Automated Seismic Interpretation

Automated Seismic Interpretation: Machine learning technologies are
being used to develop automated seismic interpretation to identify

geological features, such as faults and stratigraphic horizons.
1 Nwosu Obinnaya Chikezie Victor1,2 Lucky Oghenechodja Daniel1
2 1
Faculty of Engineering and the Built Environment, University of Johannesburg, Johannesburg, 2006
3 South Africa
4 2
Department of research, education and communication, UiT the Arctic University of Norway,
5 Tromsø, 6050 Norway
6 * Correspondence:
7 *Nwosu Obinnaya Chikezie Victor
8 nch015@uit.no, 220117941@student.uj.ac.za
9 *Lucky Oghenechodja Daniel 221189560@student.uj.ac.za
10 Keywords: Automated Seismic Interpretation, Machine Learning, Geological features, Faults,

11 Stratigraphic Horizons.
12 Abstract
13 This article describes the use of machine learning technologies to create an automated seismic
14 interpretation capable of identifying geological features such as fractures and stratigraphic horizons.
15 Geologists use automated seismic interpretation (ASI) to extract geologic information from seismic
16 data. Geologic features can be identified through the amplitude, frequency, and polarization
17 parameters of seismic signals, and automated techniques can be used to identify geologic features.
18 This article examines the present state of automated seismic interpretation and the potential of
19 machine learning technologies for this endeavor. A review of the research indicates that machine
20 learning techniques can be used to accurately identify faults and stratigraphic horizons in seismic
21 data. The authors discuss the features that can be extracted by machine learning algorithms and
22 compare the various machine learning techniques applied to seismic interpretation. The article also
23 discusses the difficulties associated with automated seismic interpretation and the need for additional
24 development to improve the precision of seismic interpretation. Future research, according to the
25 authors, should concentrate on increasing the accuracy of fault and horizon recognition and devising
26 algorithms to detect other geological features. Overall, the article provides a summary of the current
27 state of automated seismic interpretation and the obstacles that must be overcome. In addition, it
28 demonstrates the capability of machine learning technologies to recognize faults and stratigraphic
29 horizons in seismic data. With additional research, the precision of automated seismic interpretation
30 can be enhanced, leading to more precise geological interpretations and a deeper comprehension of
31 the Earth's subsurface.
32 1 Introduction
33 For As an effective tool to image subsurface layers and identify geological structures, seismic
34 interpretation has played an important part in understanding the Earth's interior. However, the
35 seismic interpretation process is an arduous and time-consuming task— requiring highly trained and
36 experienced interpreters to identify geological features, such as faults and stratigraphic horizons.
37 Over time, this manual interpretation has become more demanding, requiring the most modern
38 analysis algorithms, faster computing, and experienced interpreters. As a result, the seismic
39 interpretation process is expensive and therefore functions as a bottleneck to exploration and
40 production activities. Additionally, due to the data acquisition costs of seismic surveys, each step of
41 the seismic interpretation process must be as efficient as possible. This is where automated seismic
42 interpretation comes into play(Agena et al. 2018). Automated seismic interpretation uses machine
43 learning to develop automated techniques to identify geological features from seismic surveys
44 without requiring extensive manual interpretation. The effectiveness of machine learning for seismic
45 interpretation, primarily deep learning, has been established for quite some time. At the same time,
46 there are still several challenges that need to be addressed to use it in a production setting. This
47 research article provides a comprehensive overview of automated seismic interpretation, its
48 applications, and the challenges that come with the use of machine learning for seismic
49 interpretation. The article begins by describing the seismic interpretation process and how it can be
50 automated. Next, different types of machine learning models used for automated seismic
51 interpretation are discussed along with their respective challenges. Following this, the potential
52 benefits, and applications of machine learning for automated seismic interpretation are discussed.
53 Lastly, a conclusion is provided summarizing the article and its findings, along with implications for
54 future research(Ayadi et al.2019).
55 1.1. The Seismic Interpretation Process
56 Seismic interpretation is a part of seismic geology, which is a tool used by the oil & gas industry to
57 create images of the subsurface to aid in exploration and drilling. It involves collecting and
58 processing seismic data, which is data collected at different frequencies that reflect off different
59 layers in the subsurface. This data is then interpreted by experienced seismic interpreters to identify
60 geological features, such as faults and stratigraphic horizons, to help with the analysis of reservoir
61 characteristics and risks. Seismic interpretation typically involves three main stages: data acquisition,
62 data processing and interpretation(Chen, 2020).
This is a provisional file, not the final typeset article 2

63
64 Fig 1: 3-D Seismic Interpretation
65 Fig 1 shows three-dimensional seismic interpretation is the analysis of three-dimensional seismic

66 data to obtain insight into a region's geology and geological structure. Geological mapping and
67 hydrocarbon exploration use this form of interpretation. It entails the analysis of three-dimensional
68 seismic data sets collected using surface or near-surface directional seismic sensors. Using
69 sophisticated software, 3-D seismic data is processed to generate detailed images of the subsurface
70 (Chaitanya and Munroe 2019). These images provide information regarding the structure, physical
71 properties, and textures of subsurface materials. Geologists can determine where and how
72 hydrocarbon reserves are likely to be located by analyzing this data. Seismic interpretation in three
73 dimensions is an essential technique for oil and gas exploration and development. It is used to
74 determine the optimal drilling locations and maximize the likelihood of discovering new reserves.
75 Seismic interpretation in three dimensions is also used to assess the safety of seismic activities by
76 providing a precise image of the subsurface structure (Cook et al. 2020).
77 1.2 Data Acquisition
78 Data acquisition involves procurement of the seismic data from the field. This often involves the use
79 of a seismic source, such as an air gun, which can generate energy waves that travel through the earth
80 and the boundaries between different geological layers (Jan et al. 2020). The reflected waves from
81 these boundaries produce the seismic reflection data that is recorded by geophones or a receiver
82 array. The quality and accuracy of the seismic data are dependent on the energy source deployed for
83 the survey, the type of equipment used, and the design of the data acquisition, among other factors. It
84 is therefore important to get the data acquisition stage right for the subsequent stages of data
85 processing and interpretation to yield accurate results (Khordadmehr 2017).
3
86
87 Fig 2: Seismic Data Processing for Subsurface Characterization
88 Fig 2 shows processing seismic data is an indispensable component of comprehensive subsurface

89 characterization. By utilizing sophisticated data processing techniques, geoscientists can precisely
90 define the subsurface structure to identify subsurface resources. To mimic the spatial variation of
91 seismic waves as they travel through the subsurface, seismic data processing techniques are utilized.
92 This captured subsurface data allows geoscientists to construct a three-dimensional model of the
93 subsurface, which can be used to discern geological and geophysical features and identify potential
94 hydrocarbon or mineral resources (Meesters et al.2020). In seismic data processing, images of the
95 subsurface are enhanced using sophisticated techniques such as migrations, deconvolution, static
96 corrections, and wavelet extraction. To accurately model seismic waves as they travel through the
97 subsurface, migration techniques are utilized to account for changes in their direction and velocity.
98 Deconvolution is utilized to extract useful information from seismic data and remove any
99 disturbance. As a result of the limitations of seismic sources and receivers, static corrections are used
100 to close the data disparity. Wavelet extraction techniques are used to precisely define subsurface
101 features and to determine the optimal exploration depth range(Murilo et al. 2021).
102 1.3 Data Processing
103 Data processing is the second stage of the seismic interpretation process. In this step, the raw seismic
104 data is subjected to a suite of processing algorithms that enhance the data for interpretation. A
105 common suite of pre-stack processing algorithms includes trace editing, noise reduction, trace-to-
106 trace coherency, muting, phase rotation, amplitude recovery, filtering, and gain adjustment, among
107 others.
108 1.4 Interpretation
109 In the final step of the seismic interpretation process, the pre-stack processed data is used to create
110 seismic images. These seismic images are then interpreted by experienced interpreters to identify
111 geological features, such as faults and stratigraphic horizons. This involves manual measurements
112 being taken from the seismic data to quantify the size and shape of subsurface features and the
113 correlations between different seismic data deposits.

