Estimation of shale volume from well logging data using Artificial Neural Network

Nội dung text: Estimation of shale volume from well logging data using Artificial Neural Network

1. 46 Journal of Mining and Earth Sciences Vol. 62, Issue 3 (2020) 46 - 52 Estimation of shale volume from well logging data using Artificial Neural Network Duong Hong Vu*, Hung Tien Nguyen Faculty of Oil and Gas, Hanoi University of Mining and Geology, Vietnam ARTICLE INFO ABSTRACT Article history: th The existence of shale has a major effect on reservoir quality because it Received 11 Feb. 2021 reduces the rock’s both the porosity and permeability. There are several Accepted 25th May 2021 types of shale, and they can be distributed in the sand in four different Available online 30th June 2021 ways: laminated, structural, dispersed, or any combination of these. Each Keywords: of them has various features and physical properties. Therefore, shale Neural network, volume estimation is one of the most important and challengin tasks to be solved information evaluation. There are many equations proposed to Volume of shale, calculate shale volume from Gamma - ray log; however, none of them Well logging. could be considered the best method that can be applied to all case studies. This study aims to propose a new approach to estimate shale volume from well - logging data. Gamma - ray and other logs were used as input data for an artificial neural network (ANN) to predict the shale volume. We apply this technique to the 1143 data set of the ocean drilling program (ODP) in the East Sea. The authors compared the result to core data and recognized that utilization of several logs and ANN gives a better estimation than conventional methods (more accurate and can reflect the trend of actual shale volume). Copyright © 2021 Hanoi University of Mining and Geology. All rights reserved. solution effects. Thus, the images are usually 1. Introduction calibrated by well - logging data (Wu et al., 2017). Submarine geological images are often built This work is often time - consuming, subjective, from the processing and analysis of geophysical and lacks quantification. To overcome these data (Braitenberg et al., 2006; Huang and Wang, challenges, now computational tools such as 2006; Ding et al., 2013; Gozzard et al., 2018; Ding machine learning (ML) and artificial intelligence et al., 2018). However, the marine geophysical (AI) are being used more and more in data data is of low resolution and non - unique of processing, including geophysical processing. Specifically, the application of these tools for the ___ analysis of well log data has been published by *Corresponding author Saumen Maiti et al. (2007), Bosch et al. (2013), E - mail: vuhongduong@humg.edu.vn Dekkers et al. (2014), and Aarushi Gupta and DOI: 10.46326/JMES.2021.62(3).06 Utkarsh Soumya (2020).
2. Duong Hong Vu, Hung Tien Nguyen /Journal of Mining and Earth Sciences 62(3), 46 - 52 47 In the East Sea, there are also similar studies 2. Methodology which were proposed by Karmakar et al. (2018); In this study, a new ANN model was Tse et al. (2019). The results obtained by these developed to predict Vsh from well logging data. authors show the effectiveness when applying To demonstrate the effectiveness of the model, these ML and AI techniques to solve geological the prediction result is compared with traditional structure interpretation missions. Shale has a methods. major effect on reservoir quality because it Normally, estimating Vsh from the gamma - reduces both porosity and permeability of the ray log still remains the most preferred approach. rock. There are several types of shale, and they The procedure is easy, straighforward and likely can be distributed in the sand in four different to give reasonable results. There are two ways: laminated, structural, dispersed or any commonly used equations introduced by Clavier combination of these. Each of them has various et al. (1971) and Steiber (1973). features and physical properties. Therefore shale 2 volume estimation is one of the most essential and VshClavier=1.7 - √3.38 − (( 푅 + 0.7) ) (1) challenging tasks to be solved in formation evaluation. So far, many mathematical models 푅 VshStieber = (2) have been proposed to calculate shale volume 3−2 푅 from Gamma - ray log. However, the results are In which IGR shaliness index is defined by a not always accurate due to complicated geological function of the GR ((gamma - ray)) gamma ray log conditions. This study aims to propose a new signal: (a) record of the response of the GR log for approach to estimate shale volume from well log a nearby known shale body and a nearby known data based on the application of an artificial clean rock. neural network. We apply this technique tothe 푅푙표𝑔 − 푅 푙푒 푛푅표 1143 data set, the ocean drilling program (ODP) 푅 = (3) 푅푠ℎ 푙푒 − 푙푒 푛푅표 in the East Sea (Figure 1). Figure 1. Research area. (source: Google maps).
