Machine-Learning Approach Identifies Wolfcamp Reservoirs

You have access to this full article to experience the outstanding content available to SPE members and JPT subscribers.

To ensure continued access to JPT's content, please Sign In, JOIN SPE, or Subscribe to JPT

This paper discusses a project with the objective of leveraging prestack and poststack seismic data in order to reconstruct 3D images of thin, discontinuous, oil-filled packstone pay facies of the Upper and Lower Wolfcamp formation. The classification results were created by neural networks, which can be used as a substitute for traditional amplitude-vs.-offset, inversion, and cross-plotting techniques for seismic reservoir characterization.


The problem encountered by the operator in this oil field is that the reservoir, an oil-filled packstone, is thin and laterally discontinuous. Despite having collected a high-resolution, state-of-the-art 3D seismic survey with usable frequencies up to 138 Hz, and despite having generated seismic attribute volumes in order to assist interpretation, the operator was unable to generate an interpretation manually that matched the rock-type interpretation at the wells. Therefore, the decision was taken to supplement the human interpretation with a machine-learning methodology.

Geological Setting

The East Soldier Mount (ESM) study area is approximately 200 km northeast of Midland, Texas, USA, on the eastern shelf of the Permian Basin. The study area contains both Upper and Lower Wolfcamp oil-filled packstones, which are thin and laterally discontinuous. Bioturbation and oolitic shoals caused the initial porosity; however, much of the porosity was occluded by cementation after burial. The porosity was enhanced by fracturing that occurred after burial, caused by differential compaction beneath and tectonic faulting in the deeper formations. Many millions of years after burial, oil leaked into the Tannehill sand (Middle Wolfcamp) detrital, then migrated up the detrital zone into the delta, which is located 2 km west of the study area. Then, the oil migrated out of the delta and into the study area itself. Middle Wolfcamp deltaic sands are not collocated with the seismic survey and thus are not part of this study. They were, however, the conduit for oil migration into the ESM wells.

The total drilled depth for these vertical wells is approximately 1500 m. Each successful well produces approximately 3–4 million bbl of reserves. These wells flow naturally, without the need for hydraulic fracturing.

Theory and Definitions

The technique developed in this paper assumes the existence of a relationship between a seismic response at a given point and the rock-type distribution around this point. However, no model has been established yet and the mathematical formulation of such an operator is complex. Its determination would imply the need for a long empirical process to evaluate the consequence of rock-type distribution on seismic response. Estimating the number of parameters is challenging and is a function of the geological context, measurement constraints, and experimental design. Therefore, the work flow is to create an operator by using learning techniques.

As with any learning method, the presented technique, democratic neural network association (DNNA), needs a representative data set in order to build a robust operator. Unfortunately, rock type is not available as a volume and has to be approximated by the lithofacies distribution defined within vertical windows along the borehole.

Many limitations are identified clearly for discrete prediction using seismic data. Lithological information, interpreted and gathered at wells, is not linearly correlated with seismic data. Facies are not ordered, and there is no notion of mathematical separation between them. A specific neural network is designed to learn in a specific way; therefore, using only one supervised neural network tends to bias the results of the training.

The problem of well-to-seismic data classification renders this one-goal approach unsatisfying because classes often overlap. The use of several networks running simultaneously as an associative combination is preferred.

With regard to the data, different approaches can be considered for simultaneously training several neural networks. Usually, multiple-view learning methods are used. The application of this kind of approach to facies prediction is not optimal in a reservoir-characterization sense because seismic data are interdependent.

The second approach is to simultaneously run different neural networks to be trained with the same hard data set. This single-view colearning approach provides the ability to handle the training of associative neural networks (ASNNs) with a unique set of seismic data attributes that are not necessarily independent, paired with the well information. This is the approach preferred by the authors.

The training steps of the democratic ASNN can be summarized as follows:

  • Define a number p of neural networks.
  • Apply learning over the p neural networks with each training set and examine the training quality by analyzing misclassification rates at well locations.
  • Apply a democratic vote system over a user-defined set of soft data and add the ones that pass the majority vote test as training data, with a lower weight than hard data.
  • Apply learning over p neural networks using the expanded training set now containing hard and soft data.

The final output result of the work flow is a lithofacies distribution for each point in the area of interest, as well as the maximum probability for all facies and the probability for each facies.