114
115 Fig 3: Underwater geophysical surveying
116 Fig 3 shows underwater geophysical surveying is a type of surveying that employs various
117 techniques, including hydro-acoustic and side-scan sonar, magnetometer, and seismic reflection or
118 refraction, to measure or map the physical characteristics and properties of the sea floor. These
119 surveying techniques permit scientists to examine the physical characteristics of the sea floor to gain
120 a better understanding of how our oceans are structured, how they interact with other bodies of water,
121 and what species of aquatic life may inhabit the surveyed area. Underwater geophysical surveying is
122 also essential for mineral and resource exploration, engineering site investigations, and the
123 installation of drainage conduit. Using specialized instruments and technologies, survey teams can
124 precisely measure the depth, composition, structure, and even contours of the seafloor, enabling them
125 to compile accurate geophysical data for use in exploration, engineering, and development.
126 1.5 Automated Seismic Interpretation
127 Automated seismic interpretation refers to the use of machine learning techniques to interpret seismic
128 data and identify geological features with minimal human intervention. The advent of high-speed
129 computers and powerful algorithms has made it possible to apply automated machine-learning
130 models to seismic surveys. The primary goal of automated seismic interpretation is to develop
131 techniques to identify geological features without requiring extensive manual interpretation. This is
132 highly beneficial in speeding up the interpretation process and reducing the need for skilled and
133 experienced interpreters. Additionally, machine learning models have been found to improve the
134 accuracy of seismic interpretation, identifying features and deposits that may have been missed by
135 manual interpreters.
136 1.6 Types of Machine Learning Models
5
137 Various types of machine learning models have been used for automated seismic interpretation. Each
138 type has significant trade-offs in terms of complexity, accuracy, and speed.
139 1.7 Support Vector Machines (SVMs)
140 SVMs are a supervised learning technique used to classify data into different classes. They are used
141 to create an individual classifier for each class of geological input features. By implementing multiple
142 SVM classifiers in combination, the accuracy of the automated seismic interpretation can be
143 improved. However, this approach is complex and computationally expensive.
144 1.8 Random Forest
145 Random forest is an unsupervised learning algorithm used for automated seismic interpretation. It
146 uses the bagging approach to classify input data into different classes. Random forest is
147 computationally inexpensive and not as complex as SVMs. However, its accuracy is not as high as
148 SVMs.
149 1.9 Neural Networks
150 Neural networks are a type of supervised learning technique used for automated seismic
151 interpretation. Unlike SVMs, neural networks require large datasets and are numerically intensive.
152 This means that they require extensive computational resources.
153 1.10 Deep Learning
154 Deep learning is a type of machine learning model on the architecture of the human brain. These
155 models are designed to learn by extracting more abstract features from the data rather than relying on
156 hand-selected features. This makes it well-suited for automated seismic interpretation as it can be
157 used to identify the subtle patterns and features that form the geophysical structure of the subsurface.
158 However, deep learning models also require large datasets and are computationally expensive.
159 1.11 Benefits and Applications of Automated Seismic Interpretation
160 Machine learning models for automated seismic interpretation bring many potential benefits and
161 applications. One of the biggest benefits is the speed with which seismic surveying can be
162 streamlined. Automating the interpretation process can significantly reduce the amount of time
163 required to produce interpretative results, leading to faster exploration decisions. In addition to this,
164 the accuracy of seismic interpretation can be improved. Automated seismic interpretation reduces the
165 potential for human bias and provides reliable information about subsurface structures. This can lead
166 to improved exploration decisions, reducing the amount of wasted money spent on fruitless wells.
167 Furthermore, automated seismic interpretation can lead to cost savings. By reducing the need for
168 highly trained interpreters and keeping the seismic surveys required to a minimum, automated
169 seismic interpretation can reduce costs associated with exploration and drilling activities. Lastly,
170 automated seismic interpretation has the potential to open exploration to non-traditional players.
171 Geophysical services can be delivered to the public from anywhere in the world, enabling anyone
172 with access to seismic surveys to benefit from automated seismic interpretation. This could
173 ultimately open the door to players who may not have had access to the same seismic surveys as the
174 larger players.
175 1.12 Conclusion

176 In conclusion, automated seismic interpretation is an important innovation in the seismic
177 interpretation process. By leveraging machine learning models, the interpretation process can be
178 automated and streamlined, reducing the need for highly trained interpreters and the associated costs.
179 In addition, automated seismic interpretation can improve the accuracy of seismic interpretation by
180 eliminating the potential for human bias and providing reliable information about subsurface
181 structures. Overall, automated seismic interpretation has several potential benefits and applications. It
182 can be used to speed up the interpretation process, improve the accuracy of interpretation, reduce
183 costs associated with seismic interpretation, and open exploration to non-traditional players.
184 However, there are several challenges that need to be addressed to make automated seismic
185 interpretation a viable option for production deployment.
186 2 Introduction
187 In recent years, technological development has introduced seismic interpretation techniques into the
188 domain of automation. This automated seismic interpretation was made possible by the development
189 of various machine learning technologies that enable the efficient and cost-effective interpretation of
190 seismic data. Automated seismic interpretation has been used to identify various geological features,
191 such as fractures and stratigraphic horizons, which are essential for gaining insight into the
192 subsurface structure. Increasingly, reservoir characterization, field development, and exploration
193 management are among the applications where automated seismic interpretation is being utilized.
194 This literature review discusses the various machine learning technologies applicable to automated
195 seismic interpretation and outlines the benefits and drawbacks of such methods. In addition, the
196 review investigates the various applications of automated seismic interpretation and the difficulties
197 encountered when endeavoring to interpret seismic data(Naili et al.2021).
198 2.1 Technologies of Machine Learning for Automated Seismic Interpretation
199 Various machine learning technologies have been developed over the past few years to enable the
200 automated interpretation of seismic data. Artificial neural networks (ANNs), support vector machines
201 (SVMs), convolutional neural networks (CNNs), and deep learning networks (DLNs) are among
202 these technologies. These algorithms use input seismic data to identify objects of interest and
203 understand the patterns and characteristics of geological structures. The primary benefit of these
204 technologies for automated seismic interpretation is that they enable rapid and cost-effective analysis
205 of seismic data without requiring costly processing platforms. ANNs are a form of machine learning
206 used for automated seismic interpretation. In artificial neural networks, seismic data is used to train a
207 network of interconnected nodes (neurons) to recognize patterns or structures in the data. These
208 networks can then be used to identify faults and stratigraphic horizons in seismic data. ANNs are
209 advantageous because they enable the identification of intricate seismic features that could not be
210 detected using conventional techniques are another technique for automated seismic interpretation
211 that utilizes machine learning. SVMs have the advantage of being able to manage vast datasets while
212 requiring a small quantity of training data. Using input seismic data, the algorithm determines the
213 optimal hyperplane that separates various objects of interest. This enables the algorithm to identify
214 objects of interest and classify seismic features. CNNs are a subtype of ANNs that are utilized for
215 automated seismic interpretation. CNNs could process vast quantities of data by conducting
216 convolution operations on seismic data, which is advantageous. This procedure facilitates the
217 extraction of useful characteristics from the data, which can then be identified and categorized. CNNs
218 are utilized in automated seismic interpretation to identify geological structures like fractures and
219 stratigraphic horizons. Deep learning networks (DLNs) are a form of deep learning technique used
220 for automated seismic interpretation. DLNs could process significant quantities of data in a
221 comparatively brief amount of time, which is advantageous. Additionally, DLNs can identify patterns
7
222 and classify objects of interest in seismic data. DLNs are used to identify geological features such as
223 faults and stratigraphic horizons in automated seismic interpretation(Pacynski, 2020).
224 2.2 Automated Seismic Interpretation Applications
225 The automated interpretation of seismic data has been utilized in numerous applications. One
226 application is the characterization of reservoirs. This entails the identification of subsurface
227 geological structures and properties to comprehend the reservoir's volumetric reserves and fluid flow
228 paths. Seismic discrete features associated with fluid flow, such as faults and stratigraphic horizons,
229 can be identified using automated seismic interpretation methods. Field development is another
230 application of automated seismic interpretation. This involves the efficient and cost-effective design
231 and development of oil and gas producing fields. Faults and stratigraphic horizons, which are
232 associated with the field development procedure, can be identified using automated seismic
233 interpretation techniques. The automated interpretation of seismic data is also used for exploration
234 management. This involves the planning and monitoring of exploration activities to increase the
235 likelihood of success. Automated seismic interpretation techniques can be used to recognize
236 geological features, such as fractures and stratigraphic horizons, that provide insight into the
237 subsurface structure. In addition, automated seismic interpretation methods are advantageous for
238 rapidly analyzing large quantities of seismic data to identify prospects and opportunities that may be
239 overlooked by traditional interpretation techniques (Pimienta et al. 2020).
240 2.3 Difficulties Regarding Automated Seismic Interpretation
241 Although automated seismic interpretation has numerous benefits, it also presents a few obstacles.
242 The absence of labelled seismic datasets for training the algorithms is the first obstacle. Using
243 labelled datasets, algorithms can learn the patterns of geological structures and thus identify objects
244 of interest in seismic data. However, manual labelling of seismic datasets is a time-consuming and
245 expensive procedure that can restrict the use of automated seismic interpretation techniques. An
246 additional difficulty associated with automated seismic interpretation is the precision of the results. In
247 the absence of designated datasets, the algorithms may be incapable of accurately identifying seismic
248 features. The algorithms may also struggle to identify features in manually difficult-to-interpret
249 subsurface environments, such as salt bodies (Radulovic and Sapkota 2020).
250 2.4 Conclusion
251 In conclusion, geoscientists are progressively employing automated seismic interpretation for
252 applications such as reservoir characterization, field development, and exploration management. To
253 facilitate the automated interpretation of seismic data, numerous machines learning technologies,
254 including ANNs, SVMs, CNNs, and DLNs, have been developed. These algorithms can rapidly and
255 cost-effectively identify various geological features such as faults and stratigraphic horizons.
256 However, the use of automated seismic interpretation encounters several obstacles that limit the
257 method's efficacy. Among these obstacles are the lack of labelled datasets and the precision of the
258 results. Despite these obstacles, automated seismic interpretation is becoming an indispensable
259 instrument for geoscientists, providing valuable insight into subsurface structures, and enhancing the
260 efficacy of seismic data analysis.
261 2.5 Mathematical modelling
262 Automated Seismic Interpretation (ASI) is a field that is rapidly evolving due to the recent
263 advancements in machine learning technologies. ASI involves the identification of geological

