Machine learning-based geologic feature detection method and system

Abstract

A method for generating training seismic images for a convolutional neural network, CNN, for detecting a geological feature e.g. injectite, includes receiving 1600 seismic data d associated with a first subsurface; applying 1602 an imaging process to the seismic data d for generating an image of the first subsurface; selecting 1604 first regions from the image, which are free of the geologic feature; selecting 1606 second regions from the image, which include the geologic feature; and combining 1608 in a random manner one region of the first regions free of the geologic feature with one region of the second regions including the geologic feature to obtain the training seismic images. The method may comprise applying transfer learning.

Description

[0010] According to an embodiment, there is a method for generating training seismic images for a convolutional neural network, CNN, for detecting a geological feature. The method includes receiving seismic data d associated with a first subsurface, applying an imaging process to the seismic data d for generating an image of the first subsurface, selecting first regions from the image, which are free of the geologic feature, selecting second regions from the image, which include the geologic feature, and combining in a random manner one region of the first regions free of the geologic feature with one region of the second regions including the geologic feature to obtain the training seismic images.

[0011] According to another embodiment, there is a device for generating training seismic images for a convolutional neural network, CNN, for detecting a geological feature. The device includes an interface configured to receive seismic data d associated with a first subsurface and a processor connected to the interface. The processor is configured to apply an imaging process to the seismic data d for generating an image of the first subsurface, select first regions from the image, which are free of the geologic feature, select second regions from the image, which include the geologic feature, and combine in a random manner one region of the first regions free of the geologic feature with one region of the second regions including the geologic feature to obtain the training seismic images.

[0012] According to yet another embodiment, there is a method for detecting a geological feature in seismic data d’, and the method includes training a deep neural network, DNN, with synthetic data, corresponding to a first subsurface, receiving the seismic data d’ associated with a second subsurface, generating predicted data by feeding the seismic data d’ to the trained DNN, wherein the predicted data includes the geological feature, extracting new regions free of the geological feature from the seismic data d’, retraining the DNN with the new regions free of the geological feature, and generating additional predicted data with the retrained DNN. The additional predicted data exhibits the geological feature.

[0013] According to yet another embodiment, there is a device for detecting a geological feature in seismic data, and the device includes a processor implementing a convolutional neural network, CNN, which is trained with synthetic data, corresponding to a first subsurface and an interface connected to the processor and configured to receive seismic data d’ associated with a second subsurface. The processor is configured to generate predicted data by feeding the seismic data d’ to the trained CNN, wherein the predicted data includes the geological feature, extract new regions free of the geological feature from the seismic data d’, retrain the trained CNN with the new regions free of the geological feature, and generate additional predicted data with the retrained CNN. The additional predicted data exhibits the geological feature. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] For a more complete understanding of the present invention, reference is now made to the following descriptions taken in conjunction with the accompanying drawings, in which:

[0015] FIG. 1 is a schematic diagram of a marine based seismic acquisition system;

[0016] FIG. 2 is a schematic flow chart of a method that uses inversion for detecting injectites in seismic data;

[0017] FIG. 3 is a schematic flow chart of a method that uses DNN for detecting injectites in seismic data;

[0018] FIG. 4A illustrates seismic data that includes injectites while FIG. 4B illustrates the injectite mask extracted from the seismic data of FIG. 4A using the DNN of FIG. 3;

[0019] FIG. 5 illustrates the structure of a convolutional neural network, CNN, used to extract features from seismic data;

[0020] FIG. 6 is a flow chart of a method for generating synthetic seismic data fortraining the CNN;

[0021] FIG. 7A shows an image of input seismic data d including injectites, FIGs. 7B and 7C show selected regions of the image of FIG. 7A that are free of injectites, and FIGs. 7D and 7E show selected regions of the image of FIG. 7Athat include injectites;

[0022] FIGs. 8A and 8B illustrate regions of the image of FIG. 7A that are free of injectites, FIGs. 8C and 8D illustrate regions of the image of FIG. 7A that include injectites, and FIGs. 8E and 8F illustrate randomly combined regions from FIGs. 8A and 8B with regions from FIGs. 8C and 8D;

[0023] FIG. 9A illustrates input seismic data collected from a subsurface, FIG. 9B shows the CNN prediction for the subsurface, and FIG. 9C show the injective mask obtained from the CNN prediction;

[0024] FIG. 10A illustrates input seismic data having regions that include the geologic feature, FIG. 10B illustrates the predicted data obtained with a trained CNN, including the geologic feature and bias features, and FIG. 10C illustrates the predicted data obtained with a retrained CNN using transfer learning approach, to avoid bias;

