Formation evaluation

Machine Learning Approach Aids Source Rock Evaluation

Geochemical parameters such as total organic carbon (TOC) provides valuable information to understand rock organic richness and maturity and, therefore, optimize hydrocarbon exploration. This article presents a novel work flow to predict continuous high-resolution TOC profiles using machine learning.

Abstract Frame with 3D Geometric Shapes, Illuminated Spheres, Rock Background
Source: Akinbostanci/Getty Images

Geochemical parameters are crucial data sets to enhance prediction accuracy of organic rich zones. The current laboratory analysis methods of obtaining these measurements, however, are costly and time-consuming. While there exists a rich body of knowledge and equations for estimating total organic carbon (TOC) from wireline logs, new research efforts are continuing, especially leveraging machine learning (ML), to predict geochemical parameters from wireline logs. These methods, however, rely heavily on data availability and quality. Geochemical parameters such as TOC provide valuable information to understand rock organic richness and maturity and, therefore, optimize hydrocarbon exploration.

TOC can be defined as the amount of organic content in a rock. Organic matter is the most important component in source rock evaluation. Understanding variations in TOC, therefore, is important for evaluating hydrocarbon source rock quality, identifying organic rich zones, and enhancing unconventional reservoirs characterization. Previously, mathematical calculations using logs helped in estimating TOC values and determining source rock productivity (Passey et al. 1990). The interpretation led to the identification of organic content and mature organic rich intervals. Two of the ways to calculate TOC using logs are the sonic/resistivity ratio (Ahangari et al. 2022) and log combinations (Fertl et al. 1988). These approaches provide assessment of source rock ability to release hydrocarbons. Because of the limitations of these methods, this article presents a novel work flow to predict continuous high-resolution TOC profiles using ML, taking only few minutes. This approach helps increase the precision of the geochemical parameter predictions. It is nondestructive and requires minimal need for laboratory testing.

TOC is a critical parameter for the identification of organic rich zones and source rock characterization in a rock formation. One important limitation of laboratory-measured TOC data is the fact that these measurements are discrete and scattered and do not cover the entire area of interest because of the destructive nature of the analysis. Images containing color attributes also can be incorporated into the work flow to help in predicting geochemical parameters. An ML work flow has been introduced to detect and visualize different geochemical parameters to enhance organic rich zone characterization in a nondestructive way (Shalaby et al. 2019). Here, we showcase an ML work flow that uses core images and TOC laboratory data to generate continuous high-resolution TOC profiles in a timely manner.

Core photos were decomposed into entropy and color attributes (red, green, and blue curves). A moving average window was used to extract continuous visual curves of the attributes. These attributes were matched with their corresponding TOC measurements as measured in the laboratory using the Rock Eval pyrolysis instrument.

The work flow consists of two ML algorithms. The first algorithm is unsupervised K-means clustering, which uses the extracted entropy and color curves as inputs. This generated a continuous curve of clusters based on the extracted attributes. Based on the previous knowledge of the core TOC measurements, the number of rock clusters were selected. For example, different TOC measurements were identified, which includes high, medium, and low values. In this case, the number of clusters generated will be three. This approach is useful to classify intervals with high TOC values directly. The second algorithm applied support vector regression (SVR), with the extracted attributes tied to TOC values. This approach used 80% of the data to train the model and 20% for the blind testing and validation of the model. The final result can be used to produce a continuous high-resolution TOC profile (Fig. 1).

Fig. 1—A sketch of the ML work flow to predict a TOC profile.
Source: Saudi Aramco

The work flow generated promising results consisting of continuous high-resolution TOC profiles for source rock intervals through ML algorithms in a nondestructive manner (Fig. 2). The results show successful generation of a continuous TOC profile with 90% prediction accuracy within ±1% of measured data (Fig. 3). Using consistent and high-quality images along with adequate data distribution can help produce results with a high degree of prediction accuracy. Building and training the model based on high-quality images and a wide range of data distribution enhances the predictive results and ultimately improves the characterization of organic rich zones and unconventional resources (Peters et al. 2016).

