November 2019 Editor's Choice

‘Cybersteering’ Technology Automates Geosteering, Advanced Computing Tools

By Tristan Arbus and Steven Wilson

OKLAHOMA CITY–A new method of automated geosteering has been developed that may lighten geosteering workloads while increasing productivity and the accuracy of geosteered wellbores. Dubbed “cybersteering” and developed by Devon Energy Corp., the patent-pending method utilizes a spatial graph database and assesses the quality of matching gamma data for a large catalog of potential segments of constant bed dip (strat blocks).

With this technology, automated geosteering can construct a wellbore that, in many cases, closely mimics that of manual steering by an expert geosteerer. Conventional geosteering uses interpretations of downhole logging-while-drilling or measurement-while-drilling data to control the wellbore and stay within the pay zone. Software systems display data graphically, allowing geosteerers to stretch or squeeze gamma ray log sections to match a total vertical depth (TVD) type log from a nearby well to correlate a well’s stratigraphic depth.

The general process requires the geosteerer to review gamma and trajectory data every time a survey comes in and visually determine the best overall gamma match between the wellbore and type log by manipulating strat blocks. The gamma match is changed and determined by manually varying strat block length and angles.

With increased horizontal drilling activity in onshore resource plays, geosteering has taken a variety of forms, including dedicated geosteering centers with around-the-clock operations, onsite geosteering, remote geosteering and other operational models. With the exception of onsite geosteering, these models often require the geosteering expert to oversee multiple wells of varying status, complexity and difficulty at once. In all cases, a method of automated geosteering would be helpful to confirm and validate existing steers based on mathematical methods and also to potentially steer entire wells autonomously.

Automated geosteering can increase a geosteerer’s output and accuracy, improving costs associated with the geosteering process, while simultaneously improving wellbore placement for improved production. With gamma matching performed autonomously, geosteerers would also have more time to utilize additional information sources and experiment with data such as seismic-while-drilling or azimuthal-wave resistivity to increase their ability to proactively steer difficult wells.

Cybersteering requires nothing more than standard well log data consisting of well trajectory and gamma information, along with a gamma TVD type log. It utilizes a unique method of scoring strat blocks and storing them in a graph database structure along with a distributed “shortest-path” algorithm to determine the sequence of strat blocks that provides the best overall gamma match for the entire wellbore. These calculations are done in the cloud utilizing a computing cluster to speed processing to the point where cybersteering can be utilized on both historical wells and currently drilling wells.

Relational Graph Database

The key to cybersteering is the unique adaptation of a graph database, which is a relational NoSQL database consisting of nodes and edges. Traditionally, nodes store entities such as a person, place or object, and the edges connect these entities and describe their relationships. Common applications include fraud detection in finance and social media, product recommendation engines, identity and access management, etc.

Cybersteering takes the concept of a relational database and combines it with spatial data by arranging nodes in a grid system, with each node storing a specific measured depth (MD) and relative stratigraphic depth (RSD) combination. Therefore, each node represents a potential position within the formation at each point along the wellbore (on a prespecified interval), which is exactly the information that geosteering attempts to determine.

The graph database should be tall enough to encompass the entire target zone, along with a buffer on either side to remain within the target formation. It should be wide enough to encompass the entire length of the lateral. The intervals between nodes are adjustable, but generally speaking, a good starting point is a one-foot or half-foot interval in RSD and a 30-foot interval in MD.

FIGURE 1

Subset of Typical Node Structure for Cybersteering Graph Database

Figure 1 shows a small subset example of the graph database structure for a lateral. Of note is the placement of a node at the landing point (LP), which is a position of known MD and RSD at the beginning of the lateral. These values are basic inputs. Their accuracy does not need to be perfect since cybersteering can adjust automatically to small deviations. Also of note is the fact that a minimum strat block length of 90 feet is enforced, which is why the first full column in the graph database is 90 feet away from the LP. These columns continue to the well’s total depth.

With the graph database structure determined, edges can start to be created. If two nodes (in separate columns) are connected, it defines a section of wellbore that travels a certain distance in MD, starting at one RSD and ending at another (or potentially even the same RSD). This information is enough to define a strat block. The length of the strat block is defined by the delta in MD, and the angle of the strat block is defined by the delta RSD.

Strat Blocks And Edges

To describe this further, consider that the wellbore trajectory is fixed and defined by directional surveys. Each point along a wellbore has both an MD (from directional surveys) and an RSD (determined by the geosteering process). An edge defines these two values for two points along the wellbore from the two nodes it connects, and therefore, defines a strat block.

FIGURE 2

Example of Two Strat Blocks Defined by Graph Database Edges

An example of this is shown in Figure 2. With the well path in red fixed, and each node representing a particular MD and RSD, the strat blocks are prescribed by the nodes through which the end results pass. This example starts at LP with a wellbore 10 feet below the top of the formation, as per the LP node. It travels through a distance in MD (from left to right) and travels to node N (1, 2), which is 5 feet below the top of the formation. These points can then be plotted and connected. Similarly, it then travels to node N (3, 3), which exists 15 feet from the top of the formation, and the process continues.