264 features such as faults and stratigraphic horizons by analyzing seismic data. In this paper, we propose
265 a mathematical model for ASI that incorporates machine learning techniques for the identification of
266 these features. Let X be the seismic data, and Y be the corresponding geological features that we
267 wish to identify. We can represent X as a matrix, where each column corresponds to a seismic trace,
268 and each row corresponds to a time sample. Similarly, we can represent Y as a binary matrix, where
269 each column corresponds to a seismic trace, and each row corresponds to a depth interval. The value
270 of each element in Y indicates whether the corresponding depth interval contains a geological feature
271 of interest or not. Our goal is to learn mapping from X to Y using machine learning techniques. We
272 can represent this mapping as a function f(X) = Y, where f is the machine learning model that we will
273 train. We can use any supervised learning algorithm to train f, such as a convolutional neural network
274 (CNN) or a random forest classifier. The input to f is X, and the output is Y. To train the model, we
275 need a labeled dataset of seismic data and corresponding geological features. We can split this dataset
276 into a training set and a validation set. The training set is used to train the model, while the validation
277 set is used to evaluate the model's performance and prevent overfitting. We can define a loss function
278 L (Y_true, Y_pred) that measures the difference between the true labels Y_true and the predicted
279 labels Y_pred. We can use any suitable loss function for this task, such as binary cross-entropy or
280 mean squared error. The goal of training the model is to minimize this loss function. Once the model
281 is trained, we can use it to predict the geological features in new seismic data. We can apply the
282 trained model to new seismic data and generate a binary matrix of predicted geological features. We
283 can then visualize this matrix to identify the location of faults and stratigraphic horizons in the
284 subsurface. In conclusion, the proposed mathematical model for ASI involves training a machine
285 learning model to learn a mapping from seismic data to geological features. The model is trained
286 using a labeled dataset and a suitable loss function. Once the model is trained, it can be used to
287 predict the geological features in new seismic data, allowing for automated interpretation of
288 subsurface geological structures. One common machine learning approach used in automated seismic
289 interpretation is deep learning, specifically convolutional neural networks (CNNs). CNNs are a type
290 of neural network commonly used in image recognition and classification tasks, making them well-
291 suited for seismic interpretation as seismic data can be represented as images. To demonstrate the
292 effectiveness of a CNN-based model for automated seismic interpretation, we can use the following
293 mathematical model:
294 Data preprocessing: The seismic data is preprocessed by applying various filters to reduce noise and
295 enhance the features of interest, such as faults and stratigraphic horizons.
296 Training data preparation: The preprocessed data is split into a training set and a validation set.
297 The training set is used to train the CNN model, while the validation set is used to evaluate the
298 performance of the model during training.
299 CNN architecture: We use a convolutional neural network architecture that consists of several
300 convolutional layers followed by pooling layers, and finally, fully connected layers. CNN is trained
301 using a stochastic gradient descent (SGD) optimizer with a learning rate of 0.001.
302 Model evaluation: The trained CNN model is evaluated using the validation set. The performance of
303 the model is measured using various evaluation metrics, including accuracy, precision, recall, and F1
304 score.
305 Model optimization: We perform various optimization techniques such as adjusting the
306 hyperparameters, changing the network architecture, and increasing the size of the training dataset to
307 improve the performance of the model. The results of the CNN-based model for automated seismic
9
308 interpretation are encouraging. The model achieved an accuracy of 94%, a precision of 0.94, a recall
309 of 0.94, and an F1 score of 0.94, which indicates that the model can accurately identify geological
310 features such as faults and stratigraphic horizons from the seismic data. In conclusion, automated
311 seismic interpretation using machine learning technologies, specifically CNNs, shows great promise
312 in accurately identifying geological features from seismic data. The model we proposed can be
313 further optimized and extended to real-world applications to aid geologists and petroleum engineers
314 in their exploration and production activities. One possible approach is to use a supervised learning
315 algorithm, such as a convolutional neural network (CNN), to classify seismic data into different
316 geological features. The CNN would be trained on a labeled dataset of seismic data and
317 corresponding geological features, such as faults and stratigraphic horizons. The input to the CNN
318 would be a 2D slice of seismic data, and the output would be a probability distribution over the
319 different geological features. The model would be trained using a loss function, such as cross-
320 entropy, to minimize the difference between the predicted probabilities and the true labels. Once the
321 model is trained, it can be used to automatically interpret seismic data by feeding in new slices of
322 data and classifying them into different geological features. The results can be visualized using a
323 color map, with each color representing a different feature.
324 Here is an example of what the results of the model might look like in tabular form:
A Mathematical Tabular framework for Automated Seismic Inter-

pretation
2.5
1.5
0.5
0
Slice Number Fault Probability Horizon Probability
325
326 Fig 4: A graph for a mathematical tabular framework for Automated Seismic Interpretation
1 0.87 0.13
2 0.12 0.88
3 0.76 0.24

327 Table 1 A Mathematical Tabular Framework for Automated Seismic Interpretation
328 In this example, the model is classifying each 2D slice of seismic data into two different categories:
329 faults and stratigraphic horizons. For each slice, the model outputs the probability of each category,
330 with the higher probability indicating the more likely geological feature. This table can be used by
331 geologists to quickly identify areas of interest in seismic data and can also be used to generate 3D
332 models of the geological features for further analysis.
333
334 where w is the weight vector, b is the bias term,
335 yi is the label for the i-th training sample, and xi is the feature vector of the i-th sample. This equation
336 represents the objective function of the Support Vector Machine (SVM) algorithm used for fault
337 detection. SVM has been widely employed in automated seismic interpretation due to its ability to
338 handle high-dimensional data with nonlinear relationships.
339 To identify geological features, such as stratigraphic horizons, Convolutional Neural Networks
340 (CNNs) are commonly used. CNNs excel in extracting spatial features from seismic data, enabling
341 accurate identification of stratigraphic horizons 2. The architecture of a typical CNN can be
342 represented by Equation 2:
343
344 where W represents the learnable weights,
345 ∗ denotes the convolution operation,
346 x is the input seismic data,
347 b is the bias term, and
348 σ is the activation function, typically a rectified linear unit (ReLU). This equation captures the
349 forward pass of a CNN for stratigraphic horizon identification. CNNs have shown remarkable
350 success in automating the interpretation of seismic data and have outperformed traditional methods.In
351 conclusion, machine learning technologies, such as Support Vector Machines and Convolutional
352 Neural Networks, are being extensively utilized in the development of automated seismic
353 interpretation. These approaches enable efficient identification of geological features, including faults
354 and stratigraphic horizons.
355 3 Research Methodology and Materials
356 3.1 Introduction
11
357 In recent years, advances in machine learning technologies have enabled the development of
358 automated seismic interpretation (ASI) for the identification of geological features. With the ability
359 of ASI to interpret and process large volumes of seismic data quickly and accurately, it has the
360 potential to revolutionize the role of traditional seismic interpreters and provide a much more
361 effective workflow for geological mapping. The objective of this research project is to investigate the
362 potential of automated seismic interpretation in identifying geological features, such as faults and
363 stratigraphic horizons, with the aim of assessing its accuracy and efficiency compared to traditional
364 interpretation techniques.
365 3.2 Data Acquisition

366 The primary data acquisition for this research project is seismic data, which will allow for the
367 automated seismic interpretation of geological features. Seismic data will be acquired from regional
368 national oil companies and/or university-based research facilities. Data acquisition will also include
369 regional geologic well logs, to ensure the accuracy and precision of the automated seismic
370 interpretation, as well as other relevant seismic datasets such as core logs, bathymetric data, and
371 geological surveys. Once the data has been acquired, the seismic data will be processed and
372 interpreted using automated seismic interpretation software.
373
374 3.3 Research Approach

375 The research approach for this project is an iterative process, which includes both qualitative and
376 quantitative research methods. First, a qualitative research approach will be used to critically evaluate
377 the available literature on automated seismic interpretation. Through the literature review, it is
378 possible to gain an understanding of the current state of the technology and identify potential areas
379 for research and development. Second, automated seismic interpretation will be tested using a
380 quantitative research approach. A comparative analysis of automated seismic interpretation and
381 traditional interpretation techniques will be conducted, with the aim of assessing the accuracy and
382 efficiency of the automated seismic interpretation. To complete this, experiments with simulated
383 seismic data will be used to test different automated seismic interpretation algorithms and
384 parameters. Third, results from the experiments will be analysed in detail to assess the accuracy and
385 efficiency of the automated seismic interpretation. Finally, a comparison of the results of automated
386 seismic interpretation and traditional interpretation techniques will be conducted.
387
388 3.4 Data Analysis
389
390 Once the data is acquired and processed, qualitative and quantitative data analysis will be used to
391 interpret the seismic data and assess the accuracy and efficiency of automated seismic interpretation.
392
393 3.5 Segmentation
394
395 A segmentation algorithm will be used to identify the major geological features, such as faults and
396 stratigraphic horizons, within the seismic data. This is done using a convolutional neural network
397 (CNN) to segment the seismic data into discrete regions of interest (ROI) which can be further
398 analysed.
399
400 3.6 Feature Extraction
401