[0025] FIG. 11 is a flow chart of a method for retraining the CNN using transfer learning;

[0026] FIG. 12 illustrates the predicted injectite clusters in a 3D image;

[0027] FIG. 13 is a flow chart of a method for generating the 3D image with the injectite clusters;

[0028] FIG. 14 schematically illustrates how to determine the amplitude envelope of the signal as the predicted injectite signal is converted into injectite mask using the amplitude envelop;

[0029] FIG. 15A illustrate legacy seismic data, FIG. 15B illustrates newer seismic data, FIG. 15C illustrates the injectite extracted from the legacy seismic data, and FIG. 15D illustrates the injectite extracted from the newer seismic data;

[0030] FIG. 16 is a flow chart of a method for training the CNN with synthetic seismic data that includes a geologic feature which is randomly added;

[0031] FIG. 17 is a flow chart of a method for extracting with a trained CNN a geologic feature from seismic data; and

[0032] FIG. 18 is a schematic diagram of a computing device that is configured to implement the methods discussed herein. DETAILED DESCRIPTION OF THE INVENTION

[0033] The following description of the embodiments refers to the accompanying drawings. The same reference numbers in different drawings identify the same or similar elements. The following detailed description does not limit the invention. Instead, the scope of the invention is defined by the appended claims. The following embodiments are discussed with regard to a modified convolution neural network-based process that uses synthetically generated training data for determining the location of a geologic feature or body in a subsurface, using terminology of seismic data processing. However, the embodiments to be discussed next are not limited to seismic data, but may be applied to other types of data, for example, electromagnetic wave data or acoustic data. Also, the embodiments discussed herein are not limited to applying the modified convolution neural network, but may use other machine learning algorithms.

[0034] Reference throughout the specification to “one embodiment” or “an embodiment” means that a particular feature, structure or characteristic described in connection with an embodiment is included in at least one embodiment of the subject matter disclosed. Thus, the appearance of the phrases “in one embodiment” or “in an embodiment” in various places throughout the specification is not necessarily referring to the same embodiment. Further, the particular features, structures or characteristics may be combined in any suitable manner in one or more embodiments.

[0035] According to an embodiment, realistic synthetic seismic data is generated for various geologic bodies, for example, injectites, which are poorly imagined with the conventional methods. Then, the synthetic data is used to train a neural network so that seismic images of the geological bodies are generated instead of the typical binary masks that the existing networks are generating. A specific error function is selected for a convolutional neural network to achieve these results. Then, the trained network is used to generate a 3D seismic volume in which 2D images of the geological body are clustered together.

[0036] In the following, a ll-Net model architecture is used for exemplifying the various embodiments. However, the U-Net may be replaced with other model architectures for determining the geological bodies discussed herein. A U-Net is a type of convolutional neural network architecture commonly used in image segmentation tasks [1], The name "U-Net" comes from the network's U-shaped architecture, which consists of an encoding path (contracting path) followed by a decoding path (expansive path), resulting in a U-shaped overall structure. The U-Net architecture is particularly effective for tasks like image segmentation, where the goal is to classify each pixel in an input image into different categories or regions of interest. It is also used in other image-related tasks, such as image-to-image translation and generative tasks.

[0037] A modified U-Net architecture, which is schematically illustrated in FIG. 5, includes an encoder (also called a contracting path) 510 and this part of the network is responsible for capturing high-level features from an input image 502 that corresponds to seismic data. It typically consists of several convolutional and pooling layers (512, 514, as illustrated in the figure) that reduce the spatial dimensions while increasing the number of channels. Note that the number of layers and the size of the images at each layer are exemplary, and other values for these characteristics may be used. At the center of the modified ll-Net 500, there is a bottleneck layer 530 that retains the most critical features of the input image.

[0038] The modified network 500 also includes a decoder (also called an expansive path) 520. The decoder part of the network upsamples the features obtained from the encoder 510 to the original input image's size. It involves a series of convolutional and upsampling layers (512, 522). Skip connections 540 are part of the ll-Net architecture and these connections directly link corresponding layers Li, with “i” being an integer, from the encoder 510 to the decoder 520. This helps in preserving spatial information and assists in the segmentation task. The skip connections 540 help the modified network 500 maintain fine-grained information by connecting the encoder layers to corresponding decoder layers. These connections aid in better localization of segmented objects by combining both low-level and high-level features. The U-Net's architecture allows it to capture both local and global context information, making it well-suited for image segmentation tasks.

