Description:Field of the invention

The present invention relates, generally, to seismic data processing sub- systems employed in exploration decision systems. More particularly, the present invention relates to a deep learning based system and method for performing automatic moveout tracking of seismic data for providing an efficient and accurate exploration decision system.

Background of the invention

In oil and gas exploration, seismic data plays a pivotal role in guiding exploration decisions as they aid in evaluating potential drilling locations, reducing uncertainties, and increase the likelihood of success. One key component of seismic data processing is seismic imaging which is a multi-step workflow where moveout correction serves as a critical step. Moveout correction is a technique that aligns seismic reflections from sub-surface layers recorded at different receivers, that appear at varying times based on their distance from the seismic sources (offset). Also, moveouts are a critical input to sub-surface modelling.

Conventionally, manual techniques are used to construct sub-surface velocity models for geological analysis, which are required for conducting drilling activities in oil rig well-sites. Some of these techniques include Normal Moveout (NMO) tracking, inversion-based nonstationary NMO correction method, dynamic time warping etc., which are used for constructing the sub-surface velocity models for geological analysis. Typically, seismic waves recorded by geophones at varying source-receiver distances exhibit a characteristic moveout behavior due to travel time of the waves through sub-surface layers. The moveout pattern is influenced by velocity characteristics of sub-surface structures that provide information about geological formations. In order to track the moveout, traditional methods typically employ coherence-based measures that identify path with highest similarity (or coherence) across seismic traces recorded at different distances. The path with highest coherence is assumed to correspond to the normal moveout curve and its analysis is used to compute stacking velocity, which is a key parameter for imaging sub-surface structures. Further, the imaging is interpreted to deduce sub-surface geology. However, the existing techniques often face challenges when faced with complex seismic data characteristics including noise, multiple reflections, and overlapping events.

Further, it has been observed that conventional methods typically offer lower resolution, which leads to inaccuracies in identification of correct moveout path. Moreover, seismic data often contains interference from unwanted noise and multiple reflections that distort primary seismic events. In such conditions, coherence measures become ambiguous, as they identify paths that do not accurately represent the true moveout of primary events. Also, such ambiguities result in errors in estimation of stacking velocities, which in turn affects accuracy of sub-surface models and imaging. The inaccuracies in sub-surface models and imaging result in improper deduction of geological information, which further results in improper execution of drilling activities at oil rig sites causing environmental hazards and costing heavy execution price.

Yet further, in existing systems, the coherence measures fail to differentiate between different types of events leading to ambiguous seismic interpretations. Therefore, manual intervention becomes necessary to ensure accurate NMO tracking, which increases execution time. In this effort, seismic processing analysts often need to manually trace moveout paths to correct for the ambiguities introduced by coherence-based methods. This manual process is both time-consuming and prone to human error, especially when handling large datasets and / or complex seismic environments. This indicates a clear gap in the current seismic processing workflow and an absence of robust tools capable of detecting and resolving moveout anomalies efficiently. Therefore, there is a need for enhancements in quality and turnaround time of seismic data processing as it directly impacts efficiency and accuracy of exploration decision systems and processes, leading to more agile and informed outcomes.

In light of the aforementioned drawbacks, there is a need for a deep learning based system and method for performing automatic moveout tracking of seismic data for providing an efficient and accurate exploration decision system. Also, there is a need for a system and a method for analyzing seismic traces in real-time, detecting irregularities and anomalies, and providing corrected moveout paths, allowing for consistent and reliable stacking velocity calculations. Further, there is a need for a low-cost system and method for increasing overall efficiency and reliability of seismic data interpretation by offering higher accuracy, minimizing time required for moveout analysis and subsequent sub-surface imaging and interpretation as well as lowering environmental hazard, thereby ensuring drilling activities are carried out at appropriate oil rig sites.

Summary of the Invention

In various embodiments of the present invention, an automatic moveout tracking system (100) for performing automatic moveout tracking of seismic data is provided. The automatic moveout tracking system (100) comprises a processor (110) configured to execute an automatic moveout tracking engine (104) to identify seismic traces from the seismic data fetched from a sensor data collection unit (102). The identified seismic traces are grouped to create seismic gathers to obtain time window data. Each time window of the time window data comprises of multiple seismic traces corresponding to one or more seismic events observed at a specific time frame. The automatic moveout tracking system (100) generates semantic features from the seismic traces, employing a trained deep learning model (122), to predict relative temporal position of the one or more seismic events in the seismic traces within each of the time window relative to a specified reference seismic trace within each of the time window or a separately received reference seismic trace. The automatic moveout tracking system (100) generates a normal moveout, by the trained deep learning model (122), for each of the time window using the predicted relative temporal positions of the one or more seismic events in the seismic traces. The automatic moveout tracking system (100) is integrable with exploration decision systems for use in operations carried out by the exploration decision systems in sub-surface formations.

In various embodiments of the present invention, a method for performing automatic moveout tracking of seismic data is provided. The method comprises a processor (110) configured to identify seismic traces from the seismic data. The identified seismic traces are grouped to create seismic gathers to obtain time window data. Each time window of the time window data comprises of multiple seismic traces corresponding to one or more seismic events observed at a specific time frame. Semantic features are generated from the seismic traces, employing a trained deep learning model (122), to predict relative temporal position of the one or more seismic events in the seismic traces within each of the time window relative to a specified reference seismic trace within each of the time window or a separately received reference seismic trace. A normal moveout is automatically generated, by the trained deep learning model (122), for each of the time window using the predicted relative temporal positions of the one or more seismic events in the seismic traces.