Equipment and Processes

  • To carry out the DNNA machine-learning rock-type classification procedure, a number of different data objects had to be delivered by the exploration and production company to be used as inputs.
  • Top and base of the interval of interest, as 3D seismic interpretation horizons, in two-way time
  • Accurate time-depth curves for the three ESM wells
  • Lithofacies logs at the three ESM wells
  • Poststack seismic attributes, including:
    • High-resolution, 138-Hz prestack time-migrated stack
    • Instantaneous frequency
    • Instantaneous Q factor
    • P-impedance from seismic inversion
    • Semblance
    • Dominant frequency
    • Most-negative curvature
  • Prestack seismic data in the form of partial angle stacks (from 0 to 40°, every 5°)

Data and Results

Fig. 1 shows the well and seismic data at the well location (training data set), which are inputs to the DNNA process. From left to right appear a lithofacies log, then another lithofacies log that is a duplicate of the first one (because optional upscaling was not chosen, the upscaled log is identical to the input lithofacies log). Next, there are 15 seismic traces representing the 15 seismic attributes extracted at the well location. Together, the lithofacies logs and the 15 seismic-attribute traces comprise the training data set for DNNA. The collection of lithofacies logs, upscaled lithofacies logs, and 15 seismic traces is duplicated three times, representing the three wells collocated within the ESM 3D seismic survey.

Fig. 1—Training data set (lithofacies plus extracted seismic traces along borehole) for each well.


In the lithofacies reconstruction results at the wells (Fig. 2), there are three well tracks. The first is the input lithofacies, the second is the upscaled lithofacies (identical to the input lithofacies log), and the third is the reconstructed lithofacies log curve. By the term reconstructed, the authors mean that this is the predicted lithofacies curve that results from applying the neural network model at the well location using only seismic data as input. Therefore, it is a best-case scenario for reconstruction of lithofacies using this method. Immediately to the right of the reconstructed lithofacies curve is the maximum probability curve, which shows the probability of the most-probable reconstructed lithofacies. Finally, to the far right are the individual nine facies probabilities for each of the nine lithofacies classes. The facies prediction is extremely good compared with the input lithofacies; this is true for all three wells. There are small differences in lithofacies prediction, and occasionally the DNNA method loses a few thin beds. There are also some minor differences in terms of the positioning and thickness of the lithofacies, but overall, the reconstruction is excellent. The collection of lithofacies logs and probability logs is ­duplicated three times, again representing the three wells collocated within the ESM 3D seismic survey.

Fig. 2—Facies reconstruction results at the wells from the DNNA process.


In examining the overall quality of the lithofacies log reconstruction, each of the three wells has a well reconstruction rate varying between 96 and 98%. The lithofacies class reconstruction rate varies between 75% for Class 8 to 100% for Classes 1 through 6. The global average reconstruction rate is 97% for the entire project.

In the 3D classification results, the vertical resolution of the 3D volume matches that of the lithofacies logs and there is a near-perfect tie at the well locations, which is expected considering the 97% global facies reconstruction rate.  

When pay facies are juxtaposed against lithofacies logs in 3D space, the voxel-­visualization display allows the analyst to visualize where the pay facies may be located. A map view of a two-way time/thickness evaluation of the oil-filled packstone pay facies shows clearly that there are other drilling opportunities in this area.


The ensemble of neural networks was trained at the wells to identify a specific pay facies, using lithofacies logs and 15 3D seismic volumes as input, then propagated to the full volume using the final operator. At the wells, the training statistics are high quality; most importantly, there is no confusion about pay facies. Higher vertical and horizontal resolution was obtained compared with that obtained by conventional high-resolution 138-Hz prestack time-migrated seismic volume. This project suggests that the method has validity in Permian carbonate rocks in west Texas.

For a limited time, the complete paper SPE 193002 is free to SPE members.

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 193002, “Exploring for Wolfcamp Reservoirs, Eastern Shelf of the Permian Basin, Using a Machine-Learning Approach,” by Bruno de Ribet and Peter Wang, Emerson; Monte Meers, independent geologist; Howard Renick, independent geophysicist; Russ Creath, Hardin International; and Ryan McKee, RAM Imaging, prepared for the 2018 SPE Abu Dhabi International Petroleum Exhibition and Conference, Abu Dhabi, 11–14 November. The paper has not been peer reviewed.

Machine-Learning Approach Identifies Wolfcamp Reservoirs

01 March 2019

Volume: 71 | Issue: 3



Don't miss out on the latest technology delivered to your email weekly.  Sign up for the JPT newsletter.  If you are not logged in, you will receive a confirmation email that you will need to click on to confirm you want to receive the newsletter.