3. 48 Duong Hong Vu, Hung Tien Nguyen /Journal of Mining and Earth Sciences 62(3), 46 - 52 An ANN network using back - propagation 3.1. Data preprocessing training algorithm (BPNN) and logsig activation Abnormal data can be considered as noise function was proposed to predict the volume of because they can negatively influence the ANN shale from well - logging data. We used data from model and may restrict the model in its well A and B to train the ANN model, then used generalization. The data set of three wells are data of well C to evaluate the accuracy of the tested for anomaly by Z - score outlier detection prediction model. Finally, the proposed ANN algorithm where a threshold Z - Score of 3 is model was used to predict Vsh from input data of selected (Tripathy et al., 2013). Any data points well C; then its performance was compared to above this threshold are marked as an outlier and Clavier and Steiber model. excluded from the training data. The Z - score is the score given to the participant as per their 3. Data sets performance: In this study, we used data sources from z = |X - X |/ SD (4) International Ocean Discovery Program i mean ( At this location, Where: Xmean - the mean value of the data; SD there are three wells OPD - 184 - 1143 A, B, and C - the standard deviation of the data; SD - the (Figure 1), for brevity in this article we named standard deviation of the data. them well A, B, C, respectively (Figure 1). To simplify the interpretation of the z - scores, The data used in this study include physical the following agreements were made as: and lithological parameters measured in a z 3 implies the result is unsatisfactory. susceptibility, and volume of shale (Vsh) in the To reduce volatility and eliminate statistical core sample (Table 1). noise, the dataset is further processed and smoothened by a low - pass second – order Table 1. Summary of well log data. Butterworth filter (Selesnick and Burrus, 1998). Parameters Well A Well B Well C Figure 2 shows the example comparing raw Number of core 242 118 74 and smoothened ((P - wave)) Pwave velocity data Top 24.05 44.05 382.35 of well A. Depth (m) Bottom 399.05 249.49 499.75 Min 1550.56 1556.63 1714.9 Max 2047.57 1803.72 2163.2 Vp (m/s) Mean 1733.11 1623.25 1877.8 Stdev 119.30 56.99 77.35 Min 20.25 26.50 24.75 Max 54.75 54.00 49.00 GR (API) Mean 37.43 41.04 33.15 Stdev 6.85 5.74 4.36 Min 1.37 1.47 1.61 Density Max 1.82 1.83 1.93 (g/cc) Mean 1.66 1.65 1.77 Stdev 0.09 0.08 0.07 Figure 2. Pwave velocity smoothened. Min 1.00 5.50 3.20 Magnetic Max 54.00 29.50 14.50 susceptib - Mean 14.66 18.73 8.04 3.2. Data analysis ility (SI) Stdev 8.89 6.80 2.10 The selection of input parameters for the Min 0.00 5.00 5.00 training process is an important step, which Max 98.00 98.00 94.00 Vsh (%) determines the accuracy of the ANN model. To Mean 72.39 78.36 60.89 decide which parameter to be used as input data, Stdev 26.89 24.02 30.58 the interrelationships between parameters were