Brief description of the accompanying drawings

The present invention is described by way of embodiments illustrated in the accompanying drawings wherein:
FIG. 1 is a block diagram of an automatic moveout tracking system for sub-surface imaging, in accordance with an embodiment of the present invention;

FIG. 1a illustrates a model for seismic data acquisition and a sample normal moveout, in accordance with an embodiment of the present invention;

FIG. 1b illustrates two stages of a deep learning model employed for moveout tracking, in accordance with an embodiment of the present invention;

FIG. 1c illustrates continuous trace wise updating of parameters of a discriminative correlation filter during moveout prediction, in accordance with an embodiment of the present invention;

FIG. 1d illustrates steps of designing a deep learning model, in accordance with an embodiment of the present invention;

FIG. 1e illustrates feature map attention module and squeeze-and-excitation network, in accordance with an embodiment of the present invention;

FIG. 1f illustrates rectified linear activation function and a sigmoid activation function, in accordance with an embodiment of the present invention;

FIG. 1g illustrates steps to train the deep neural model, in accordance with an embodiment of the present invention;

FIG. 1h illustrates the design of the training dataset, in accordance with an embodiment of the present invention

FIG. 2 is a flowchart illustrating a method for automatic moveout tracking for sub-surface imaging, in accordance with an embodiment of the present invention;

FIG. 3 illustrates predicted normal moveout of a sample time window, in accordance with an embodiment of the present invention;

FIG. 4 illustrates various time windows cropped from a real seismic gather and the predicted normal moveouts in accordance with an embodiment of the present invention;

FIG. 5 illustrates an architecture of a siamese neural network, in accordance with an embodiment of the present invention;

FIG. 6 illustrates a sample depth domain interval 2-dimensional velocity model, in accordance with an embodiment of the present invention;

FIG. 7 illustrates positions of receivers and sources in a simulated acquisition geometry, in accordance with an embodiment of the present invention;

FIG. 8 illustrates a time sampled ricker wavelet used as a source wavelet, in accordance with an embodiment of the present invention;

FIG. 9 illustrates sample generated seismic traces for deep neural network model training, in accordance with an embodiment of the present invention;

FIG. 10 illustrates a sample common midpoint offset sorted seismic gather and its corresponding velocity spectrum, in accordance with an embodiment of the present invention;

FIG. 11 illustrates a sample time-domain interval velocity model corresponding to a depth domain velocity model, in accordance with an embodiment of the present invention;

FIG. 12 illustrates a sample time window, in accordance with an embodiment of the present invention;

FIG. 13 illustrates seismic traces in a training sample, in accordance with an embodiment of the present invention; and

FIG. 14 illustrates a training and validation loss graph, in accordance with an embodiment of the present invention.

Detailed description of the invention

In various embodiments of the present invention, a system for automatic moveout tracking of seismic events using neural networks is provided. The system aids in ensuring that drilling activities are carried out at appropriate sites, thereby minimizing the extent of environmental hazard caused due to unnecessary drilling at improper sites. The system aids in generating sub-surface images from processed sensor data collected from sensor data collection units which are employed for seismic data interpretation and drilling site selection. The system generates semantic features from the collected sensor data and tracks the semantic features resulting in prediction of relative temporal positions of seismic events required for sub-surface imaging. In various exemplary embodiments of the present invention, the automatic moveout tracking system (100) employs a unique deep learning model which is built on an architecture comprising siamese neural network and a discriminative correlation filter, thereby automating moveout tracking process and enhancing the sub-surface imaging. The present invention represents a significant advancement in seismic processing technology, offering far-reaching benefits to exploration decision systems by improving both quality of seismic imaging and agility of exploration workflows.

Exemplary embodiments herein are provided only for illustrative purposes and various modifications will be readily apparent to persons skilled in the art. The underlying principles defined herein may be applied to other embodiments and applications without departing from the spirit and scope of the invention. The terminology and phraseology used herein is for the purpose of describing exemplary embodiments and should not be considered limiting. Thus, the present invention is to be accorded the widest scope encompassing numerous alternatives, modifications and equivalents consistent with the principles and features disclosed herein. For purposes of clarity, details relating to technical material that is known in the technical fields related to the invention have been briefly described or omitted so as not to unnecessarily obscure the present invention.

FIG. 1 is a block diagram of an automatic moveout tracking system (100) for performing automatic moveout tracking of seismic data, in accordance with various embodiments of the present invention. The automatic moveout tracking system (100) is integrable with exploration decision systems for use in various operations carried out by the exploration decision systems in sub-surface formations. In an exemplary embodiment of the present invention, normal moveout generated by the automatic moveout tracking system (100) is used for making accurate seismic data interpretation for various sub-surface operations. In an embodiment of the present invention, the automatic moveout tracking system (100) comprises a sensor data collection unit (102), an automatic moveout tracking engine (104) and a sub-surface imaging unit (114). The automatic moveout tracking engine (104) comprises an extraction unit (106) and a neural network unit (108). In an embodiment of the present invention, the units of the system (100) operate in conjunction with each other and are operated via a processor (110) specifically programmed to execute instructions stored in a memory (112) for executing respective functionalities of the units of the system (100).

In an embodiment of the present invention, the automatic moveout tracking engine (104) communicates with the sensor data collection unit (102) and with the sub-surface imaging unit (114) via a communication channel (not shown). The communication channel (not shown) may include, but is not limited to, a physical transmission medium, such as, a wire, a fiber optic cable or a wireless connection over a multiplexed medium, such as, a radio channel in telecommunications and computer networking. The examples of radio channel in telecommunications and computer networking may include, but are not limited to, a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN) and a campus area network (CAN). In various embodiments of the present invention, the automatic moveout tracking engine (104) performs automatic prediction of relative temporal position of seismic events for faster sub-surface imaging. The automatic moveout tracking engine (104) employs the neural network unit (108) which is uniquely and particularly trained to identify temporal patterns in seismic data that correlate with specific seismic events corresponding to geological structures. In an exemplary embodiment of the present invention, the automatic moveout tracking engine (104) may be implemented in a cloud computing architecture in which data, applications, services, and other resources are stored and delivered through shared datacenters.

In an embodiment of the present invention, the sensor data collection unit (102) includes seismic data received from one or more waveform receivers (receivers 118) positioned at one or locations of sub-surface geological formations of the earth to receive one or more reflected waves from sub-surface earth materials, as shown in FIG. 1a. Waveforms are received by the waveform receivers (118) from one or more waveform sources (sources 120) corresponding to the sub-surface earth materials, as shown in FIG. 1a. The waveform receivers (118) may include, but are not limited to, geophones or hydrophones configured to capture one or more reflected seismic waves. The extraction unit (106) fetches the seismic data from the sensor data collection unit (102) and identifies seismic traces from the sensor data. The extraction unit (106) groups the seismic traces to create seismic gathers, as shown in Fig. 1a. From the seismic gathers, one or more time windows are obtained, as shown in Fig. 1a. Each time window comprises multiple seismic traces corresponding to one or more seismic events observed at a specific time frame. Seismic events are reflection and refraction events of interested sub-surface geological formations. Such reflection signals are received from different earth materials such as rock layers, along with background noise. The time windows may be of an arbitrary length of time.