[0039] The choice of the cost function (also known as loss function) for a ll-Net or any neural network depends on the specific task and the type of data being used. For image segmentation tasks, such as those commonly addressed by U-Net architectures, a common choice of loss function is the pixel-wise cross-entropy loss or its variants. The pixel-wise cross-entropy loss compares the predicted segmentation mask (output of the network) with the ground truth mask (target) on a per-pixel basis. It measures the dissimilarity between the predicted pixel values and the true pixel values, encouraging the network to generate accurate segmentations.

[0040] Mathematically, the pixel-wise cross-entropy loss for a single pixel, which is used in the traditional U-Net is defined as: Lpixeity, y) = -(y ■ iog(y) + (1 - y) ■ iog(i - yj), (1) where y is the true label for the pixel (0 or 1, indicating background or foreground) and y is the predicted probability for the pixel to belong to the foreground (class of interest). For a full image or a batch of images, the loss is typically computed as the average of pixel-wise losses over all pixels. The total loss for the entire segmentation task is the sum of the pixel-wise losses. Additionally, to address class imbalance, some modifications might be made to the loss function, such as adding class weights to give more importance to underrepresented classes.

[0041] In a U-Net architecture, the cost function is typically applied at the output layer 524, which corresponds to the final layer of the network where the predicted segmentation mask is generated. More specifically, the encoding path 520 of the U-Net 500 consists of several layers of convolutional and pooling operations Li that progressively reduce the spatial dimensions of the input image 502 while increasing the number of channels. These layers are responsible for extracting features from the input image. The output layer of the U-Net corresponds to the final layer of the decoding path. It produces the predicted segmentation mask for the input image. The cost function is applied to the predicted mask at this layer.

[0042] While the cost function is applied at the output layer, the entire U-Net architecture, including both the encoding and decoding paths 510 and 520, contributes to the computation of the loss by learning meaningful features and relationships between pixels. The output layer 524 is where the predicted results is generated, and the cost function evaluates the accuracy of these predictions. However, the traditional cost function does not allow to determine the amplitude of the signal in the seismic data for the geologic feature of interest. Thus, the modified network 500 uses a different cost function, which preservers the amplitude of the signal. This cost function is discussed later.

[0043] Based on the architecture of the ll-Net discussed above, a specific model training and evaluation process is now discussed. First, a process of seismic data generation necessary for training the network is introduced. As noted above, the available real seismic data d available for training a network for determining injectites (or other geological bodies as noted above) is limited because this data is expensive to generate, and also proprietary, which means that there are no publicly available sites of such data. To generate more training datasets from the limited number of manually picked injectites, the inventors propose in this embodiment a novel approach of synthesizing realistic training datasets. As illustrated in FIG. 6, this novel approach includes a step 600 of receiving seismic data d from one or more seismic surveys. In step 602, a large number of 2D slices are extracted from 3D seismic images generated based on the seismic data d, without injectites, as the background. In this regard, note that FIG. 7A shows an image obtained from the seismic data d (any known process may be used for generating the image) and the image includes regions 710 in which the injectite is not present and regions 720 in which the injectite is present. The presence of the injectite in the image of FIG. 7A is determined by visual observation of the operator of the seismic data. The slices 710 extracted in step 602 are illustrated in FIGs. 7B and 7C. These slices are specifically selected by the operator to not include any injectite. Note that from one single image 700 in FIG. 7A, it is possible to extract plural slices 710 of data without injectite. To further generalize the model, the background should cover various types of geologies, signatures and signal-to-noise ratio (SNR). For this reason, the slices 710 are selected from various locations of the image 700.

[0044] Then, in step 604, a 2D seismic image dictionary is generated based on manually interpreted injectites 720, which are also extracted from image 700 of FIG. 7A. Slices 720 including the injectites are shown in FIGs. 7D and 7E. Note that these slices are selected by the operator of the process, based on experience and / or previous algorithms. In one application, the number ofclean images 710 should be greater than the manually interpreted images 720. While the current example uses manually interpreted injectite images 720 from the same geology 700, in one application, it is possible to use injectites images from different geologies, i.e., from another seismic survey that surveyed a different region of the earth.