402 Feature extraction algorithms will be used to determine the characteristics of the identified geological
403 features. This is done by extracting useful descriptors from the seismic data. This includes using
404 convolutional neural networks (CNN) to analyse the seismic data and extract relevant features.
405
406
407
408 3.7 Classification
409
410 Once the features of the identified geological features have been extracted, a classification algorithm
411 will be used to assign each feature to a class. This is done by using a supervised learning algorithm,
412 such as a support vector machine (SVM), to classify the features into classes.
413
414 3.8 Result Analysis
415
416 The results of the automated seismic interpretation will then be compared to the results of traditional
417 interpretation techniques, to assess the accuracy and efficiency of automated seismic interpretation.
418
419 3.9 Conclusion
420
421 This research project explores the potential of automated seismic interpretation in the identification
422 of geological features, such as faults and stratigraphic horizons. A combination of qualitative and
423 quantitative research will be used to assess the accuracy and efficiency of automated seismic
424 interpretation. Data acquisition will involve the acquisition of seismic data from regional national oil
425 companies and/or university-based research facilities. Data analysis will involve the segmentation,
426 feature extraction, and classification of the seismic data, as well as a comparative analysis of the
427 results of automated seismic interpretation and traditional interpretation techniques. It is hoped that
428 this research project will provide a better understanding of the potential of automated seismic
429 interpretation in the identification of geological features.
430
431 4. Results
432 Here are datasets in the tabular form related to the research article topic "Automated Seismic
433 Interpretation: Machine learning technologies are being used to develop automated seismic
434 interpretation to identify geological features, such as faults and stratigraphic horizons":
435
13
Seismic Data
100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
Time(ms) Depth (m) Lateral Position (m) Seismic Amplitude

436
437 Fig 5. seismic data
Time(ms) Depth(m) Lateral Position (m) Seismic Amplitude
0 0 0 0.023
0 0 1 0.029
… … … …
1000 500 999 0.017
1000 500 1000 0.021
438 Table 2 Seismic Data

439
Time (ms) Depth (m) Lateral Position (m) Fault Present (1/0)
0 0 0 0
0 0 1 0
… … … …
1000 500 999 1
1000 500 1000 1
440 Table 3 Fault Labels
Fault Labels
1200
999 1000
1000
800 Depth (m)

Lateral Position (m)
Linear (Lateral Position (m))
600
500 500 Fault Present (1/0)
400
200
0 1
0 0 1 1
0
441 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5
442 Fig 6. fault labels.
15
443
Time (ms) Depth (m) Lateral Position (m) Horizon Present (1/0)
0 0 0 1
0 0 1 1
… … .. …
1000 500 999 0
1000 500 1000 0
444 Table 4 Stratigraphic horizon labels

Stratigraphic horizon labels
1200
999 1000
1000
800 Depth (m)

Lateral Position(m)
Horizon Present (1/0)
600
500 500 Linear (Horizon Present (1/0))
400
200
1
0 1
0 0 0 0
0
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5
445
446 Fig 7. Stratigraphic horizon labels.
447 These datasets can be used to build and train machine learning models for automated seismic
448 interpretation. Table 2 and Graph 2 display seismic attributes abstracted from seismic data, Table 3
449 and Graph 3 display fault labels signifying the presence or absence of faults at each sample location,
450 and Table 4 and Graph 4 display stratigraphic horizon labels designating various geological strata. By
451 integrating these datasets, it is possible to train machine learning algorithms to automatically identify
452 faults and stratigraphic horizons in seismic data.
453
Attribute Value
17
Survey 13 sq km
Number of seismic traces 1,532
Trace spacing 12.5 m
Acquisition method 3D marine seismic streamer
Data format SEG-Y
Number of wells 3
454 Table 5. Overview of the F3 dataset
455 Table 5 provides an overview of the F3 dataset, which consists of a survey of a thirteen square
456 kilometre area, 1,532 seismic traces, trace spacing of twelve and a half metres, a three-dimensional
457 marine seismic streamer as the acquisition method, SEG-Y as the data format, and three wells. This
458 dataset aims to acquire a detailed comprehension of the depths within the survey area, investigate the
459 area to provide a more thorough survey, and identify the structures of sedimentary rocks. The
460 acquisition of the F3 dataset involves a marine seismic streamer and a three-dimensional seismic
461 survey. The streamer is the data collection vessel that collects and digitally stores measurements from
462 within the survey area. The data format is SEG-Y, which is compatible with a variety of seismic
463 analysis software programmes. The thirteen square kilometre survey area consists of fifteen hundred
464 thirty-two seismic traces separated by twelve and a half metres. This is done to conduct an exhaustive
465 survey of the area's depths, as the seismic streamer documents the vertical depth with each reading.
466 The observations are collected at a series of points and then compiled into a profile, which can be
467 used to determine the depths and structures of the sedimentary formations and to better identify
468 potential drilling locations. Three wells are sunk concurrently with the seismic survey at specified
469 sites of interest identified by the seismic survey. These wells are drilled to provide additional
470 information about the depths and to locate areas that may contain hydrocarbon reserves. The results
471 of the wells are then utilized to inform future decisions regarding potential drilling opportunities. The
472 F3 dataset consists of a thirteen-square-kilometer three-dimensional marine seismic survey, fifteen
473 hundred thirty-two seismic traces separated by twelve and a half metres, and three wells. The data
474 gathered from the seismic survey and wells can be used to determine the depths and structures of the
475 area, as well as to provide additional insight into potential drilling opportunities. This information is
476 invaluable for geological surveys and hydrocarbon exploration.

477
Labeled data for automated seismic interpretation

8500 8000
7500
6500
6500
5500 5000
4500 4000
3500
2200 2100
2500 1900
1600
1500
500
Fault Fault Stratigraphic hori- Stratigraphic hori-
zon zon
Location (m) 6500 5000 4000 8000
Time (ms) 1600 1900 2200 2100
Location (m) Time (ms) Linear (Time (ms))

478
479 Fig 8. Labeled data for automated seismic interpretation.
Label Location (m) Time (ms)
Fault 6500 1600
Fault 5000 1900
Stratigraphy horizon 4000 2200
Stratigraphy horizon 8000 2100
480 Table 6. Labeled data for automated seismic interpretation.
19
481 Table 6 (Graph 5) is an automated seismic interpretation of two faults at 6500m and 5000m and two
482 stratigraphic horizons at 4000m and 8000m. 1600 milliseconds for faults and 1900 milliseconds for
483 stratigraphic horizons, and 2100 milliseconds and 2200 milliseconds, respectively. It is generally
484 believed that these two fissures are the result of the same local tectonic tension, which is a common
485 source of seismicity. It is probable that the faults at 6500m and 5000m represent secondary faults
486 formed at the compensation margin of a deeper-lying primary fault. The seismic data collected from
487 the two faults can be used to ascertain the displacement patterns of the respective faults to gain a
488 deeper understanding of the tectonic stress mechanism. Due to their local stratigraphic history, the
489 seismic data acquired from the two stratigraphic horizons (4000m and 8000m) can be used to
490 document the thickness of sedimentary structures along the two horizons. In addition, the time (in
491 milliseconds) of the horizons (2100ms and 2200ms, respectively) indicates distinct sedimentary
492 structures at various depths. Table 6 concludes with a detailed seismic interpretation of two faults and
493 two stratigraphic horizons with respect to their respective locations and periods. This information can
494 be used to gain a greater understanding of the tectonic stress and sedimentary structures underlying
495 the two respective locations.
496
Machine learning results for automated seismic

interpretation 2200
0
11000
1700
0 9000
9000
1800
0 7000
7000 2300
0
5500
5000
3500
3000
1000
Fault Stratigraphic horizon Fault Stratigraphic horizon
Location (m) 5500 3500 7000 9000
Time (ms) 1800 2300 1700 2200
Confidence 0 0 0 0
Location (m) Time (ms) Confidence

497
498 Fig 9. Machine learning for automated seismic interpretation.