In an embodiment of the present invention, the neural network unit (108) processes the time window data received from the extraction unit (106) for detection and tracking of seismic events from the seismic traces. In an exemplary embodiment of the present invention, the time window data comprises discrete segments of two or more seismic traces which include data corresponding to seismic events observed within a specific time frame. From each time window, the neural network unit (108) selects one or more seismic traces to serve as a reference seismic trace. The neural network unit (108) designates other seismic traces in the time window as test seismic traces. The neural network unit (108) predicts a relative temporal position of the seismic events in the test seismic traces relative to a specified reference seismic trace within each received time window or a separately received reference seismic trace. Using the reference seismic trace, the neural network unit (108) establishes the time reference of the seismic events against which the relative temporal position of the seismic events is automatically determined in the test seismic traces. The relative temporal positioning of seismic events across multiple seismic traces is crucial for constructing accurate sub-surface models. In particular, the neural network unit (108) determines temporal offset of seismic events between reference seismic trace and the test seismic traces, which indicates how the seismic events detected in the test seismic traces are related to the reference seismic trace in terms of timing. This relative temporal information helps in predicting correct depth and location of the geological layers within the sub-surface. Based on the predicted relative temporal positions, normal moveout is generated as per received time window for all received time windows, as explained in detail herein below.

In an exemplary embodiment of the present invention, the seismic trace containing seismic event of interest is denoted as reference seismic trace (x) and the seismic trace where the related seismic event needs to be temporally positioned is denoted as test seismic trace (z). Temporal position of the seismic event in the reference seismic trace is denoted as ‘y’, which is used as the baseline for comparison. The relative temporal position of the related seismic event in the test seismic trace (z) is denoted as (r). The neural network unit (108) determines (r) given (x,y,z). The neural network unit (108) automatically generates a normal moveout of this time window through prediction of the relative temporal position (r) of seismic events in each seismic trace (z) in the time window, with respect to a reference seismic trace (x). Fig. 1a illustrates predicted relative temporal positions of seismic events and the generated normal moveout, in accordance with an embodiment of the present invention. The neural network unit (108) predicts the relative temporal position (r) using semantic features. Examples of semantic temporal features include one or more learnt features derived from arrival times, signal amplitudes, and frequency content of reflection signals received at the waveform receivers (118). The semantic features are highly discriminative against seismic events from background noise, and aid in distinguishing between seismic signals of interest which are related to sub-surface geological formations and noise or unwanted artifacts that may be present in the seismic data. The semantic features enable the system (100) to perform seismic event detection, determining presence or absence of seismic events, and enabling the system (100) to make accurate predictions about the relative temporal location of the seismic events within the test seismic traces (z).

In various embodiments of the present invention, the neural network unit (108) employs a deep learning model and an online update method to process the time window data and generate a normal moveout for all the received time windows. The deep learning model is a neural network trained on synthetic data and optimized for performing automatic moveout tracking by training on seismic events at one or more travel times received from travel paths through sub-surface, thereby ensuring accuracy of seismic event positioning even when data may have been recorded using varying acquisition geometry. The deep learning model generates semantic features optimized for temporal localization of seismic events, through training on synthetic data, which is explained in the following paragraphs and description of FIGs. 1d and 1h. In an exemplary embodiment of the present invention, the deep learning model comprises of two components viz. a convolutional siamese neural network and a discriminative correlation filter with parameters optimized for accurate discrimination of seismic events in each seismic trace from the background noise and accurate temporal localization of seismic events in the test seismic trace(z) with respect to the reference seismic trace(x), as explained in detail herein below.

In the exemplary embodiment of the present invention, the neural network unit (108) is configured to pass both the reference seismic trace (x) (103 A) and the test seismic trace (z) (103 B) through the siamese neural network to obtain corresponding semantic features. As shown in FIG. 1(b), the neural network unit (108) includes two stages. In a first stage, the siamese neural network (105) generates a first set of semantic features (109A) and a second set of semantic features (109B) using a first set of learnt parameters (105A) that discriminates seismic events from other spurious signals and noise. In a second stage, the discriminative correlation filter (107) uses a second set of learnt parameters (107 A) needed for accurate localization of a seismic event in the test seismic trace (z) (103B) relative to the reference seismic trace (x) (103A), using the semantic features (109 B) of the seismic event in the test seismic trace (z) (103B) and the semantic features (109A) of the seismic event in the reference seismic trace (x) (103A).

In particular, in this exemplary embodiment of the present invention, the discriminative correlation filter (107) takes the semantic features (109A, 109B) and generates a gaussian response map with peak at temporal position corresponding to the seismic event in the reference seismic trace (x) (103A) (referred as reference seismic event (y)). The discriminative correlation filter (107) learns filter weights (w) from the reference seismic trace (x) (103A) and then convolves it with test seismic trace (z) (103 B) to generate the relative temporal position (r) (111) of seismic event in the test seismic trace (z) relative to the reference seismic trace (x). In an example, the discriminative correlation filter (107) takes the semantic features of the reference seismic trace (x) (103A) and the temporal position of the reference seismic event (y) and produces filter weights (w) by solving a ridge regression problem as provided herein below:

min_w ||w*x-y?||?_2^2+?||w?||?_2^2

where ? is the regularisation parameter, * denotes circular convolution. The regression is efficiently computed in fourier domain and the weights (w) are computed as provided herein below:

w=F^(-1) {(F(x)?F^* (y))/(F^* (x)?F(x)+?)}

where ? is the element-wise product, * denotes complex conjugate, F is the Discrete Fourier transform (DFT) and F -1 is the inverse DFT. The relative temporal position (r ) for the test seismic trace (z) is computed as provided herein below:

r=F^(-1) {F^* (w)?F(z)}

FIG. 1c illustrates an online model update method which entails continuous trace wise updating of the parameters (107 A) of the discriminative correlation filter (107) to parameters (107 B) during moveout prediction based on each new received reference seismic trace (103A), in order to adapt to the changing characteristics of the seismic events on the respective reference seismic trace (x) (103A), before prediction of the relative temporal position (r) (111) of the seismic event in the test seismic trace (103B), in accordance with an another embodiment of the present invention. In this embodiment of the present invention, weights of the discriminative correlation filter 107 when tracking a reference seismic trace (x) is Wx. The weights for tracking the reference seismic trace (x + 1) is W*x+1 which is obtained by a linear interpolation equation, W*x+1 = (1 - a)Wx+1 + aWx ,where ’a’ is an interpolation coefficient set to 0.01. Advantageously, through repeated relative temporal position prediction using the deep learning model coupled with the adaptive online update method, normal moveout of the time window is automatically determined. This online fine-tuning capability of the neural network unit (108) allows the deep learning model to continuously provide good results even with change in seismic event characteristics.