4. Duong Hong Vu, Hung Tien Nguyen /Journal of Mining and Earth Sciences 62(3), 46 - 52 49 studied using cross plots as shown in Figure 3. A train the network, 15% is used for testing and, regression coefficient closer to 1 represents a 15% for the validation. Four parameters: P - wave positive correlation and closer to - 1 represents a velocity (m/hr), GR (API), Density (g/cc), negative correlation between variables. From Magnetic suscept (SI) are considered as input data Figure 3, we can see that all parameters: P - wave and the output value of the ANN model is the velocity (m/hr), GR (API), Density (g / cc), volume of shale. The calculated output from ANN Magnetic suscept. (SI) are suitable and can be after a cycle (or iteration) is compared with the retained in the ANN model development. actual output given in the sample dataset (Vsh of core) to trace the error. This error is propagated 4. Predict Vsh model development back to output neurons and hidden neurons so An ANN network using back - propagation that these neurons adjust their weights. This training algorithm (BPNN) and logsig activation bidirectional propagation is carried out function was proposed to predict the volume of repeatedly, until the error reaches a minimum shale from well logging data (Figure 4). In this value less than a certain allowable value or until study, we used data from well A and B to train the the number of loops reaches a predetermined ANN model, then used data from well C to value. The accuracy of the data model is evaluate the accuracy of the prediction model. demonstrated by the root mean square which A training data set of 316 samples from data serves as a metric to score the predictions of the of two wells A and B includeseveral parameters: ANN model with the expected result. longitudinal wave velocity, gamma - ray, density, 2 ∑ ( − ) magnetic susceptibility from well logging data, √ 푠ℎ 푒 𝑖 푡 푠ℎ 푡 푙 (5) RMSerror = and volume of shale from core data. This database 푛 is divided into 3 sets: 70% of the sample is used to Figure 3. Crossplot between well logging parameters.
5. 50 Duong Hong Vu, Hung Tien Nguyen /Journal of Mining and Earth Sciences 62(3), 46 - 52 Determining the number of neurons in the 5. Results and discussions hidden layer is a challenging step in model design, The proposed ANN model was used to and there is no rigid rule to do it. In this study, to predict Vsh from input data of well C, then determine the optimal number of hidden neurons, compare its performance with Clavier and Steiber different scenarios were carried out with variable model. It can be seen that ANN model predictions numbers of neurons in the hidden layer and tests are much more accurate than the other two for their effect on the final prediction (Figure 5). It traditional models, as confirmed by both RMSE is worth noting that the number of neurons in the and R2 in Table 2. It is observed from Figure 6 and hidden layer should be chosen carefully since too Figure 7 that prediction from the ANN model many neurons in the hidden layer can lead to follows the trend of actual Vsh. overfitting, making the network lose its generalization. Therefore, we decided to use 30 Table 2. Model performance comparison. neurons in one hidden layer, Figure 5 shows that Model RMSE R2 the ANN model with one hidden layer including 1 Clavier model 0.2616 0.022 30 neurons given the RMSE = 0.002 and R2 = 2 Steiber model 0.29 0.019 0.956. 3 ANN model 0.0037 0.92 Figure 5. (a, b) - Optimum number of neural in hidden layer. Figure 6. Vsh prediction by ANN for well C.