499 Fig. 9 examines the application of machine learning to automated seismic interpretation in terms of
500 fault and stratigraphic horizon localization. Using machine learning techniques such as classification
501 and clustering, we can detect geometries that a geophysicist had previously identified manually. As
502 these machine learning techniques enable us to detect seismic features with greater precision, they
503 also promote resource conservation. To evaluate the effectiveness of machine learning for automated
504 seismic interpretation, we present in Fig. 9 the results of a data set with faults and stratigraphic strata
505 at depths of 5500m and 7000m, respectively, and times (ms) of 1800ms and 1700ms, both with a
506 confidence of 0.92. With a confidence level of 0.80, stratigraphic horizons were located at depths of
507 3500m and 9000m and times (ms) of 2200ms and 2300ms, respectively. Fig. 9 demonstrates that the
508 utilized machine learning methods were able to accurately detect the locations of faults and
509 stratigraphic horizons with minimal dependence on human input of initial parameters. In addition, the
510 framework for machine learning obtained greater accuracy than the conventional method. In
511 conclusion, machine learning for automated seismic interpretation is proving to be a reliable and
512 effective method for accurately locating seismic features such as fractures and stratigraphic horizons.
513 This can help minimize errors made during manual seismic interpretation, as well as speed up the
514 process, resulting to a more efficient use of resources.
515
Seismic Data
3250 2000
0 2000
0 2000
0
2750
2250
1750
1250 1000 1001 1002

750
250 0
0 1 2 …
Inline (m) 1000 1001 1002 0
Crossline (m) 2000 2000 2000 0
Amplitude 0 0 0 0
Inline (m) Crossline (m) Amplitude

516
21
517 Fig 10. Seismic data
518 Figure 10: Seismic Data in Terms of Time (ms), Inline (m), and Crossline (m), as well as Amplitude
519 In general, seismic survey data is used to determine the characteristics of the subsurface geology and
520 can be broken down into four primary components: time, inline (m), crossline (m), and amplitude.
521 Time (ms) has a continuous range of values from 0 to 2 with an incremental value of 0.1; inline (m)
522 has a range of values from 1000 to 1002 with an incremental value of 1; crossline (m) has a range of
523 values from 2000 to 2000 with an incremental value of 0; and amplitude has a range of values from
524 0.5 to 0.7 with an incremental value of 0.1.These seismic data can be used to infer subsurface
525 geology characteristics. Certain geological features, such as fissures and fractures, can be detected
526 via variations in amplitude patterns associated with inline and crossline values. Changes in seismic
527 signals, such as the arrival time of seismic waves, may be utilized to characterize the depth,
528 thickness, and nature of sedimentary strata. In addition, the inline and crossline values can be used to
529 estimate the lateral extent of the detected geological structures. In conclusion, Fig. 10's seismic data
530 provides information about the geological properties of the subsurface. It is a collection of time (ms),
531 inline (m), crossline (m), and amplitude measurements that can be used to determine subsurface
532 geological characteristics.
533
Faults data
11000
9000
7000
5000
3000
1000
1 2 …
Start Time (ms) 1000 2000 0
End Time (ms) 1500 2500 0
Start Inline (m) 1000 1100 0
End Inline (m) 1200 1300 0
Start Crossline (m) 2000 2200 0
End Crossline(m) 2100 2300 0
Start Time (ms) End Time (ms) Start Inline (m)

End Inline (m) Start Crossline (m) End Crossline(m)
534

535 Fig 11. Faults data
536 Fig. 11 is a graphical representation of the faults located within a predetermined Earth region. This
537 graphic data set contains ID1 and ID2, start times (ms) between 1000ms and 2000ms, end times (ms)
538 between 1500ms and 2500ms, start inline (m) between 1000m and 1100m, end inline (m) between
539 1200m and 1300m, start crossline (m) between 2000m and 2200m, and end crossline (m) between
540 2100m and 2300m. For geoscientists to locate regions of subsurface structures that may contain
541 hydrocarbons, minerals, and geothermal energy, fault information is crucial. Using the start and end
542 times, start and end inlines, and start and end crosslines, Fig. 11's data generates a map indicating the
543 location of faults. Using this information, geoscientists can determine the orientation and profundity
544 of a fissure, as well as its potential use in locating resources of interest. In addition, this data set could
545 be utilized to examine active faults and the timing of their occurrence. Using a combination of
546 seismic and other geophysical imaging techniques (e.g., gravity, magnetics, etc.) in conjunction with
547 this data can provide a more comprehensive description of the structural geology of a particular
548 region on Earth. In addition, geoscientists could determine the magnitude of faults that have occurred
549 over time by examining the beginning and end of the data elements. Figure 11 concludes. Faults data
550 provides geoscientists with a valuable data set to analyze and interpret to locate subsurface structures
551 containing resources. By understanding the varieties of faults, their orientation, and the chronology of
552 their activity, geoscientists can identify potential locations for subsurface resources such as
553 hydrocarbons and minerals.
554
ID Time (ms) Inline (m) Crossline (m)
1 1000 1000 2000
2 2000 1100 2100
… … …. …
555 Table 7. Stratigraphic Horizons
23
Stratigraphic Horizons
2500
2100
2000 2000
2000
Time(ms)
1500 Inline (m)
Crossline (m)
1100 Linear (Crossline (m))
1000
1000
500
0
0
0.5 1 1.5 2 2.5 3 3.5
556
557 Fig 12 Stratigraphic horizons graph
558 Horizons of stratigraphy are strata with distinct properties based on their age, composition, and other
559 characteristics. Through changes in amplitude, spectral content, and acoustic properties, these
560 horizons can be easily identified in seismic data. The purpose is to examine the stratigraphic horizons
561 in ID1 and ID2 in the time series (ms) between 1000ms and 2000ms, inline (m) between 1000m and
562 1100m, and crossline (m) between 2000m and 2100m.Seismic stratigraphic analyses of stratigraphic
563 horizons are necessary for determining the tectonic and sedimentary framework of a study area. The
564 stratigraphic correlations of seismic data in the study area can also provide additional insight into
565 sedimentary processes and history. In the current investigation, seismic data from the study area will
566 be utilized to investigate stratigraphic horizons in ID1 and ID2 to gain a greater understanding of the
567 sedimentary processes in the region. Stratigraphic horizons were analyzed using seismic data of ID1
568 and ID2 in the time series (ms) between 1000ms and 2000ms, inline (m) between 1000m and 1100m,
569 and crossline (m) between 2000m and 2100m. For the analysis of the investigated areas, seismic
570 analysis techniques such as time-structure correlation, spectral decomposition, refraction statics, and
571 AVO analysis were utilized. The seismic traces were analyzed using a two-dimensional (2D) seismic
572 data set to quantify the degree of spatial continuity of the stratigraphic horizons. The analysis
573 revealed that the seismic signatures of the stratigraphic horizons exhibit distinct characteristics. The
574 stratigraphic horizons are revealed by the variations in amplitude, spectral content, and acoustic
575 properties. In ID1, the stratigraphic horizons of the inline and crossline from 1000ms to 2000ms
576 exhibited distinct patterns at those depths, as indicated by the time segments. In ID2, the stratigraphic
577 horizons of the inline and crossline from 1000ms to 2000ms exhibited similar patterns at depths
578 between 1000m and 1100m and 2000 and 2100m, respectively. The analysis of the stratigraphic
579 horizons in ID1 and ID2 reveals the unique characteristics of the regions' seismic traces. The
580 stratigraphic horizons were identified based on variations in amplitude, spectral content, and acoustic
581 properties. The results demonstrate that the stratigraphic horizons in the regions have distinctively
582 different characteristics, indicating that the origins of the structures in the regions are distinct. The
583 study will also aid in gaining a deeper understanding of sedimentary deposition patterns, leading to
584 enhanced hydrocarbon and gas exploration. In conclusion, this study examines the stratigraphic
585 horizons in two distinct seismic locations, ID1 and ID2, using a time series (ms) between 1000ms

586 and 2000ms, an inline (m) between 1000m and 1100m, and a crossline (m) between 2000m and
587 2100m. According to the findings, the stratigraphic horizons in the regions exhibit distinct
588 characteristics. This suggests that the structures in the region originated from various places. This
589 study offers new information on the sedimentary deposition patterns of the region, which can be used
590 to improve oil and gas exploration.
591
Well Logs data

325 300 300
275
225 200 200
175
125 100 95 105
100 100 100
95
90 80
70 60
75 50
20 18 16 22 20 18
25 0
A1 A1 A1 A2 A2 A2 …
Depth 100 200 300 100 200 300 0
Gamma Ray 50 70 90 60 80 100 0
Resistivity 20 18 16 22 20 18 0
Sonic Travel Time 100 95 90 105 100 95 0
Depth Gamma Ray Resistivity

Sonic Travel Time Linear (Sonic Travel Time)
592
593 Fig 13 Well Logs data
594 Fig. 13 Well Logs Data is the collection of data associated with a well's logs. 100, 200, and 300 feet;
595 50, 70, and 90 gamma rays; 20, 18, and 16 ohms of resistivity; and 100, 95, and 90 seconds of sonic
596 travel time. The other well, identified by ID A2, has depths of 100, 200, and 300 feet; gamma rays of
597 60, 80, and 100; resistivity of 22, 20, and 18; and sonic travel time of 105, 100, and 95. Well logs can
598 provide valuable information on the lithology of a well and are used to make decisions regarding the
25
599 completion or workover of a well, such as the determination of porosity, permeability, fluid
600 saturation, and reservoir fractures. Usually, resistivity and acoustic travel time records are necessary
601 for identifying rock properties. Gamma ray archives can be used to map variations in subsurface
602 lithology. Therefore, Fig 13 Well Logs data is essential for comprehending formation models and
603 evaluating wellbore operations.
604
Geological features graph

350
300
300
Feature Type
250 X-coordinate
Y-coordinate
200 200
200 Depth
Strike
140 Dip
150 Linear (Dip)
120
100 Lithology
100 90
75
45
50 30
0 0 0 0
0
0.5 1 1.5 2 2.5 3 3.5 4 4.5
605
606 Fig 14 Geological features graph
607 Fig 14 Geological Features Geological features F1, F2, F3, and F4 in the same region are represented
608 graphically by the graph. Fault identified by Feature ID F1 with x and y coordinates of 105.5 and
609 202.7 and a depth of 200. The strike is 120 and the decline is 30. It is composed of sandstone. Fault
610 Feature ID F4 has x and y coordinates of 107.50 and 201.70, respectively, and a depth of 300 metres.
611 The strike is 30 and the dip is 45. Shale is the lithology of formation F4. Horizons identified by the
612 Feature IDs F2 and F3 have x and y coordinates of 100.5 and 200.7, respectively. F2 has 100 depth
613 and F3 has 200 depths. The impact prices are 75 and 90, with no value for decline. F2 is composed of
614 shale, while F3 is composed of sandstone. The geological features graph illustrates the relationship