In various embodiments of the present invention, as demonstrated above, the neural network unit (108) is trained to process each test seismic trace in parallel with the reference seismic trace, analyzing how the seismic event in the test seismic trace aligns with the seismic event in the reference seismic trace. This alignment is crucial for determining how seismic reflections and other seismic events shift over space and time, thereby enabling the neural network unit (108) to provide an accurate estimate of one or more seismic event's position relative to other traces in the seismic gather. FIG. 9 illustrates sample generated seismic traces for deep neural network model training, in accordance with an embodiment of the present invention. The siamese neural network (105) generates the learnt semantic features for each input seismic trace. Further, the siamese neural network (105) generates feature transformations that differentiate between the seismic event and background noise. The feature transformation provides same/similar output for, for instance, a seismic event A and a seismic event B which is a shifted/scaled version of seismic event A. The feature transformation provides a completely dissimilar output for a different seismic event C / unrelated noise D.

After semantic features are generated, the automatic moveout tracking engine (104) utilizes the extracted semantic features to predict the relative temporal position (r) of seismic events in the test seismic traces. The neural network unit (108) is trained to associate specific semantic features with known seismic event positions and assess temporal offset of seismic events between the reference seismic traces (x) and test seismic traces (z). The neural network unit (108) outputs a predicted relative temporal position (r) for each seismic event in the test seismic trace (z) relative to the reference seismic trace (x) for mapping the seismic event’s location in time. Fig.1a further illustrates the predicted relative temporal positions of seismic events and the generated normal moveout, in accordance with an embodiment of the present invention.

FIG. 1d illustrates the method of development of the deep learning model (122), in accordance with various embodiments of the present invention. The various steps in building the deep learning model 122 comprise:
Step 1: preparing a synthetic training dataset (117) containing data samples of time windows simulated using wavefield propagation through simple depth domain interval velocity models;
Step 2: designing an architecture comprising two stages viz. the siamese neural network and discriminative correlation filter as explained above; and
Step 3: training the deep learning model using the synthetic training dataset (117) such that forward and backward tracking of seismic events are performed for each training data sample. The training is performed using an unsupervised learning model composed of results of forward tracking and backward tracking with minimized consistency loss, without the need for use of training labels i.e. reference information indicating the true temporal position of the seismic events per seismic trace of either the reference seismic trace or the test seismic trace.
Referring to step 2, the siamese neural network (105) is designed using a 1d convolutional layer (124)) with several feature channels, a rectified linear activation function (126) and a feature map attention module using squeeze-and-excitation network (128) as illustrated in FIG. 1e. The squeeze-and-excitation network (128) provides feature map attention through scaling operation performed on each feature channel produced after the linear activation function (126), via a part of learnt parameters (105 A) as illustrated in FIG. 1e. The squeeze-and-excitation network (128) is designed using an adaptive pooling layer (130-1), fully connected layer (130-2), rectified linear activation function (130-3) and a sigmoid activation function (130-4) as illustrated in FIG. 1f. The semantic features (109 A, 109 B, as shown in Fig. 1b) are weighted (through feature attention/channel attention) in terms of their learnt importance to the task of temporal position determination of reference seismic event (y) in the test seismic trace (z) (103 B) relative to the reference seismic trace (x) (103A), through the squeeze-and-excitation network (128).

Referring again to step 2, the discrete correlation filter (107) is designed to perform the steps as illustrated in FIG. 1g to train the deep neural model, in accordance with an embodiment of the present invention, as provided below:
a. solving a regression problem between a reference seismic trace (x) (103A) and a gaussian response map (r1) (113), the gaussian response map (113) having its peak at the temporal location (115A) of a seismic event, which is the actual temporal position of reference seismic event (y), as mentioned above; and
b. generating, as a solution to the regression problem, learnt parameters (107A,107 B) (as shown in Figs. 1c and 1e) which are used to predict the temporal position of reference seismic event (y) (115B) in the test seismic trace (z) (103B), relative to the reference seismic trace (x) (103A).

Referring to step 3, the synthetic training dataset (117) is prepared using multiple training samples (119), as illustrated in FIG. 1h and as demonstrated herein below in accordance with an embodiment of the present invention, by:
a. using simulated wavefield propagation (121) of a simulated wavelet (133) generated by a simulated seismic source positioned at different distances from simulated receivers;
b. selecting depth-domain seismic interval velocity models (123);
c. virtually recording wavefields that are received at the simulated receivers, where each simulated receiver is placed at fixed, regular positions with equal spacing;
d. recording seismic traces at each simulated receiver for different positions of the simulated seismic source, referred to as full trace record (125);
e. grouping the full trace records (125) received by the simulated seismic source - simulated receiver pair by a common midpoint bin between the simulated seismic source and the simulated receiver, and sorted by spatial distance between the same simulated seismic source and the simulated receiver to form a seismic gather (139);
f. deriving time-domain seismic interval velocity models (127) using depth-domain seismic interval velocity models (123), waveform source position (129) and simulated receiver position (131) and characteristics of the waveform wavelet (133), where the wavelet (133) is a ricker wavelet of a single frequency;
g. cropping fixed length time windows (137) starting at the predicted temporal positions (135) of the simulated seismic reflections, per seismic gather (139);
h. selecting three or more seismic traces through random sampling from all the seismic traces in each of the seismic gathers (139);
i. scaling an amplitude of the seismic traces of each training sample (119) to the range [-1,1]; and
j. predicting relative temporal positions (135) of the seismic reflections in the full trace records (125) which are received at the simulated receiver closest to the simulated seismic source, per position of the simulated seismic source, through first order differentiation in time dimension.

As demonstrated in FIG. 1h, the depth-domain interval velocity models (123) are selected through random sampling from a database comprising depth-domain interval velocity models (123) of 2-10 layers of random thickness and a range of interval velocities. The training of the deep learning model 122 (Fig. 1d) takes place via the parameters (105A, 107A, 107B) (FIG. 1c) using several training samples (119) through several iterations to minimize the consistency loss (141) function (FIG. 1g) representation of difference between the location of the seismic event in the first seismic trace of each training data sample (119) and the predicted location (135) of the related seismic event in the first seismic trace. The prediction takes place through forward tracking (143) and backward tracking (145).

