EXPLAINABLE ARTIFICIAL INTELLIGENCE (XAI) - ONE OF THE MAIN CHARACTERISTICS OF PETROLEUM DATA ANALYTICS (PDA); Section - 4

History of eXplainable AI (XAI) in the Petroleum Data Analytics

Since 2001 seven SPE papers have been published that include what today is called Explainable Artificial Intelligence (XAI). These seven SPE papers include total of 38 figures that use the Predictive Analytics models in order to explain how these models can explain the physical phenomena that was modeled purely based on field measurements rather than mathematical equations. To show the historical application of XAI in Petroleum Data Analytics, this section of the article shows most of the figures from these seven SPE technical papers.

The first technical paper that included the first version of what today is called eXplainable Artificial Intelligence (XAI) was an SPE paper (SPE 72385) that was published in 2001[1]. This paper is about the application AI and Machine Learning in modeling hydraulic fractures. Figure 24 shows couple of XAI related figures from this SPE paper. In this paper the data-driven model that was developed (trained, calibrated, and validated) using artificial neural network was used to explain the impact of three hydraulic fracturing related parameters. These hydraulic fracturing related parameters are number of perforations, injection rate, and total amount of water that is injected. The XAI in this paper explains the impact of these parameters on the well productivity (5 years Cumulative gas production). Such explanations of the influence of these parameters on well productivity are used to identify well productivity optimization as a function of number of perforations, injection rate, and total amount of water that is injected.

Figure 24, that shows the Figure 7 of this paper, explains the conditions under which the well productivity (5 year cumulative gas production) can reach its maximum value. For this particular well (Carlson, Thomas #075) the 5 year cumulative gas production can reach its maximum value of 57,050 MSCF as long as the number of perforations and the rate of injection are kept at their minimum values (injection rate at 25 bbls/min and 8 perforations) and can reach its maximum value of 63,400 MSCF, as long as the number of perforations and the rate of injection are kept at their maximum values (60 bbls/min for injection rate and 26 perforations).

As shown Figure 24 while these parameters (the number of perforations and the rate of injection) are kept at their minimum value the highest 5-year cumulative gas production (57,050 MSCF) is achieved at 680 barrels of water, on the other hand when these values are kept at their maximum the highest 5-year cumulative gas production (63,400 MSCF) is achieved at about 725 barrels of water injected. The conclusion may be that for this well the ideal water injection volume is about 700 barrels. This SPE paper include two more similar figures showing such explanation for two more wells in the same field.

Similar to the paper that was published in 2001, the second technical paper that included eXplainable Artificial Intelligence (XAI) was an SPE paper (SPE 77597) that was presented at ATCE and was published in 2002[2]. This paper is also about the application AI and Machine Learning in modeling re-fracturing (re-stimulation) in DJ-Basin, Colorado. Figure 25 is the explanation of the modification of the well productivity as a function of parameters such as amount of injected proppant (sand) and fluid as well as number of perforations. In this figure the monthly barrel of oil equivalent as well productivity is show for Bohlender 8-5 well during the re-frac process.

The third technical paper that included eXplainable Artificial Intelligence (XAI) was an SPE paper (SPE 77659) that was also published in 2002[3]. This paper covers purely data-driven modeling of BP’s Prudhoe Bay (Alaska, USA) surface facility using Artificial Intelligence and Machine Learning. This paper includes figures that generate eXplainable AI covering separator pressure, temperature, hydrocarbon rates, and compressor inlet suction pressure of eight different three-phase separator facilities. Figure 26 and Figure 27 show two of such figures from this SPE paper.

The fourth technical paper that included eXplainable Artificial Intelligence (XAI) was an SPE paper (SPE 89033) that was published in 2004[4] that included more examples of XAI from the Prudhoe Bay’s surface facility’s purely data-driven model. This paper includes 7 figures that covers XAI of the Prudhoe Bay’s surface facility. Figure 28 and Figure 29 show two of such figures from this SPE paper.

The fifth technical paper that included Explainable Artificial Intelligence (XAI) was an SPE paper (SPE 95942) that was published in 2005[5]. This paper covers XAI of hydraulic fracturing of Golden Trend filed of Oklahoma. This paper includes two figures that show XAI results for the hydraulic fracturing of multiple individual wells in this field.