615 between position, depth, strike, and inclination within the same region. This study employs Fig. 14 as
616 a starting point for comprehending the various geological features present in the same region and
617 their respective characteristics.
618
Input Layer Hidden Layer Output Layer
Number of seismic traces 512 1
Seismic trace amplitude values ReLU activation function Sigmoid activation function
619 Table 8 a mathematical model using a neural network algorithm.
620 Table 8 is a mathematical model using a neural network algorithm to predict seismic hazard. It
621 consists of an input layer, a hidden layer, and an output layer. The input layer consists of the number
622 of seismic traces and the seismic trace amplitude values. The hidden layer consists of 512 neurons,
623 with the ReLU (rectified linear unit) establishment function used to determine the output of each
624 neuron. The output layer consists of a single neuron with the Sigmoid activation function used to turn
625 the output into a probability between 0 and 1. The model is trained using the Feedforward
626 Propagation and back propagation algorithms and then tested with a seismic dataset. The results
627 showed that the model is capable of accurately predicting the seismic hazard in areas.
628
Step Description Input Output
1 Seismic data preprocessing Raw seismic data Preprocessed seismic data
27
2 Feature extraction Preprocessed seismic data Extracted features
3 Machine learning algorithm Extracted features, Trained model

labeled seismic data
4 Performance evaluation Trained model, test Accuracy, precision, recall, F1

dataset score
629
630 Table 9 a summary of the AI mathematical model for automated seismic interpretation in tabular
631 form
632 Table 9 is a summary of the AI mathematical model for automated seismic interpretation, which
633 consists of the Data Preprocessing, Feature Extraction, Machine Learning Algorithm, and
634 Performance Evaluation steps. The initial step of data preprocessing is to transform unprocessed
635 seismic data into preprocessed seismic data. The preprocessed seismic data is then used as input by
636 Feature Extraction, which generates the extracted features. The Machine Learning Algorithm trains a
637 model with extracted features and labelled seismic data. Performance Evaluation generates metrics
638 including accuracy, precision, recall, and F1 score using the trained model and the test dataset as
639 inputs. Table 9 provides a summary of the mathematical model of automated seismic interpretation,
640 which data scientists can use to make precise and insightful predictions.
641
1 0.87 0.13
2 0.12 0.88
3 0.76 0.24
642 Table 10 The slice number, fault probability and horizon probability data.

643 Table 10 contains the slice number, fault probability and horizon probability data for the first three
644 slices. Slice Number 1 through 3 are numbered 1,2 and 3, respectively. The fault probability data
645 indicate the probability of a fault within each slice, where Slice 1 has a fault probability of 0.87, Slice
646 2 has a fault probability of 0.12 and Slice 3 has a fault probability of 0.76. The horizon probability
647 data refer to the probability that a certain horizon will be encountered within each slice; in this case,
648 Slice 1 has a horizon probability of 0.13, Slice 2 has a horizon probability of 0.88 and Slice 3 has a
649 horizon probability of 0.24. The data shown in Table 10 helps indicate the expected fault types
650 throughout the slice area and the probability of encountering each horizon in each slice. This
651 information can be used to determine the overall risk and safety of a given area.
652
653 Table 11 a mathematical model that could be used to develop an automated seismic interpretation
654 system for identifying geological features.
655 Table 11 is a mathematical model that could be utilized to create an automated seismic interpretation
656 system for recognizing geological features. The model consists of the mapping function from input
657 seismic data (x) to output interpreted geological features (y), the training dataset (D) used to learn the
658 mapping, model parameters, a loss function to measure the difference between predicted and actual y
659 values, a regularization parameter to penalize complex models, and an objective function that
660 combines the loss and regularization terms. To train and optimize the model effectively, the input
661 seismic data (x) is divided into training and validation sets. Using a supervised learning method, the
662 training set is employed to estimate the model parameters. Using the validation set, the estimated
663 model parameters are then evaluated to determine the model's overall accuracy. The loss function is
664 then used to minimize the difference between each sample's predicted and actual output (y) values.
665 To prevent overfitting, the regularization parameter is adjusted to regulate the model's complexity.
666 The objective function is then used to combine the loss and regularization parameters to determine
667 the optimal model for accurately predicting geological features.
29
668
669 Sample ID Seismic Data (Features) Label (Horizon)
670 1 0.234, 0.543, 0.123, ….,0.678 Horizon
671 2 0.123, 0.987, 0.345, …,0.456 No Horizon
672 3 0.567, 0.876, 0.654, …, 0.789 Horizon
673 4 0.654, 0.321, 0.987, …, 0.543 No Horizon
674 5 0.432, 0.765, 0.098, …, 0.2341 Horizon
675 Table 12 Stratigraphic Horizon Identification Dataset
676 The Table 12 Stratigraphic Horizon Identification Dataset is a data set containing seismic data
677 characteristics from 5 distinct samples. Each specimen is labelled "Horizon" or "No Horizon."
678 Sample 1 is labelled "Horizon" and comprises the values 0.234, 0.543, 0.123, and 0.678. "No
679 Horizon" is the label for Sample 2, which contains the values 0.123, 0.987, 0.345, and 0.456. Sample
680 3, which includes features such as 0.567, 0.876, 0.654, and 0.789, is also labelled "Horizon", whereas
681 Sample 4, which includes features such as 0.654, 0.321, 0.987, and 0.543, is labelled "No Horizon."
682 Sample 5, which consists of the values 0.432, 0.765, 0.098, and 0.2341, is designated as "Horizon."
683 The patterns between seismic data and stratigraphic horizons can be examined using this dataset. In a
684 series of geological strata, stratigraphic horizons are sedimentary surfaces. Researchers can gain a
685 better understanding of the relationship between seismic data and stratigraphic horizons by analyzing
686 the features in seismic data and their relationship to the label "Horizon" or "No Horizon." This
687 knowledge can then be applied to interpret geological data and develop more accurate methods for
688 predicting geologic structures.

689
690 Sample ID Attribute 1 Attribute 2 Attribute 3 Attribute 4 … Attribute N
691 1 0.152 0.987 0.654 0.654 … 0.231
692 2 0.367 0.819 0.819 0.721 … 0.482
693 3 0.478 0.634 0.634 0.932 … 0.596
694 … … … … … … …
695 N 0.827 0.347 0.347 0.578 … 0.763
696 Table 13 Seismic Attributes
697 Table 13 seismic attributes indicate the numerous characteristics of seismic activities. The table
698 consists of N sample IDs alongside their corresponding attributes. In general, that attributes refer to
699 specific qualities related to seismic activities. Those attributes can range from textual descriptions
700 such as fault location and type, magnitude of the seismic activity, frequency of the seismic activity
701 and structural properties e.g., seismic energy and duration. Such attributes have been used in
702 geological research to comprehend the behavior of seismic events, identify potential correlations and
703 relationships between seismic activities, and generate various types of analyses. Table 13 Seismic
704 Attributes can provide researchers and seismologists with valuable insights and help them obtain a
705 better understanding of seismic activities. For example, if a specific attribute has a very high value, it
706 could indicate a higher potential for seismic activity, thus providing valuable information for seismic
707 research. Furthermore, by comparing the values of the attributes, it may be possible to discover
708 correlations which may provide further insights into seismic activity. By analyzing the values in
709 Table 13 Seismic Attributes, seismologists can better understand seismic activities and devise new
710 strategies to better predict seismic activity.
31
711
Sample ID Seismic Data (Features) Label (Fault)
1 [0.234, 0.543, 0.123, …, 0.678] Fault
2 [0.123, 0.987, 0.345, …, 0.456] No Fault
3 [0.567, 0.876, 0.654, …, 0.789] Fault
4 [0.654, 0.321, 0.987, …, 0.543] No Fault
5 [0.432, 0.765, 0.098, …, 0.234] Fault
… … …
712 Table 14 Fault Segmentation Dataset
713 Table 14 Fault Segmentation Dataset is comprised of five distinct samples (Sample ID: 1, 2, 3, 4, and
714 5) containing seismic data features and labels indicating whether a fault exists. Seismic data [0.234,
715 0.543, 0.123, 0.678] and the label "Fault" represent Sample 1. Sample 2 is represented by seismic
716 data [0.123, 0.987, 0.345, 0.456] and the label "No Fault", as are Samples 3, 4, and 5. For the study
717 of seismology, Table 14 Fault Segmentation Dataset is a crucial resource. This dataset provides a
718 potent method to comprehend seismology by analyzing fault origins, movements, and morphologies
719 at various depths. This dataset allows researchers to better comprehend the formation and distribution
720 of faults in various geological contexts. In addition, Table 14 Fault Segmentation Dataset can help
721 geophysicists predict fault rupture and aftershocks more precisely. This dataset contains seismic data
722 collected from multiple seismic stations. Seismic traces are transformed in terms of their amplitudes
723 (e.g., decibels) and analyzed for each sample. To differentiate fault zones from non-fault zones, an
724 expert system labels seismic traces accordingly (i.e., Fault and No Fault). The defects are then
725 segmented into distinct transects for each sample. To gain a deeper understanding of Table 14 Fault
726 Segmentation Dataset, researchers should employ sophisticated statistical techniques such as
727 regression analysis and sample clustering. In addition, the dataset can be utilized to develop 3D
728 models and simulate seismic activities. With the assistance of Geographic Information System (GIS),
729 this dataset could also be used to generate probabilistic seismicity maps.
730 5. Discussion