During the training phase, relative temporal positions of seismic events may be forward tracked by the neural network unit (108) in subsequent test seismic traces. This generates predicted relative temporal positions on all test seismic traces concerning the seismic trace before them. For instance, forward tracking (143) as illustrated in FIG. 1g takes place by prediction of relative temporal position of the seismic event in a first seismic trace of each training data sample (119) on each subsequent seismic trace of the respective training data sample (119), until the last seismic trace of the respective training data sample (119). Backward tracking (145) is accomplished by prediction of the relative temporal location of the seismic event present at the predicted relative temporal location on the last seismic trace of each training sample (119) obtained through forward tracking (143) on the first seismic trace of the respective training sample (119). That is, the last seismic trace is taken as a reference and moveouts are predicted backwards till zero offset trace. The estimated moveout in the first frame obtained via backward tracking is expected to be identical to its actual moveout (= 0). The difference in forward-tracked moveouts and backward-tracked moveouts is calculated by the neural network unit (108) and the neural network weights are trained to minimize the consistency (141) loss or maximize the consistency of forward- backward moveout trajectory in an unsupervised manner. The parameters (105 A, 107A, 107B) are updated after each iteration such that the consistency loss (141) is minimized.

In an embodiment of the present invention, the sub-surface imaging unit (114) utilizes the predicted relative temporal features output from the neural network unit (108) to construct a 3D sub-surface image. The image is generated by constructing a sub-surface velocity model from the predicted relative temporal positions of seismic events across all test seismic traces in the time window, thereby providing a high-resolution spatial representation of the sub-surface structure. Advantageously, the predicted seismic event positions serve as the foundation for building an accurate sub-surface model, which can be used in downstream processing tasks for geological interpretation and further exploration. In an exemplary embodiment of the present invention, the 3D image generated by the sub-surface imaging unit (114) incorporates information about the depth, composition, and structure of geological formations, making it a valuable tool for oil and gas exploration, mining, and other applications where sub-surface characterization is critical. The accuracy of the relative temporal positioning directly influences the quality of the resulting 3D image, ensuring that geological features are correctly represented in terms of their spatio-temporal relationships.

Advantageously, processing of the time window data by employing the deep learning models, in accordance with various embodiments of the present invention, ensures that the reference seismic trace is representative of the seismic events in time window which is necessary to predict relative temporal positions of one or more seismic events. The deep learning model is trained to produce semantic features within the time window data that correlate with temporal position of seismic events in reference seismic trace. The deep learning model is critical for deducing temporal positioning of seismic events, which in turn provides crucial information about sub-surface structure. Also, the neural network unit (108) may be automatically updated in an online mode which enables continuous refinement of the deep learning model as new data may be added, thereby adapting to changing field conditions or new seismic scenarios. This adaptability is particularly useful in dynamic field conditions where geological features or acquisition techniques may change over time. The online update process involves retraining the neural network unit (108) with new seismic data and incorporating feedback from the system’s (100) performance to refine the prediction algorithms. This ensures that the system (100) remains accurate and reliable even as new seismic events are encountered, providing consistent and up-to-date sub-surface images. Further, advantageously, the neural network unit (108) detects and tracks seismic events from the seismic traces with high accuracy. Also, prediction of temporal events by the neural network unit (108) facilitates construction of accurate three-dimensional (3D) sub-surface images which is crucial for various applications such as, but not limited to, oil and gas exploration, mineral exploration, and geotechnical studies.

In various embodiments of the present invention, the automatic moveout tracking system (100) is connected to exploration decision systems and facilitates the exploration decision systems to make and accurate decisions to identify potential drilling locations. For example, the moveout generated by automatic moveout tracking system (100) enables generation of an initial velocity model automatically, streamlining one of the most labor-intensive phases of seismic processing. This enables closer grid velocity analysis, regardless of increased data volume, leading to more precise models and faster iteration cycles in velocity analysis. In another example, the moveout generated by automatic moveout tracking system (100) introduces the potential for new inversion misfit functions, which can be applied to invert sub-surface properties such as velocity and anisotropy. This opens new avenues for improving the accuracy of seismic inversion workflows, which are used to infer the physical properties of the sub-surface based on seismic data.

In yet another example, the automatic moveout tracking system (100) enables novel Quality Control (QC) methods, such as detecting moveout anomalies and automatically identifying first-order multiples based on computed moveout similarity. This enhances the QC of seismic data, allowing for more efficient identification of processing errors or sub-surface complexities that may impact data quality. In another example, automatic moveout tracking system (100) enhances exploration decisions related to Amplitude Variation with Offset (AVO) analysis, a key technique for identifying fluid types and lithology variations in the sub-surface. By accurately aligning seismic reflections across different offsets, it allows for precise amplitude extraction, decoupling the need for a perfectly accurate velocity model. The automatic moveout tracking system (100) reduces variability across offsets and gathers, leading to more reliable AVO results. These improvements translate into better fluid discrimination and more accurate sub-surface condition predictions, enhancing the value of AVO-based exploration efforts.

In yet another example, in fields where 4D seismic data is used to monitor reservoirs over time, this automatic moveout tracking system (100) enables more accurate and faster detection of sub-surface changes. The automatic moveout tracking system (100) ensures accurate time-lapse data alignment, which is crucial for identifying small yet significant changes in the reservoir over time. These changes could include fluid flow patterns, pressure distribution, and compaction effects, allowing operators to optimize reservoir management and improve enhanced oil recovery techniques. Therefore, the deep learning-based automatic moveout tracking system delivers substantial time savings and quality improvements across the seismic processing workflow. By automating moveout correction, it improves the speed and precision of velocity model building, inversion processes, AVO analysis, and reservoir monitoring. Additionally, it introduces several novel methods within seismic processing, such as new misfit functions, anomaly detection, and automated QC techniques, further enhancing the overall accuracy of seismic data interpretation.

FIG. 2 is a flowchart illustrating a deep learning based method for performing automatic moveout tracking for sub-surface imaging, in accordance with an embodiment of the present invention.

