisplay indicate the presence of stackedchannel system bands oriented in thenorth-northwest-south-southwest direction.These patterns correlate reasonably well withgeologic information available from well dataanalysis in this area. The color distributionseen on the display, exhibits the seismic faciesclassification, which could be correlated withthe lithology information available at thelocation of the available wells. Data courtesy ofPan American Energy S.L., Argentina.
The dashed outlines in black overlaid on the display indicate the presence of stacked channel system bands, oriented north-northeast to south-southwest and represent the main reservoirs linked to the fluvial deposition in low-relief plains. These patterns also correlate reasonably well with geologic information available from well data analysis in this area. Besides, the color variations in the magenta, red, yellow and bluish colors suggest that the facies distribution is compatible with deposits from fluvial systems. The interpretation of these channels was described in detail in the November 2018 installment of Geophysical Corner.
Likewise at other depth/time levels the other colors corresponding to different lithology/facies can be determined and the overall facies/lithology variation in the interval of interest can be ascertained.
Conclusions
The choice of attributes used as input for the computation of unsupervised facies classification using any of the machine learning techniques can have a bearing on the results. We have shown the comparison of two sets of attributes chosen for a seismic dataset from the Golfo San Jorge Basin. We find that the use of a Gaussian filter on the input attributes has a smoothing effect on the facies results. Similar analysis carried out on the spectrally-balanced seismic data showed a better overall interpretation of seismofacies at the level of interest.
One might wonder how or why the “texture” of the channels gets differentiated from the background reservoir facies using the GTM technique. A plausible explanation would be that it is possible to see the different bits and pieces of the highlighted channels on the different input attribute displays, which we believe are all integrated by the GTM technique.