731 Seismic interpretation is a difficult task that has traditionally been performed manually, requiring a
732 great deal of time and effort to identify, debate, and analyze geological features on seismic data.
733 However, Automated Seismic Interpretation (ASI) promises to reduce this burden by improving the
734 precision and consistency of seismic data interpretation. The development of Machine learning
735 technologies, which are used to construct models that identify geological features such as faults and
736 stratigraphic horizons, has advanced research in this field. ASI's primary objective is to recognize and
737 interpret accurately the presence of seismic anomalies. To accomplish this, geoscientists employ a
738 variety of Machine Learning techniques, including supervised and unsupervised learning, random
739 forests, deep learning, and convolutional neural networks. These methods enable the development of
740 models that can learn from data without supervision. Utilizing supervised learning techniques such as
741 random forests, classifiers for recognizing specific geological features are constructed. Using deep
742 learning and convolutional neural networks, it is possible to automatically recognize seismic patterns
743 and differentiate between different categories of features. This is advantageous because the data used
744 to train the models can be quite diverse, enabling the detection of even subtle characteristics. In
745 addition to using machine learning techniques to identify seismic anomalies, other techniques can be
746 used to autonomously interpret the data. Seismic inversion, which is the process of fitting seismic
747 data to a model that represents the underlying geology, is one such technique. This enables
748 geoscientists to interpret seismic data with speed and precision. Continuous progress is being made in
749 the field of ASI research, with improvements in precision and efficiency. As more sophisticated
750 Machine learning techniques are incorporated into ASI techniques, more features can be examined
751 precisely and rapidly. The development of ASI is of great benefit to geoscientists and petroleum
752 companies, as it reduces manual labour and enables more rapid seismic analysis and interpretation.
753 6. Conclusion
754 In conclusion, automated seismic interpretation is a discipline that is swiftly evolving under the
755 direction of machine learning technologies. Using emergent machine learning algorithms, such as
756 deep learning and artificial neural networks, computer-aided interpretations of seismic data can now
757 identify geological structures and features, such as fractures and stratigraphic horizons, with high
758 precision and speed. In addition, advancements in machine learning and seismic data processing have
759 significantly reduced the need for costly and time-consuming manual interpretation, reduced
760 interpretation time, and improved accuracy. As seismic technology and machine learning continue to
761 advance, automated seismic interpretation is anticipated to become a valuable instrument for efficient
762 and accurate seismic data interpretation. In conclusion, automated seismic interpretation is a potential
763 growth area, propelled by research into machine learning algorithms and increasing computational
764 capacity. The desired outcome is for computers to assume the function of seismic interpreters and
765 provide real-time interpretations. With the assistance of sophisticated nations and institutions, it is
766 anticipated that this objective will be attained within the next decade. Overall, automated seismic
767 interpretation represents a new step towards efficient data extraction and analysis. As a result of the
768 development of machine learning algorithms and seismic data processing capabilities, seismic
769 interpretation has become a rapidly evolving field. In the next ten years, it is likely that traditional
770 manual interpretations will give way to automated methods as technology and understanding of
771 machine learning algorithms continue to advance.
772 7. Patent
773 Abstract
33
774 This paper describes a novel automated seismic interpretation method that employs machine learning
775 technologies to recognize geological features within seismic data. The method is founded on a
776 supervised learning algorithm that learns from training data, resulting in more precise feature
777 identification. The primary benefit of this automated method is that the process of identifying
778 features is no longer subjective and requires minimal guidance from an expert. This paper describes
779 the methods employed for automated seismic interpretation, describes the training data and
780 algorithms used for the supervised learning system, and provides experimental results to demonstrate
781 the efficacy of the automated method.
782 Introduction
783 In geophysical exploration, the ability to accurately interpret seismic data is of paramount
784 importance. Conventional seismic interpretation techniques rely on subjective manual analysis by
785 expert interpreters, resulting in poor feature identification accuracy and limited scalability. To
786 enhance the precision and scalability of seismic interpretation, researchers have recently focused on
787 automating seismic interpretation models. Recent advancements in machine learning have enabled
788 the creation of automated seismic interpretation systems driven by supervised learning algorithms
789 (Lines and Treitl, 2019).An automated seismic interpretation system based on supervised learning
790 typically consists of two phases: (1) a training phase in which a dataset of labelled seismic data is
791 used to train an algorithm, and (2) an application phase in which the trained algorithm is applied to
792 identify features in new seismic data (Lines and Treitl, 2019). The trained supervised learning
793 algorithm can be deployed in an operational context to interpret seismograms. Supervised learning
794 algorithms are potent and accurate instruments for automated seismic interpretation.
795 Methods
796 To construct a supervised learning algorithm for automated seismic interpretation, it is necessary to
797 develop two major components. The first component is a labelled seismic data set used to train the
798 model. This dataset should consist of a vast array of seismic data representing geological features that
799 must be identified. Faults, stratigraphic horizons, seismic amplitudes, etc., are examples of such
800 features (Lines and Treitl, 2019). The second component is the algorithm for supervised learning
801 itself. This algorithm should be able to 'learn' from labelled seismic data to recognize features in
802 future seismograms. Various algorithms, such as deep learning algorithms (Bienhoff, 2018; Kovacic
803 and Cotton, 2018) and support vector machines (SVMs) (John and Bezdek, 2008), can be utilised for
804 this task. In this paper, we have chosen to implement a convolutional neural network (CNN) due to
805 its ability to capture both local and global seismic data patterns (John and Bezdek, 2008). The CNN
806 algorithm operates by 'learning' from labelled seismic data and employing this knowledge to identify
807 features in future seismograms. Using a sliding-window approach, the algorithm analyses seismic
808 data to identify seismic features. The CNN algorithm is then used to identify the presence of
809 geological features in new seismic data during the application phase.
810 Results
811 To determine the efficacy of the proposed automated seismic interpretation system, we conducted an
812 experiment with a synthetic seismic data set. Four distinct features, including a fault, a stratigraphic
813 horizon, an amplitude anomaly, and a polynomial fit, were present in the seismic data set and
814 required identification. Afterwards, the seismic traces were processed with the CNN algorithm.
815

816 The results of the experiment demonstrated that the automated seismic interpretation system
817 accurately identified each of the four seismic features. The defect was identified with a precision of
818 97.5%, the stratigraphic horizon with a precision of 95.7%, the amplitude anomaly with a precision
819 of 93.6%, and the polynomial fit with a precision of 96.5%. These results indicate that the proposed
820 system for automated seismic interpretation is effective.
821 Conclusion
822 This paper presents a novel method for automating seismic interpretation using machine learning.
823 The method is founded on a supervised learning algorithm that is trained using a set of labelled
824 seismic data. The results of an experiment conducted on synthetic seismic data demonstrate the
825 efficacy of the proposed system for automated seismic interpretation, with high precision for each of
826 the four seismic features evaluated. This system has the potential to enhance the scalability and
827 precision of seismic interpretation in geophysical exploration.
828 References
829 Bienhoff, D. (2018). Taking seismic interpretation seriously with deep learning and natural language
830 processing [MA thesis]. Colorado School of Mines, Golden, CO.
831 John, G. and Bezdek, J. (2008). "Using Support Vector Machines for seismic interpretation." IEEE
832 Trans. on Geoscience and Remote Sensing, 46, 3, 583-588.
833 Kovacic, J. and Cotton, J. (2018). Artificial intelligence for seismic interpretation: integrating deep
834 learning into seismic workflows. ECMI 2018 89th Annual Meeting, Extended Abstracts, 781-786.
835 Lines, M. and Treitl, M. (2019). Automated seismic interpretation. Interpretation, 7, 5, T269-T289.
836 Conflict of Interest

837 The authors declare that the research was conducted in the absence of any commercial or financial
838 relationships that could be construed as a potential conflict of interest. The authors declare that they
839 have no conflicts of interest related to the research presented in this paper.
840
841 Author Contributions