Fig. 2—ML work flow prediction results for TOC. Core photos and discrete laboratory measurements are the main inputs (left). Image attribute analysis and entropy (middle). Final outputs are K-means clustering of TOC values based on image characteristics and continuous high-resolution prediction of TOC profile (right).
Source: Saudi Aramco

The ML work flow predicted accurate TOC profiles for source rock intervals. In addition, obtaining a large data set for training helps in generating accurate results in a cost- and time-effective ways for TOC determination, with fewer laboratory testing and nondestruction of rock samples. This nondestructive and fast approach optimized accurate identification of TOC, which can directly affect organic rich zones classification and, therefore, improve source rock characterization. Another implication of the work flow is to use the generated TOC profile for a single well and apply it to calibrate other TOC measurements for the offset wells with similar geology and geochemical attributes. Ultimately, this prediction can aid in enhancing hydrocarbon exploration in a cost-effective way within a few minutes.

Fig. 3—Measured vs. predicted TOC cross plot. The prediction accuracy is 90% within ±1% of measured data; where y = x is the 1:1 line where predicted values equal true values, y = x+1 is the line where predicted values are above the true values by 1%, and y = x−1 is the line where predicted values are below the true values by 1%.
Source: Saudi Aramco

Summary and Conclusions To understand source rock richness and its ability to release hydrocarbons, it is critical to analyze geochemical parameters, which can distinguish the amount of organic matter in a rock and detect rock maturity (Tissot et al. 1987). Thus, analyzing TOC and other geochemical parameters is essential to understanding the possibility of hydrocarbon expulsion rate (Carvajal-Ortiz et al. 2018). The presence of different geochemical parameter measurements within a rock unit determines variations in organic rock content and, therefore, enhances the identification of organic rich zones (Peters et al. 2010). These parameters play an important role in identifying the hydrocarbon source and type.

Although laboratory geochemical data provide a good understanding of rock chemical content, using ML applications to identify and quantify geochemical parameters can be a cost- and time-effective tool, providing inputs and parameters to enhance source rock characterization and field development. The work flow showcases a successfully generated continuous high-resolution TOC profile with a prediction accuracy of up to 90% in a timely and cost-effective way. The result shows a great potential to estimate geochemical parameters using an ML work flow. The work flow is implemented to evaluate and predict TOC in a nondestructive way and in a timely manner. This approach can also be used to predict and calibrate TOC values (Bath 2002) for the nearby wells with the recommendation to use sufficient data distribution and superior image quality for higher prediction.

Ahangari, D., Daneshfar, R., Zakeri, M., Ashoori, S., and Soulgani, B.S. 2022. On the Prediction of Geochemical Parameters (TOC, S1 and S2) by Considering Well Log Parameters Using ANFIS and LSSVM Strategies. Petroleum, 8 (2), 174–184.

Bath, A. 2002. Geochemical Parameters Required From the SKB Site Characterisation Programme. Sweden.

Carvajal-Ortiz, H., and Gentzis, T. 2018. Geochemical Screening of Source Rocks and Reservoirs: The Importance of Using the Proper Analytical Program. International Journal of Coal Geology, 190, 56–69.

Fertl, W.H., and Chilingar, G.V. (1988). Total Organic Carbon Content Determined From Well Logs. SPE Formation Evaluation, 3 (02), 407–419.

Passey, Q. R., Creaney, S., Kulla, J. B., Moretti, F. J., & Stroud, J. D. 1990. A Practical Model for Organic Richness From Porosity and Resistivity Logs. AAPG Bulletin, 74 (12), 1,777–1,794.‏

Peters, K.E., Walters, C.C., and Moldowan, J.M. 2010. The Biomarker Guide (Vol. 1). Cambridge University Press.‏

Peters, K.E., Xia, X., Pomerantz, A.E., and Mullins, O.C. 2016. Geochemistry Applied to Evaluation of Unconventional Resources. Unconventional Oil and Gas Resources Handbook, 71–126. Gulf Professional Publishing.

Shalaby, M.R., Jumat, N., Lai, D., and Malik, O. 2019. Integrated TOC Prediction and Source Rock Characterization Using Machine Learning, Well Logs, and Geochemical Analysis: Case Study From the Jurassic Source Rocks in Shams Field, NW Desert, Egypt. Journal of Petroleum Science and Engineering, 176, 369–380.

Tissot, B.P., Pelet, R., and Ungerer, P.H. 1987. Thermal History of Sedimentary Basins, Maturation Indices, and Kinetics of Oil and Gas Generation. AAPG Bulletin, 71 (12), 1,445–1,466.‏