6. Duong Hong Vu, Hung Tien Nguyen /Journal of Mining and Earth Sciences 62(3), 46 - 52 51 Figure 7. Vsh prediction by Clavier, Steiber model for well C. Meanwhile, Vsh calculated by Clavier and Steiber equations can not reflect this trend and References has significant gaps with the actual Vsh. It can be Aarushi Gupta and Utkarsh Soumya, (2020). Well explained that these traditionally predicted log interpretation using deep learning neural equations usually are proposed from a limited networks. International Petroleum Technology database in a particular research area; therefore, Conference, Dhahran, Kingdom of Saudi Arabia. when applying them to other cases which have different geological properties, the result is Bosch, D., Ledo, J., and Queralt, P., (2013). Fuzzy usually inaccurate. Logic Determination of Lithologies from Well Log Data: Application to the KTB Project Data 6. Conclusion set (Germany). Surveys in Geophysic 34(4), 413 - 439. This study presented the practical use of data analysis and AI applications to solve geological Braitenberg, C., Wienecke, S., and Wang, Y., interpretation problems. The ANN model is (2006). Basement structures from satellite- developed to predict Vsh from longitudinal wave derived gravity field: South China Sea ridge. velocity, gamma-ray, density, and magnetic Journal of Geophysical Research: Solid Earth B5 susceptibility. The ANN model shows its (111). advantages compared to other traditional Clavier, C., Hoyle, W. R., Meunier, D., (1971). methods. Therefore it can be recommended as an Quantitative interpretation of TDT logs: Parts I effective and suitable method to determine the and II. J. Pet. Technol. 23, 743 - 763. volume of shale presenting in rock in the vicinity of the East Sea. Recommendations for future work Dekkers, M. J., Heslop, D., Herrero - Bervera, E., are to update data from other wells to the ANN Acton, G., and Krasa, D., (2014). Insights into model to increase its accuracy. magmatic processes and hydrothermal alteration of in situ superfast spreading ocean Author contributions crust at ODP/IODP site 1256 from a cluster analysis of rock magnetic properties. The author Duong Hong Vu proposes ideas Geochemistry, Geophysics, Geosystems 8(15), and contributes to the manuscript. The author 3430 - 3447. Hung Tien Nguyen constructs the manuscript and contributes to the material analyses. The authors Ding, W., Franke, D., Li, J., and Steuer, S., (2013). both declare no conflict of interest. Seismic stratigraphy and tectonic structure
7. 52 Duong Hong Vu, Hung Tien Nguyen /Journal of Mining and Earth Sciences 62(3), 46 - 52 from a composite multi - channel seismic modelling and classification of lithofacies using profile across the entire Dangerous Grounds, well log data: a case study from KTB borehole South China Sea. Tectonophysics 582, 162 - site. Geophysical Journal International 169, 733 176. - 746. Ding, W., Li, J., and Clift, P. D., (2016). Spreading Selesnick, I. W. & Burrus, C. S., (1998). Generalized dynamics and sedimentary process of the digital Butterworth filter design. IEEE Southwest Sub - basin, South China Sea: Transactions on Signal Processing, 46(6), 1688 Constraints from multi - channel seismic data - 1694. and IODP Expedition 349. Journal of Asian Steiber, R. G., (1973). Optimization of shale Earth Sciences 115, 97 - 113. volumes in open hole logs. J. Pet. Technol. 31, Gozzard, S., Kusznir, N., Franke, D., Cullen, A., 147 - 162. Reemst, P., and Henstra, G., (2018). South Tse, K. C., Chiu, H. - C., Tsang, M. - Y., Li, Y., and Lam, China Sea crustal thickness and oceanic E. Y. J. F. o. E. S., (2019). Unsupervised learning lithosphere distribution from satellite gravity on scientific ocean drilling datasets from the inversion. Petroleum Geoscience 25, 112 - 128. South China Sea. Frontiers of Earth Science Huang, W., and Wang, P., (2006). Sediment mass 1(13), 180 - 190. and distribution in the South China Sea since Tripathy, S. S., Saxena, R. K., & Gupta, P. K., (2013). the Oligocene. Science in China Series DEarth Comparison of statistical methods for outlier Sciences 11(49), 1147 - 1155. detection in proficiency testing data on Karmakar, M., Maiti, S., Singh, A., Ojha, M., and analysis of lead in aqueous solution. American Maity, B. S. J. M. G. R., (2018). Mapping of rock Journal of Theoretical and Applied Statistics types using a joint approach by combining the 2(6), 233 - 242. multivariate statistics, self-organizing map and Wu, H., Shi, M., Zhao, X., Huang, B., Zhang, S., Li, H., Bayesian neural networks: an example from Yang, T., and Lin, C., (2017). IODP 323 site. Marine Geophysical Research Magnetostratigraphy of ODP Site 1143 in the 39, 407 - 419. South China Sea since the Early Pliocene. Saumen Maiti., Ram Krishna Tiwari and Hans - Marine Geology 394, 133 - 142. Joachim K¨umpel., (2007). Neural network