At step 202, seismic traces are identified from sensor data and time window data is generated. In an embodiment of the present invention, the seismic traces are grouped to create seismic gathers. From the seismic gathers, one or more time windows are obtained. Each time window comprises multiple seismic traces corresponding to one or more seismic events observed at a specific time frame. At step 204, the time window data is processed to select one or more seismic traces. In an embodiment of the present invention, from each time window, one or more seismic traces are selected to serve as the reference seismic trace (x), and the other seismic traces are designated in the time window as test seismic traces (z). At step 206, semantic features are extracted from each of the seismic traces within a particular time window. At step 208, relative temporal position of one or more seismic events is predicted in the seismic traces of the time window data relative to a specified reference seismic trace within the time window or a separately received reference seismic trace using semantic features. At step 210, a normal moveout is generated to construct a 3D sub-surface image. In various embodiments of the present invention, the deep learning model is optimized for generating an automatic moveout tracking by training on seismic events at one or more travel times received from one or more travel paths through sub-surface, thereby ensuring accuracy of seismic event positioning even when data may have been recorded using varying acquisition geometry. The generated 3D image includes information about the depth, composition, and structure of geological formations, making it a valuable tool for oil and gas exploration, mining, and other applications where sub-surface characterization is critical. The accuracy of the relative temporal positioning directly influences the quality of the resulting 3D image, thereby ensuring that geological features are correctly represented in terms of their spatio-temporal relationships. Details of steps 202 to 210 are provided in the description provided above with respect to system (100).

FIG. 3 illustrates the predicted normal moveout of a sample time window cropped from a synthetic seismic gather where moveout (a) shows the time window containing seismic traces, each with related seismic events and the predicted relative temporal position in each seismic trace with respect to the first seismic trace, in accordance with an embodiment of the present invention. Moveout (b) shows the moveout correction applied across each seismic trace based on the predicted moveout, where the perfect horizontal alignment shows efficacy of the automatic normal moveout tracking method, in accordance with an embodiment of the present invention.

FIG. 4 illustrates various time windows cropped from a real seismic gather and corresponding predicted normal moveouts, in accordance with an embodiment of the present invention. Time windows (a) illustrate time windows cropped from a seismic gather followed by application of normal moveout correction using different stacking velocities. Time windows (b) show the corresponding predicted normal moveout, showing changing moveout predictions, following actual moveout of each time window.

FIG. 5 illustrates various layers used in the architecture of the siamese neural network for a sample input batch of 1024 and the seismic trace of 44 samples along-with their inter relationships and input-output dimensions at each layer, in accordance with an embodiment of the present invention. Automatic moveout tracking, in accordance with this embodiment of the present invention, has been tested for several time windows containing ‘n > 3’ seismic traces each of 40-time samples length. Time windows of any arbitrary length were supported. Moveout tracking was carried out in batches of time windows, in accordance with this embodiment of the present invention. In this example, batch size was 1024 and in each time window, a first trace was considered as reference seismic trace, with position of seismic event in the trace to be at the zeroth time sample. Further, the seismic trace was considered as the reference seismic trace. The seismic event may be at any position, since prediction of temporal position is relative to the reference seismic trace. The seismic traces except the first seismic trace were considered as test seismic traces. The time window within the batch was padded with zeros at the start and end, to create a temporal length of 44 samples. The batch of padded time windows were given as input to the deep learning model. The deep learning model produced a one-dimensional array for each time window of the entire batch, with each element of the array containing the relative temporal position (in time samples) of the test seismic trace with respect to reference seismic trace. The output features of the siamese neural network for two seismic traces, were then passed through discriminative correlation filter which produced a gaussian response map. The discriminative correlation filter is configured to learn weights through a ridge regression between the semantic features of the reference seismic trace and gaussian response corresponding to the temporal position of the seismic event, as described above. The ridge regression was performed over semantic features of reference seismic trace shifted / translated to all possible temporal positions using circular convolution.

FIG. 6 illustrates a depth domain interval in terms of 2-dimensional layer cake velocity model (3km depth, 9 km lateral extent) containing 10 layers and corresponding velocity profile. As described above, 2-10 layered 2D layer cake depth domain velocity models were forward modelled by the neural network unit (108) through wavefield propagation utility, using mixed acquisition geometry including several sources and receivers. Full length seismic traces obtained from the waveform receivers were sorted in order of common midpoint between waveform sources and waveform receivers and then by order of distance between the waveform source and waveform receiver. The group of seismic traces obtained was in terms of a seismic gather. Fixed length time windows were cropped by the neural network unit (108) from each seismic gather, around times of arrival of primaries. From each time window, three seismic traces were randomly chosen by the neural network unit (108) as data samples used for training the automatic moveout tracking system (100), as shown in FIG. 13. A training dataset of more than 1,16,000 such data samples were used to train the neural network unit (108). Depth domain velocity models with zero dip layers were forward modelled for generating the training dataset. The velocity models were randomly generated with an interval velocity range of 1500 m/s to 6000 m/s with each layer of varying thickness in the range 75 to 100 grid points. A velocity step was randomly chosen from the set {100,200,400,600}. The velocity grid dimensions were 2301 x 750 with a grid spacing of 4m. A total of 90 such models were generated.

FIG. 7 illustrates an acquisition geometry along with a sample depth domain velocity model, used for forward wavefield propagation, in accordance with an embodiment of the present invention. A total of 230 receivers were positioned in a regular spacing of 40m. A total of 115 shots were simulated such that each shot has one active waveform source and the waveform receivers were in recording mode. In order to capture patterns in training data, representative of types of acquisition geometries, a mixed acquisition geometry was used for forward wavefield propagation where the waveform receiver positions were fixed and waveform source moved in regular intervals for each shot. Each waveform source wave is simulated using a 750 time sampled ricker wavelet of 25 Hz dominant frequency as shown in FIG. 8. Forward modelling was carried out using `Deepwave` forward wave propagation library using finite difference method in space and time to step through the 2-Dimensional scalar wave equation, in accordance with an embodiment of the present invention. The waveform receivers were configured to record for a period of 3 seconds with a sampling frequency of 4ms, recording a total of 750 time samples. The recorded seismic traces were sorted by midpoint between the waveform source and waveform receiver positions and distance between the waveform source and waveform receiver. One such sample sorted seismic gathers with absolute value of offset and its corresponding velocity spectrum is illustrated in FIG. 10.

