In recent years, machine learning and specifically deep learning has captured the attention of all industries, and Oil & Gas is no exception. Obviously, the end goal of petroleum companies is to extract oil; yet, one must first find prominent locations inside the Earth to do so.
That is where seismic exploration comes in: it is the study of subterranean formations to help us locate structures underneath. There are a lot of stages of seismic exploration, from processing the raw acoustic signal traces into a regular fixed-grid volume to detecting subterranean facies inside of it. Neural networks could immensely reduce the time taken by these procedures while simultaneously improving the accuracy of results.
Today I will tell you about one specific part of the field development workflow: horizon detection, and use it as an example of integrating ML into a well-established scheme of seismic exploration. Overall, in a two-part article, we will discuss:
- manual approach to detect horizons
- metrics to track the progress of the models
- data generation in seismic exploration tasks
- neural network architectures
- techniques to enhance detected surfaces
- infrastructure for model deployment
Note that in this publication (and in the next part) we are just briefly introducing all of these concepts, overviewing the entire pipeline of our solution. Each of them deserves its own article for an in-depth dive, which are currently underway.
Let’s dig in!
Manual horizon detection
A seismic horizon is a boundary between two layers of rock; such changes are visible in seismic images even for an untrained eye. A set of horizon surfaces is used to either create a structural model of the field to outline potential oil reservoirs. As horizons are the basic blocks of a 3D model and contain essential information about the area, seismic experts track a lot of them in every region during the interpretation stage.
For the past 30 years, surface tracking in Oil & Gas has hardly changed. In most companies, one or more geological experts manually label every point of the target horizon, with some aid from auto-correlation software. The latter allows extending the label from a point to its neighboring traces; yet, it works only where the reflection is distinctly visible, the level of noise is low and the geology itself is trivial. In complex seismic cube areas (often the most interesting ones!), specialists still resort to hand-labeling, spending up to half a year on each field.
Another drawback of such automated methods is that they analyze information in a tiny window: usually, no more than 30 units. That does not fit the bill of tracking seismic properties very well: all of the lower layers affect the shape and properties of the upper ones, requiring a thorough analysis of the entire seismic volume.
Quality assessment
Before prototyping any model, we need to define metrics of our success and utilize them to track progress. In the task of horizon detection, one must make sure that tracked surface:
- corresponds and correlates with actual seismic data
- covers a large area
- (optional) is close to a hand-tracked reference horizon. This metric is relevant only when we have the hand-labeled reference and can compare our result against it.
I can’t stress enough the importance of having quantitative metrics before implementing any (machine learning or not) models.
Moreover, it is desirable to have some thresholds for each metric: for example, the coverage of a horizon must be more than 90% of the field area. Without metrics, it is virtually impossible to be sure that the performance of the proposed model is good, not to mention tracking the progress and impact of our code and architecture improvements.
Unfortunately, current geology workflows don’t care much for quality evaluation: most of the results are passed to the further stages of exploration as they are. At best, labeled surfaces are briefly looked on by other specialists, and this measure is highly subjective and unreliable. Moreover, it requires an enormous amount of time and involvement from the side of geology experts.
The Oil & Gas industry is in dire need of quality control. Even one of the fundamental tasks, such as horizon detection, is not constrained enough, making all the results highly subjective and non-reproducible. Every mistake in the early stages of exploration can cost millions of dollars later, and recent history with the ever-growing complexity of developed projects knows too much of such blunders.
The absence of quantitative metrics is even worse for machine learning: this renders all of the model results useless. Yes, some parts of the detected horizon may be looking (very) good, but we can never be sure that the quality of tracking is consistent on the entire field. And even a human expert can’t save this situation: to make sure that the model-produced tracking is adequate, he would need to manually inspect each of the horizon slices very carefully, essentially spending the same time as to label it by hand.
We overcome this problem by creating a set of easy-to-compute metrics and don’t require human interaction whatsoever. They assess quality from different points of view:
- one metric, called local, is used to assert that nearby traces are tracked in the same fashion. It is computed by comparing the correlation coefficient in a lateral window of each of the traces
- the other metric reflects global patterns. It makes sure that the horizon in different parts of the field is tracking the same reflection along the same phase. We calculate it by comparing each trace to a set of pre-defined reference traces
- the last one combines both local and global properties and serves as a quick quality assessment.
Each of the proposed metrics is trace-wise: that means that the output is an entire map of quality, which makes it easy to find places where the tracking is unsatisfactory. Importantly, by taking a (weighted) average of it, one can easily rank the performance of our models and select the best one! 🎉
Our metrics are, essentially, a digitized expert knowledge, speeding up the process of evaluation of any tracked surface immensely. We are preparing to publish a separate article on the matter, as metrics are invaluable and allow for express quality control. Stay tuned to learn more about the exact mechanics, formulae, and implementation details!
Summary
In this article, we introduced the task of horizon detection, which is a crucial part of the seismic exploration pipeline. To this day, it is solved manually, spending hundreds of hours of highly qualified geologists. Despite seeming simplicity, the ever-growing complexity of developed fields calls for more sophisticated methods, which in turn demand easy-to-compute yet strict quality control.
Our advanced metrics provide such quality control, allowing us to evaluate our to-be proposed models' success. At the same time, our quality maps are clear and explicable, giving invaluable insight into the seismic data and labeled horizons on their own: with their help, an experienced geologist can locate faults, data defects, and other anomalies in a matter of seconds.
To be continued
Despite the evident success of our metrics, there is yet room to improve them; for example, we would love to distinguish different kinds of anomalies on the created maps. Another priority is to assess the quality of seismic data itself, locating promising areas or, on the contrary, noisy and unclear places. Our team will investigate these ideas in the near future.
In the second part, we’ll dive into even more details, namely, data generation and neural networks configuration, and look at the entire workflow of our proposed method.
