Faults play a crucial role in the exploration and development of oil and gas reservoirs. They can range from large faults that create oil traps to smaller ones that act as baffles or conduits, influencing the flow of hydrocarbons. Regardless of their size, seismic attributes can aid in mapping these faults, whether interpreted on vertical seismic sections or time slices from 3-D seismic volumes. The manual process of interpreting faults on vertical seismic sections or time slices can be time-consuming and results in fault sticks, which are then connected to form fault surfaces. Seismic attributes useful for this purpose include coherence and fault likelihood. While curvature attributes could also be considered, they will be discussed in a separate article.
To accelerate fault interpretation, several image processing algorithms, such as the swarm intelligence-based ant tracking procedures, have been developed. These algorithms provide detailed fault mapping and are faster once the user is proficient with the associated software. In recent years, as the size of seismic data volumes has increased, the application of machine learning and artificial intelligence techniques has grown significantly. These advancements are noteworthy as they reduce the workload of seismic interpreters and offer valuable insights, as demonstrated in the following data example.
We begin with an overview of seismic coherence and fault likelihood attributes, followed by a discussion of the AI fault probability attribute.
Coherence Attribute
The coherence attribute is a widely used tool in seismic interpretation, particularly for imaging geological discontinuities in 3-D seismic data. First introduced in 1995, the original cross-correlation-based coherence algorithm has evolved over time, with various other algorithms having been developed, including semblance, eigenstructure, prediction-error filter, gradient structure tensor and energy ratio-based methods. These algorithms differ in how they process amplitude, phase and waveform variations, influencing their sensitivity to geological features, spectral bandwidth and seismic noise.
The computation of coherence involves constructing an analysis window along the structural dip, followed by solving for eigenvalues and eigenvectors in eigenstructure-based methods. The ratio of the first eigenvalue to the sum of all eigenvalues provides the coherence value. This process is repeated for each sample to generate a coherence volume for interpretation. The energy ratio-based method, on the other hand, computes the ratio of the coherent component’s energy of the analytic trace to the total energy of the analytic trace within the analysis window and applies this across the entire seismic volume.
For more details on the applications of the coherence attribute, refer to our Geophysical Corner article published in the July 2018 issue of the EXPLORER.
Fault Likelihood Attribute
To compute fault likelihood in seismic data, the process begins by creating a reflector dip-oriented coherence attribute volume. Fault likelihood is then determined by scanning for fault dip angles (both negative and positive) and strike. The results often highlight the fault likelihood corresponding to faults within the seismic data. The next step involves retaining only the local maxima in the fault likelihood images, setting all other values to zero. This thinning process allows for the extraction of fault surfaces using quadrilateral meshes, which are then connected to form the fault surfaces.
While this method is effective for identifying single faults, it encounters challenges when dealing with intersecting faults. At fault intersections, determining which reflectors to correlate in order to estimate fault slip becomes problematic, often leading to gaps or holes in the resulting fault surfaces. One potential solution to address these gaps is to filter the likelihood images in both the inline and crossline directions and then combine the two filtered results. This can improve the continuity of the fault surfaces. For more details on the fault likelihood attribute, refer to our May 2021 installment of Geophysical Corner.
AI Fault Probability Attribute
The computation of seismic fault attributes is typically framed as an image segmentation problem, where deep learning techniques are employed. Some of the commonly used image segmentation methods include fully convolutional networks, SegNet, U-Net, and Mask-RCNN. In our December 2021 article, we described the application of U-Net for automated fault interpretation. The U-Net encoder-decoder architecture is again highlighted in the present seismic image segmentation application.
The AI-assisted seismic interpretation workflow primarily leverages the U-Net convolutional neural network, which automates tasks such as fault detection, channel detection, horizon interpretation and salt dome detection. A major challenge in this field is the lack of sufficient training data, especially for subsurface geoscience applications. To address this, two main approaches are usually explored: improving AI algorithms to work effectively with fewer labeled data and using synthetic data for training. However, synthetic seismic data often leads to false positives, as it cannot fully capture the complexities of real-world seismic data.
To overcome these challenges, an interactive training workflow has been developed that customizes U-Net models for specific seismic tasks, accounting for both general object characteristics and the geological specifics of each dataset. This approach enhances model predictions by addressing the inherent limitations of seismic data, such as limited resolution and noise, ultimately ensuring more accurate and realistic results.
The AI-assisted seismic interpretation workflow consists of five key steps:
- Interpret fault lines on interpreter-selected vertical seismic sections.
- Generate (i.e., “train”) a model on the interpreted vertical sections. Then use that model to predict the faults on the same sections. Repeat this for other interpreted sections. If the results are unsatisfactory, adjust the model configuration or training parameters (e.g., iterations, image size, vertical slice orientation) and retrain.
- Validate the trained model by using it to predict faults on several uninterpreted vertical slices. If the validation results are satisfactory, the process concludes. If not, interpret additional vertical slices and retrain. Both validation and training errors should be carefully monitored to prevent overfitting.
- Apply the trained model to the entire 3-D seismic volume, creating an AI-interpreted volume that indicates fault probability at each seismic sample.
- Within the AI interpretation volume, the predicted faults are joined to form fault surfaces, which can then be extracted.
The key benefit of interactive training is that it allows the AI model to integrate the interpreter’s expertise by tailoring it to the specific geological characteristics of a dataset. This is a crucial advantage, as demonstrated in the real data example discussed in the following section. Moreover, the model retains the general characteristics of fault interpretation through the base model, which can be trained using synthetic data, previous interpretations, or both. This approach significantly enhances the model’s ability to deliver more accurate, contextually relevant predictions.