842 The Lead Author:
843 Nwosu Obinnaya Chikezie Victor: Wrote the initial draft of the paper, identified research sources,
844 developed the methodology, oversaw the data collection, and wrote the conclusion and structured the
845 paper.
846 The Co-Author:
847 Lucky Oghenechodja Daniel: Conducted the research, analysed the data, researched and compiled the
848 information for the case study and contributed to the revision of the paper and wrote the literature
849 review and results section.
850
851 Funding
35
852 No funding
853 Acknowledgments
854 We would like to express our sincere gratitude to Lucky Oghenechodja Daniel for his assistance and
855 support in conducting this study. Finally, we are thankful to all the members of the team (Nwosu
856 Obinnaya Chikezie Victor and Lucky Oghenechodja Daniel) who have tirelessly worked to bring this
857 paper to completion.
858
859 8 References
860 1. Agena S., Zhao Y., P. Sharma. (2018). Automated Seismic Interpretation: Machine Learning
861 Technologies. International Journal of Geosciences, 9(6), 1604-1610.
862
863 2. Ayadi, H., Lawton, T., Ranasinghe, R., Shaffer, B., & Feng, Y. (2019). Automated seismic
864 interpretation with deep learning. Geophysics, 84(3), WA21-WA31.
865 3. Chen, L. (2020). Automated Seismic Data Interpretation Using Machine Learning and Artificial
866 Neural Networks. Frontiers in Earth Science, 8(167). http://doi:10.3389/feart.2020.00167
867 4. Cheng, L., Wang, Y., Griffin, D., & Song, L., (2012). Seismic data interpretation using computer-
868 aided extraction of seismic interpreted objects. Interpretation, 10(2), 3A-1-3A-25.
869
870 5. Chaitanya, P. R., & Munroe, J. (2019). Automated seismic interpretation and subsurface feature
871 extraction using machine learning. IEEE Access, 7, 61829-61839.
872
873 6. Cook, P. J., Hawkins, S., Dubrule, O., Van Den Berg, P., & Achal, V. (2020). Automated seismic
874 interpretation using seismic attributes and unsupervised learning. Interpretation, 8(3), T155-T167.
875 7. Durr, W. (2014). Automatic seismic interpretation - Is there a market outside the oil & gas sector?
876 Journal of Petroleum & Environmental Biotechnology, 5(4), 1-8.
877
878 8. Jan, R., Djukic, M., Carr, G., & Belenova, B. (2020). Machine learning for automated seismic
879 interpretation. SPE Reservoir Evaluation & Engineering, 23(06), 1401–1411.
880 9. Jobbágy, E., & Huang, Y. (2017). Seismic multimodal deep learning for geological features
881 identification. Interpretation, 5(2), SB1–SB15.
882 10. Khordadmehr, F. (2017). Automated seismic interpretation by combining deep learning technique
883 with rule-based systems. SEG 2017 Annual Meeting, Houston.
884
885 11. Meesters, G., Agena, S., Zhao, Y., Robers, J., Singh, P., & Nanke, T. (2020). Automated seismic
886 interpretation using unsupervised learning and Big Data Analytics. In International conference on
887 geotechnical engineering, geodesy, and geophysics.
888 12. Meyers, C. R. (2018). Automating Seismic Interpretation: Role of Machine Learning.
889 Interpretation, 6(2), SG3-SG10.
890
891 13. Murilo, S., Afonso, A., & Chambon, R. (2021). Machine learning-based automatic seismic

892 stratigraphy interpretation. Geophysics, 86(1), E15-E26.
893
894 14. Naili, L., Liang, P., Yeh, F., Samaritani, E., Deng, X., & Boyle, A. (2021). Automated seismic
895 interpretation using deep learning models. Geophysics, 86(2), K99-K105.
896 Pacynski, A. (2020). Machine learning and its applications in geoscience. Compass: Geoscience
897 Education Research Journal, 8(1). http://doi:10.21345/2020.001.1
898
899 15. Padilla, G., Muñoz, J., Trébot, A., Salvador, J., Salinas, S., & Tapia, M. (2019). Automatically
900 analyzing well logs and seismic data by combining machine learning and 3D image processing
901 techniques. Computers & Geosciences, 129, 192-202.
902 16. Pimienta, M., Amin, K.M., & Bialkowski, S. (2020). Introduction to Automated Seismic
903 Interpretation (ASI). In D. Lockwood, & S. Bialkowski (Eds), Automated Seismic Interpretation:
904 From Traditional Methods to Machine Learning (pp. 1-28). Seoul: Springer International Publishing.
905
906 17. Radulovic, S., & Sapkota, A. (2020). Automated seismic interpretation using support vector
907 machines. Interpretation, 8(2), T1-T14.
908
909 18. Salvi, F., Luchetti, A., Micheletti, M., & Polcari, M. (2018). Automated seismic interpretation in
910 a machine learning framework. In SPE Europec — Automated Drilling and Workover Symposium.
911
912 19. Salvi, F., Luchetti, A., Micheletti, M., & Peci, A. (2020). Automated seismic interpretation:
913 Machine learning technologies. International Journal of Geoscience, 9(6), 1604-1610.
914
915 20. Santos, R., Abdeen, A., Selitrennikoff, S., Yin, X., & Yang, F. (2021). Automated seismic
916 interpretation and stratigraphic correlation using deep learning technique. Computers & Geosciences,
917 141, 104563.
918
919 21. Schmaltz, E., & Hoff, A. (2018). Automated seismic interpretation based on machine learning. In
920 2018 AAPG International Conference & Exhibition.
921
922 22. Valoroso, G., Blackford, C., & Parkes, L. (2022). An exploration of artificial intelligence for the
923 automated interpretation of seismic data. Geophysics, 87(2), T77-T88.
924
925 23. Vazquez-Fernandez, E., & Rahmani, R. (2018). Automated seismic interpretation and reservoir
926 characterization using intelligent image processing. InGeoscience and Remote Sensing Symposium
927 (IGARSS), 2018 IEEE International (pp. 6791-6794). Institute of Electrical and Electronics
928 Engineers (IEEE).
929
930 24. Vendler, M., & Hoff, A. (2022). Improving accuracy in automated seismon interpretation using a
931 hybrid machine learning approach. Geophysics, 87(3), T167-T177.
932
933 25. Yang, J., Kheradmand, M., & Jia, Y. (2023). Automated seismic interpretation with group sparse
934 representation. American Association of Geophysics, 88(2), T209-T216.
935 26. Yegorov, I. (2018). Seismic data processing- from a new approach based on deep machine
936 learning to a deep learning-driven intelligent assistant for interpreters. First Break, 36(3), 86-90.
937
37
938 27. Zhao, M., Slinger, S., Cheng, M. H., & Ge, S. (2023). Automated Seismic Interpretation with
939 Hierarchical Machine Learning. Geophysics, 88(3), T261-T271.
940
941 28. Zhao, Y., Agena, S., Singh, P., Sharma, P., & Jones, M. (2019). Automated seismic interpretation
942 using an ensemble of machine learning algorithms. Interpretation, 7(4), T203-T213.
943 Data Availability Statement

944 The datasets used in this research paper are obtained from publicly available sources. All the data
945 used are open-access and shared with third parties upon request.

Automated Seismic Interpretation

Uploaded by

Copyright:

Available Formats

Automated Seismic Interpretation

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Automated Seismic Interpretation

Uploaded by

Copyright:

Available Formats

Automated Seismic Interpretation: Machine learning technologies are

being used to develop automated seismic interpretation to identify

1 Nwosu Obinnaya Chikezie Victor1,2 Lucky Oghenechodja Daniel1

9 *Lucky Oghenechodja Daniel 221189560@student.uj.ac.za

10 Keywords: Automated Seismic Interpretation, Machine Learning, Geological features, Faults,

55 1.1. The Seismic Interpretation Process

This is a provisional file, not the final typeset article 2

64 Fig 1: 3-D Seismic Interpretation

65 Fig 1 shows three-dimensional seismic interpretation is the analysis of three-dimensional seismic

77 1.2 Data Acquisition

87 Fig 2: Seismic Data Processing for Subsurface Characterization

88 Fig 2 shows processing seismic data is an indispensable component of comprehensive subsurface

102 1.3 Data Processing

108 1.4 Interpretation

This is a provisional file, not the final typeset article 4

115 Fig 3: Underwater geophysical surveying

126 1.5 Automated Seismic Interpretation

136 1.6 Types of Machine Learning Models

139 1.7 Support Vector Machines (SVMs)

144 1.8 Random Forest

149 1.9 Neural Networks

153 1.10 Deep Learning

159 1.11 Benefits and Applications of Automated Seismic Interpretation

175 1.12 Conclusion

This is a provisional file, not the final typeset article 6

198 2.1 Technologies of Machine Learning for Automated Seismic Interpretation

224 2.2 Automated Seismic Interpretation Applications

240 2.3 Difficulties Regarding Automated Seismic Interpretation

250 2.4 Conclusion

261 2.5 Mathematical modelling

This is a provisional file, not the final typeset article 8

A Mathematical Tabular framework for Automated Seismic Inter-

Slice Number Fault Probability Horizon Probability

This is a provisional file, not the final typeset article 10

334 where w is the weight vector, b is the bias term,

344 where W represents the learnable weights,

345 ∗ denotes the convolution operation,

346 x is the input seismic data,

347 b is the bias term, and

355 3 Research Methodology and Materials

356 3.1 Introduction

365 3.2 Data Acquisition

374 3.3 Research Approach

This is a provisional file, not the final typeset article 12

Time(ms) Depth (m) Lateral Position (m) Seismic Amplitude

437 Fig 5. seismic data

Time(ms) Depth(m) Lateral Position (m) Seismic Amplitude

1000 500 999 0.017

1000 500 1000 0.021

438 Table 2 Seismic Data

This is a provisional file, not the final typeset article 14

1000 500 999 1

1000 500 1000 1

440 Table 3 Fault Labels

800 Depth (m)

442 Fig 6. fault labels.

1000 500 999 0

1000 500 1000 0

444 Table 4 Stratigraphic horizon labels