In order to extract data samples containing time windows, the depth-domain velocity model was converted to its corresponding time-domain velocity model, based on the acquisition geometry parameters and forward modelling parameters, in accordance with an embodiment of the present invention. A sample time-domain velocity model corresponding to the depth domain velocity model in FIG. 6 and its velocity gradient is illustrated in FIG. 11. The position of change in the time -domain velocity gradient was used to obtain the times of arrival of primaries in each seismic gather by the neural network unit (108). For each seismic gather, a time window of 40 samples was used around the time of arrival of each primary in the seismic gather. A sample window cropped around a primary at the 271st time sample in a seismic gather is illustrated in FIG. 12. The final data sample in the training dataset comprised of three randomly sampled seismic traces for each time window. FIG. 13 illustrates the seismic traces in a training sample from the training dataset, in accordance with an embodiment of the present invention. A total of 15481 seismic gathers were processed, extracting 10 data samples from each seismic gather, resulting in a dataset of 1,54,810 gathers. Out of these, 75% of the samples were used for training the deep learning model and 25% for model validation and testing.

In an embodiment of the present invention, the deep learning model was trained by the neural network unit (108) through optimization of neural network parameters by minimizing a smooth L1 based consistency loss computed between the forward tracking moveout prediction and backward tracking moveout prediction. In another embodiment of the present invention, a general purpose ‘Adam’ optimizer was used to train the deep learning model for minimizing the consistency loss for 100 epochs with a batch size of 1024 data samples. FIG. 14 illustrates loss curve dynamics for training data and validation data, showing that the deep learning model converges to low loss, in accordance with an embodiment of the present invention. After the training, a deep learning model with accurate moveout tracking capability was obtained.

Advantageously, the automatic moveout tracking system (100), in accordance with various embodiments of the present invention, automates and accelerates traditionally manual and time-intensive process of moveout correction. The system (100) provides significant speed and efficiency benefits across various phases of seismic data processing. The system (100) provides an advanced and highly automated system for the relative temporal positioning of seismic events, leveraging the power of deep learning and neural networks. By automating the selection of reference and test seismic traces, extracting discriminative semantic features, predicting relative temporal positions, and filtering out noise, the system (100) delivers highly accurate results for 3D sub-surface imaging. Furthermore, its ability to be updated ensures that the system (100) can continuously adapt to new data, making it an invaluable tool for modern seismic exploration and sub-surface characterization.

The present invention may be implemented in numerous ways including as a system, a method, or a computer program product such as a computer readable storage medium or a computer network wherein programming instructions are communicated from a remote location.

While the exemplary embodiments of the present invention are described and illustrated herein, it will be appreciated that they are merely illustrative. It will be understood by those skilled in the art that various modifications in form and detail may be made therein without departing from the scope of the invention.
, C , Claims:We Claim:

1. An automatic moveout tracking system (100) for performing automatic moveout tracking of seismic data, the system (100) comprising:
a processor (110) configured to execute an automatic moveout tracking engine (104) to:
identify seismic traces from the seismic data fetched from a sensor data collection unit (102), wherein the identified seismic traces are grouped to create seismic gathers to obtain time window data, each time window of the time window data comprises of multiple seismic traces corresponding to one or more seismic events observed at a specific time frame;
generate semantic features from the seismic traces, employing a trained deep learning model (122), to predict relative temporal position of the one or more seismic events in the seismic traces within each of the time window relative to a specified reference seismic trace within each of the time window or a separately received reference seismic trace; and
generate a normal moveout, by the trained deep learning model (122), for each of the time window using the predicted relative temporal positions of the one or more seismic events in the seismic traces, wherein the automatic moveout tracking system (100) is integrable with exploration decision systems for use in operations carried out by the exploration decision systems in sub-surface formations.

2. The system (100) as claimed in claim 1, wherein the automatic moveout tracking engine (104) comprises a neural network unit (108) that comprises the deep learning model (122) which is trained to process the time window data for:
selecting one or more seismic traces within each of the time window to serve as a reference seismic trace, wherein the time window data represents discrete segments of two or more seismic traces including data corresponding to the seismic events observed within the specific time frame;
designating other seismic traces within each of the time window as test seismic traces; and
determining a temporal position of the seismic events in the reference seismic trace, wherein the deep learning model predicts the relative temporal position of the seismic events in the test seismic trace relative to the reference seismic trace based on the reference seismic trace, test seismic trace and the temporal position of the seismic events.

3. The system (100) as claimed in claim 2, wherein the deep learning model (122) comprises a siamese neural network (105) and a discriminative correlation filter (107) and is trained to pass the reference seismic trace and test seismic trace through two stages, wherein:
in a first stage, the siamese neural network (105) automatically generates the semantic features for each of the reference seismic trace and test seismic trace within each of the time window received such that the reference seismic trace are transformed to a first set of semantic features and the test seismic trace is transformed to a second set of semantic features employing a first set of learnt parameters; and
in a second stage, a discriminative correlation filter (107) automatically localizes the relative temporal position of the seismic events in the test seismic trace relative to the reference seismic trace using the generated first and second set of semantic features employing a second set of learnt parameters,
wherein the generated semantic features include one or more learnt features derived from arrival times, signal amplitudes, and frequency content of reflection signals received at waveform receivers (118) positioned at one or locations of sub-surface geological formations of the earth from one or more waveform sources (120) corresponding to the sub-surface earth materials, and wherein the generated first and second set of semantic features discriminate the seismic events in the respective seismic traces from other spurious signals and noise within the respective seismic trace.

4. The system (100) as claimed in claim 3, wherein the discriminative correlation filter (107) receives the generated first and second set of semantic features and generates a gaussian response map with peak at the temporal position corresponding to the seismic events in the reference seismic trace using filter weights learnt from the reference seismic trace and its convolution with the test seismic trace to generate the relative temporal position of the seismic events in the test seismic trace relative to the reference seismic trace.

5. The system (100) as claimed in claim 3, wherein the deep learning model (122) is updated with each prediction such that parameters of the discriminative correlation filter (107) are updated based on a new received reference seismic trace in order to adapt to changing characteristics of the seismic events on the respective reference seismic trace before prediction of the relative temporal position of the seismic event on the test seismic trace.

6. The system (100) as claimed in claim 1, wherein the deep learning model (122) is developed by:
preparing a synthetic training dataset containing data samples of time windows simulated using wavefield propagation through simple depth domain interval velocity models;
designing an architecture comprising of two stages including siamese neural network and discriminative correlation filter; and
training the deep learning model using the synthetic training dataset such that forward and backward tracking of the seismic events are performed for each training data sample, wherein the training is performed using an unsupervised learning model composed of results of forward tracking and backward tracking with minimized consistency loss, without the need for use of training labels comprising reference information indicating a true temporal position of the seismic events per seismic trace of either the reference seismic trace or the test seismic trace.