Application to a 3-D Volume from New Zealand
Based on the discussion above, we applied the previously mentioned fault delineation workflows to the Nimitz 3-D seismic volume in order to compare their performance. The seismic survey is situated along the western continental shelf of New Zealand’s North Island, in the northernmost region of the Taranaki Basin. The seismic data has a bin size of 12.5 x 25 meters, a sample interval of 2 milliseconds, and a record length of 6 seconds.
Figure 1a shows vertical slice AA’ through the seismic amplitude volume where we see several faults are visible in the middle, along with a prominent fault on the right, marked with yellow block arrows. There are two faults on the left marked with green and cyan block arrows. While the fault represented by the green arrows looks to be linear and clearly interpretable, the fault marked with the cyan arrows has a curved signature which is not so clear. Figure 1b shows a time slice through the same volume at t= 2.0 seconds (the location of which is indicated on the crossline section in figure 1a). The sharp amplitude breaks in this display highlight the fault locations. The location of the faults marked with green and cyan arrows are shown with the dots of the same individual colors.
Coherence attributes are typically the first to be computed for fault detection, so we generated an energy ratio coherence volume for the input seismic data and used the corendered volume to create displays similar to figure 1. Figure 2a shows vertical line AA’ where low coherence values computed using a vertically oriented 3-trace by 3-trace by 11-sample analysis window aligned with structure identify the faults, as expected. However, rather than the sharp, clear faults seen in figure 1a, we observe broken values along the fault locations. We discussed such stair-step artifacts in the March 2017 Geophysical Corner article. Note that the steeply dipping “coherent” fault plane reflector indicated by the yellow arrow does form part of a coherence fault-plane anomaly. The time slice display in figure 2b shows some faults clearly on the left side, but in the noisier part of vertical slice on the right side the low coherence anomalies are more sensitive to seismic data quality than they are to the underlying geology. The faults indicated by the green and yellow block arrows are not as prominent, though they can be discerned with some effort.
Next, we generated the fault likelihood attribute and replicated the figures produced for the coherence attribute. Figure 3a shows the corresponding seismic section, corendered with the fault likelihood attribute computed using a 3-trace by 3-trace by 21-sample (i.e. non-vertical) analysis window aligned with structure. Many of the fault likelihood lineaments align with the seismic amplitude breaks associated with the faults. However, a key observation was that the prominent fault marked by the cyan block arrows is still not clearly interpreted. Instead, we see broken lineaments along the amplitude break that defines the fault. The time slice in figure 3b shows a clearer definition of some faults, but the prominent fault marked with the cyan arrow remains difficult to interpret.
At this point, we reassessed our objective, expecting the faults to be similar to those shown on the crossline section in figure 4, as an experienced interpreter would expect. By now, it was clear that simply applying any discontinuity attribute might not be sufficient to achieve our goal. We realized that we needed to incorporate some interpretation input into the process to generate more accurate fault predictions from the seismic data volume. This led us to focus on computing the AI fault attribute volume.
We manually interpreted the faults on ten inlines and ten crosslines within the 3-D seismic volume to create a training model. Each time we did this, we also tested the model’s predictions to evaluate its effectiveness. Initially, the model was less effective, as some faults were not being detected while others were. However, we continued refining the model by picking additional faults and testing the predictions on uninterpreted inline and crossline sections. After several iterations, the results improved, and we arrived at the final fault probability volume after thinning.
Figure 5 displays both the corendered crossline section and time slice displays computed using AI and convolutional neural networks using a 64 by 64 2-D analysis window. The predicted faults are now clearly defined and continuous, which was the desired outcome. It may be mentioned specifically, how, among others, the faults marked as 1, 2 and 3 (with their ends indicated with white block arrows in figure 5b) stand out clearly as compared with the displays in figures 1b, 2b and 3b, and their interpretation is likely to impact the bottom-line.
Finally, we generated a chair display from the 3-D corendered sub-volume for the seismic and AI fault probability volumes, as shown in figure 6. The faults at the locations marked by the orange and cyan block arrows are well-defined, and the fault plane corresponding to the cyan block arrow is now clearly interpretable.
Conclusions
Coherence measures discontinuities in a small analysis window and provides localized estimates of structural, stratigraphic and diagenetic edges. Fault likelihood is a modification of coherence tuned to compute more through-going fault anomalies, providing more continuous fault images with reduced stair-step artifacts. CNN fault prediction uses a much larger window, and having been trained by synthetics and human interpretation is able to “jump” across low-amplitude and noisy zones in the data to “connect” faults seen in the upper and lower part of the analysis window. Furthermore, human interpretation allows the incorporation of fault indicators other than reflector discontinuities, including fault-plane reflection events and rapid change in dip across a fault.
We plan to extend this work by connecting the generated faults into fault planes, which can then be exported as a deliverable. The results of this next step will be the focus of our future article.
Acknowledgements
We would like to express our gratitude to New Zealand Petroleum and Minerals for providing the Nimitz 3-D seismic dataset used in this study. The AI fault interpretation workflow and fault likelihood attribute generation discussed in this work were implemented using the seismic interpretation software from subsurfaceAI Inc., under the names InterpAI and subsurfaceAI. The energy ratio coherence was computed using software from the AASPI Consortium at the University of Oklahoma.