Figure 30 explains how Shot per foot (x axis) and injection rates – BPM/ft. (y axis) influence 30 Year EUR well productivity, for three individual wells. These wells show different types of responses as the number of perforations and average rate of injection per foot of pay thickness changes. The production response is different for each of these wells as number of perforations and the average injection rates start to increase.

The sixth technical paper that included Explainable Artificial Intelligence (XAI) was an SPE paper (SPE 101474) that was published in 2006[6]. This paper covers XAI of Smart Proxy Modeling of traditional numerical simulation models. Figure 31 provides Type Curves of the model that demonstrates the impact of a specific reservoir parameter on oil production while Figure 32 provides Type Curves of the model demonstrating the impact of a another specific reservoir parameter on water cut.

The seventh technical paper that included Explainable Artificial Intelligence (XAI) was an SPE paper (SPE 139032) that was published in 2010[7]. This paper covers completion design characteristics of hydraulic fracturing in Bakken Shale. The Explainable Artificial Intelligence model that is shown in this paper was developed using IMagine[8] Software Application that is used to develop Top-Down Models (TDM)[9], [10]. Figure 33 demonstrates the use of XAI to explain the oil production characteristics as a function of “Lateral Length” and “Injected Fracturing Fluid” (graph on the left) and the oil production characteristics as a function of “Pay Thickness” and “Lateral Length” (graph on the right).

Conclusions

Petroleum Data Analytics is a solid Engineering Application of Artificial Intelligence and Machine Learning that has the capability of addressing all Petroleum Engineering related problems. Engineering Application of Artificial Intelligence and Machine Learning has many differences with the Non-Engineering Application of this technology (Artificial General Intelligence - AGI). One of the major characteristics of Petroleum Data Analytics is its use of today is called Explainable AI (XAI). Purely data-driven models that are developed through Petroleum Data Analytics are not “Black Box”. Petroleum Data Analytics started to go above and beyond extraction and recognition of “Correlation” from field measurements and started to address “Causation”. While this was done in late 1990s and early 2000, it was not called XAI at that time. Now, especially since 2016, XAI has become an important topic in non-engineering applications of AI & Machine Learning while it was addressed more than a decade before that time by Engineering Application of AI&ML through Petroleum Data Analytics.

[7] Field Development Strategies for Bakken Shale Formation. SPE-139032. SPE Eastern Regional Meeting held in Morgantown, West Virginia, USA, 12–14 October 2010.

[8] https://intelligentsolutionsinc.com/Products/IMagine.shtml

[9] https://intelligentsolutionsinc.com/Technology/TDM-2.shtml

[10] https://intelligentsolutionsinc.com/Technology/TDM-1.shtml

[6] Uncertainty Analysis of a Giant Oil Field in the Middle East Using Surrogate Reservoir Model. SPE 101474. Abu Dhabi International Petroleum Exhibition and Conference held in Abu Dhabi, U.A.E., 5–8 November 2006.

[5] Analysis of Best Hydraulic Fracturing Practices in the Golden Trend Fields of Oklahoma. SPE 95942. SPE Annual Technical Conference and Exhibition held in Dallas, Texas, U.S.A., 9 – 12 October 2005.

[4] Recent Development in Application of Artificial Intelligence in Petroleum Engineering. Shahab D. Mohaghegh, West Virginia University & Intelligent Solutions, Inc. SPE 89033. JPT Distinguished Author Series, April 2004.

[3] Prudhoe Bay Oil Production Optimization: Using Virtual Intelligence Techniques, Stage One: Neural Model Building. SPE 77659. SPE Annual Technical Conference and Exhibition held in San Antonio, Texas, 29 September–2 October 2002.

[2] Identification of Successful Practices in Hydraulic Fracturing Using Intelligent Data Mining Tools; Application to the Codell Formation in the DJ –Basin. SPE Annual Technical Conference and Exhibition held in San Antonio, Texas, 29 September–2 October 2002.

[1] Identifying Best Practices in Hydraulic Fracturing Using Virtual Intelligence Techniques, SPE 72385, Mohaghegh, et. al. SPE Eastern Regional Meeting held in Canton, Ohio, 17–19 October 2001.