7. The system (100) as claimed in claim 1, wherein the synthetic training dataset is prepared using multiple training samples by:
using simulated wavefield propagation of a simulated wavelet generated by a simulated seismic source positioned at different distances from simulated receivers;
selecting depth-domain seismic interval velocity models;
virtually recording wavefields that are received at the simulated receivers, where each simulated receiver is placed at fixed, regular positions with equal spacing;
recording seismic traces at each simulated receiver for different positions of the simulated seismic source as a full trace record;
grouping the full trace records received by the simulated seismic source - simulated receiver pair by a common midpoint bin between the simulated seismic source and the simulated receiver and sorted by spatial distance between the same simulated seismic source and the simulated receiver to form a seismic gather;
deriving time-domain seismic interval velocity models using depth-domain seismic interval velocity models, waveform source position and simulated receiver position and characteristics of the waveform wavelet, where the wavelet is a ricker wavelet of a single frequency;
cropping fixed length time windows starting at the predicted temporal positions of the simulated seismic reflections, per seismic gather;
selecting three or more seismic traces through random sampling from all the seismic traces in each of the seismic gathers;
scaling an amplitude of the seismic traces of each training samples to a range [-1,1]; and
predicting relative temporal positions of the seismic reflections in the full trace records, which is received at the simulated receiver closest to the simulated seismic source, per position of the simulated seismic source, through first order differentiation in time dimension.

8. The system (100) as claimed in claim 6, wherein the deep learning model (122) is trained using the synthetic dataset comprising several training samples employing learning parameters through several iterations to minimize the consistency loss function representation of a difference between location of the seismic events in a first seismic trace of each training data sample and a predicted location of a related seismic event in the first seismic trace predicted through forward tracking and backward tracking.

9. The system (100) as claimed in claim 8, wherein the forward tracking includes prediction of the relative location of the seismic event in the first seismic trace of each of the training data sample on each subsequent seismic trace of the respective training data sample, until the last seismic trace of the respective training data sample, and backward tracking includes prediction of a temporal location of the seismic event present at the predicted temporal location on the last seismic trace of each training sample obtained through forward tracking on the first seismic trace of the respective training sample.

10. The system (100) as claimed in claim 2, wherein the automatic moveout tracking engine (104) comprises a subsurface imaging unit (114) configured to generate a three-dimensional (3D) sub-surface image by constructing a sub-surface velocity model from the predicted relative temporal positions of the seismic events across all test seismic traces in each of the time windows, thereby providing a high-resolution spatial representation of the sub-surface formations, wherein the 3D sub-surface image includes information related to depth, composition, and structure of geological formations.

11. A method for performing automatic moveout tracking of seismic data, the method comprising:
a processor (110) configured to:
identify seismic traces from the seismic data, wherein the identified seismic traces are grouped to create seismic gathers to obtain time window data, each time window of the time window data comprises of multiple seismic traces corresponding to one or more seismic events observed at a specific time frame;
generate semantic features from the seismic traces, employing a trained deep learning model (122), to predict relative temporal position of the one or more seismic events in the seismic traces within each of the time window relative to a specified reference seismic trace within each of the time window or a separately received reference seismic trace; and
generate a normal moveout, by the trained deep learning model (122), for each of the time window using the predicted relative temporal positions of the one or more seismic events in the seismic traces.

12. The method as claimed in claim 11, wherein the deep learning model (122) is trained to process the time window data for:
selecting one or more seismic traces within each of the time window to serve as a reference seismic trace, wherein the time window data represents discrete segments of two or more seismic traces including data corresponding to the seismic events observed within the specific time frame;
designating other seismic traces within each of the time window as test seismic traces; and
determining a temporal position of the seismic events in the reference seismic trace, wherein the deep learning model predicts the relative temporal position of the seismic events in the test seismic trace relative to the reference seismic trace based on the reference seismic trace, test seismic trace and the temporal position of the seismic events.

13. The method as claimed in claim 11, wherein the deep learning model (122) passes the reference seismic trace and test seismic trace through two stages, wherein:
in a first stage, a siamese neural network (122) automatically generates the semantic features for each of the reference seismic trace and test seismic trace within each of the time window received such that the reference seismic trace are transformed to a first set of semantic features and the test seismic trace is transformed to a second set of semantic features employing a first set of learnt parameters; and
in a second stage, a discriminative correlation filter (107) automatically localizes the relative temporal position of the seismic events in the test seismic trace relative to the reference seismic trace using the generated first and second set of semantic features employing a second set of learnt parameters,
wherein the generated semantic features include one or more learnt features derived from arrival times, signal amplitudes, and frequency content of reflection signals, and wherein the generated first and second set of semantic features discriminate the seismic events in the respective seismic traces from other spurious signals and noise within the respective seismic trace.

14. The method as claimed in claim 13, wherein a gaussian response map with peak at the temporal position corresponding to the seismic events in the reference seismic trace is generated by the discriminative correlation filter (107) using the generated first and second set of semantic features, filter weights learnt from the reference seismic trace and its convolution with the test seismic trace to generate the relative temporal position of the seismic events in the test seismic trace relative to the reference seismic trace.

15. The method as claimed in claim 13, wherein the step of predicting relative temporal position of the one or more seismic events comprises employing the deep learning model (122) and an online update method wherein the discriminative correlation filter (107) is updated based on a new received reference seismic trace in order to adapt to changing characteristics of the seismic events on the respective reference seismic trace before prediction of the relative temporal position of the seismic event on the test seismic trace.

16. The method as claimed in claim 11, wherein the deep learning model (122) is trained using a synthetic dataset comprising several training samples employing learning parameters through several iterations to minimize the consistency loss function representation of a difference between location of the seismic events in a first seismic trace of each training data sample and a predicted location of a related seismic event in the first seismic trace predicted through forward tracking and backward tracking.

17. The method as claimed in claim 11, wherein the method comprises generating a 3D sub-surface image by constructing a sub-surface velocity model from the predicted relative temporal positions of the seismic events across all test seismic traces in each of the time windows, thereby providing a high-resolution spatial representation of the sub-surface formations, wherein the 3D sub-surface image includes information related to depth, composition, and structure of geological formations.