Taking this idea back to the original concept of geosteering, each strat block defines a particular way in which the wellbore gamma log lines up on the type well gamma log. Historically, the quality of this match has been done visually, but a mathematical formula for strat block quality can be determined. It can be as simple as a modified mean squared error, or a more complicated formula that takes things like overall strat block length into consideration (a very slight preference for longer strat blocks is ideal since it more closely emulates what would be expected from actual stratigraphy and more closely mimics the choices a geosteerer would make).

By creating a graph database large enough with sufficient edges (representing enough potential strat blocks), the strat blocks can be scored to determine the combination of strat blocks from start to finish of the wellbore that has the best overall quality of gamma match. This, of course, presents challenges in computation because a graph database with reasonable granularity can produce upward of 1 million potential strat blocks. That is where distributed cloud computing can help.

Technical Solution

The technical solution runs in Azure Databricks™ on Apache Spark™, and primarily uses the Python programming language (pySpark). The only exceptions are the distributed graph algorithms that use the Scala programming language upon which Spark GraphX™ is built.

The general solution is based on a forward-directed, acyclic, weighted graph with no edges connecting vertices at the same measured depth. The graph vertices are spaced uniformly into rows and columns according to user-defined intervals for the MD and RSD. Single vertices are added to the beginning and end of the graph, allowing algorithmic traversal from an initial point to a final point. Each vertex is numbered and identifies the MD and RSD interval, respectively. Forward-only edges are subscripted by the source and destination vertex. With this generalized model, the program logically follows the sequence of:

  • Building a graph;
  • Calculating edge costs (weights);
  • Finding the shortest path between the start and end vertices; and
  • Visualizing the results.

Graph construction is partitioned across Spark nodes and scaled linearly with cluster size. This enables fast graph creation for very large graphs. A 10,000-foot lateral with hundreds of thousands of vertices and hundreds of millions of edges can be generated in only a few minutes.

Edge cost evaluation also is partitioned, but individual edges require iteration across measured depths within a block. This results in linear time execution for each partition and requires optimized partition sizing to find solutions in a reasonable time.

The most technically complex part of the solution occurs when traversing the graph to find the shortest path. Shortest-path algorithms proceed sequentially by nature and depend on previous path information to solve. This is difficult to accomplish in a distributed environment when a graph is partitioned. To overcome this limitation, cybersteering uses a unique distributed shortest-path algorithm.

Promising Results

FIGURE 3

Manual Geosteering versus Cybersteering Results in Same Well

FIGURE 4

Example of Impact of Poor Data Quality on Cybersteering

The initial results of the cybersteering approach are very promising. Devon Energy performed a study in which 60 wells across multiple formations were steered with the new approach. Previously, these wells had been manually geosteered and the results were compared. In regions with well defined, clean gamma traces, cybersteering performed extremely well, producing steers very similar to its human counterparts (Figure 3). Both steers leave the target formation and re-enter at similar locations after landing the curve. Both steers also tend to follow the formation slightly up dip along the lower quarter of the target zone.

Data quality and parameter choices have an impact on cybersteering (of course, human geosteerers also can be sensitive to data quality). In areas with muddy or erratic gamma traces, it is important to more aggressively smooth incoming gamma traces. Normalization should be as accurate as possible, especially in areas where the type log shows little to no variation in gamma along the vertical depth, making it difficult to confidently place wellbore position. Figure 4 shows the potential impact of poor-quality inputs.

While cybersteering doesn’t come without faults, the areas in which it falls short seem to have clear paths for improvement. Overall trends could be easier to detect with more aggressive smoothing or other trending methods. It also is currently unable to deal with faults, but vertical edges with an associated cost could be constructed within the graph database. These ideas, along with incremental improvements in other areas, could guide cybersteering to usability in an even wider variety of formations and drilling operations.

Cybersteering comes with numerous advantages. The geosteering process as the industry knows it today is widely variable. The same person geosteering the same well can produce different results of varying quality based on a range of factors. Cybersteering can provide accurate, repeatable results. As with any automated process, building trust in the automated system takes time, so gradual implementation and adoption is recommended. Of equal importance is end-user understanding of how the system works. Understanding why cybersteering makes the decisions that it does along with its potential flaws is incredibly important to interpret results.

Cybersteering has major potential in its current iteration. Further improvements to the technology only will expand its usability as a primary geosteering methodology in a wider variety of geologies. Using a graph database to represent spatial relationships in drilling operations also has proven to be a valuable tool and has potential application outside of geosteering.

TRISTAN ARBUS is a data scientist at Devon Energy Corp. He designs and builds production-ready machine learning and statistical models for drilling, supply chain, human resources, communications and other applications. Before joining Devon in 2014 as a drilling and project engineer, Arbus served as a drilling engineer at Consol Energy and as a mechanical engineer at BUG-O Systems International. He holds B.S. degrees in physics and in mechanical engineering from Johns Hopkins University.

STEVEN WILSON is an advanced analytics and data sciences engineer at Devon Energy. He designed and developed the cybersteering automated geosteering software. Before joining Devon in 2015, Wilson served as information technology manager at Chesapeake Energy and as a software developer and air quality and water treatment engineer at OGE Energy Corp. He holds a B.S. in chemistry from the University of Oklahoma and an M.S. in chemical engineering from Oklahoma State University.

This story came from the print edition of the American Oil & Gas Reporter. For other great articles about exploration, drilling, completions and production, subscribe.