[0045] In step 606, augmentation may be applied, which modifies copies of already-existing datasets to artificially expand the training set. For example, in one embodiment, the augmentation step makes minor changes to the dataset or adds new data points. While training the above discussed model, the inventors added noise, scaled, flipped, and cropped the data in the data augmentation stage. Then, in step 608, the method randomly pairs a region 710 from the background and a region 720 from the injectite dictionary and sums the two regions to create the input training data 502 of the CNN 500, as shown in FIGs. 8A to 8F. While this embodiment is discussed with regard to CNN 500, one skilled in the art would understand that other machine learning algorithms may be used, for example, deep neural network (DNN). In one application, a visual transformer (ViT) architecture may be used. Thus, in this and the other embodiments, the CNN 500 may be replaced with a DNN. The target of the CNN 500 is the 2D seismic image of injectites. By the pairing of step 608, a large number of training datasets 502 are obtained, and they cover a large variety of injectites and background geologies. In one application, 80% of the obtained datasets were used for training, and 20% for validation. While FIGs. 8A to 8F show an exemplary methodology using 2 sample datasets for injectite free regions 710 and two manually interpreted regions 720 that include injectites, this is not a limitation and the method can use as many samples as need to create the input training dataset 502.

[0046] Next, a method for model training and evaluation based on the generated training dataset 502 is discussed. The method uses the modified 2D U-Net 500 for delineating injectites from input seismic images. The traditional U-Net has been modified as discussed above with regard to FIG. 5 so that (1) the model predicts the seismic images of injectite bodies rather than the binary masks with pixels inside the injectites set to true and pixels of background set to false, in order to preserve the true amplitude of injectite’s seismic responses for future AVO analysis, and (2) consequently, the inventors have chosen the mean squared error (MSE) loss instead of the binary cross entropy (BCE) loss (see equation (1) above), as the MSE loss is used for prediction continuous values whereas BCE is more often used for binary mask usage. The change of the loss function takes place in the final layer 524 of the network. The MSE loss is defined as: n 1=1 where n is the number of pixels. One skilled in the art would understand that instead of the BCE loss of the traditional U-Net, it is possible to use other loss functions, such as but not limited to Mean Absolute error (MAE), as long as the new loss function provides an output that captures the true amplitude of the injectite. With these modifications, the model 500 shown in FIG. 5 is trained with the training dataset 502 generated based on the method of FIG. 6.

[0047] In one application, the inventors applied the pretrained modified model 500 to unseen real seismic data. As shown in FIGs. 9A to 9C, the pretrained model 500 has generalized well. FIG. 9A shows the input data 902, which corresponds to actual seismic data recorded during a seismic survey. FIG. 9B shows the predicted data 904 that indicates the true amplitude of the injectites 910, and FIG. 9C shows the injectite masks 906. The modified model 500 delineates not only strong seismic images of injectites 910, but also subtle amplitude variations 912 at the edge of each injectite body, which are missed by conventional ll-Net approaches (see FIG. 4B).

[0048] However, after analysing these results (see, for example, input data 902 used in FIG. 10A for generating predicted data 904 shown in FIG. 10B), the inventors observed that the modified model sometimes delineates reflections from the fault surfaces as false positives, as highlighted by the dashed boxes 1000 in FIG. 10B. This is due to the “domain shifting” problem, in which the geology of the test dataset 902 is different from the training dataset 502. In this context, “domain shifting” refers to the challenge of dealing with differences between training data and real-world data when applying a machine learning model to a specific problem, such as determining the location of sand injectites in the subsurface. More specifically, when using machine learning for this task, the model 500 is usually trained using a dataset that includes various features and attributes from known locations of sand injectites. However, the real-world geological conditions, data collection methods, and other factors might differ from the training dataset. This misalignment or shift between the training data distribution and the distribution of real-world data is what is referred to as “domain shift.” Domain shifting can pose challenges to the performance and generalization of machine learning models. If the model encounters data in the real world that is significantly different from the training data, its performance might degrade, as illustrated in FIG. 10B. This is because the model has learned patterns and relationships specific to the training data distribution, and it may struggle to adapt to new or different patterns in the real-world data.

[0049] To mitigate domain shifting and improve the model's ability to generalize to new environments, a transfer learning technique may be applied. Transfer learning involves pre-training a model on a related task or dataset and fine-tuning it on the specific task of interest. This can help the model capture more general features that are applicable across domains.

[0050] Thus, to improve accuracy and reduce the number of false positives, the inventors applied transfer learning to the modified model 500 as now discussed with regard to FIG. 11. More specifically, a modified model training and evaluation method starts in step 1102 with training the modified model 500 with the synthetic training data 502 obtained with the method discussed in FIG. 6. Note that this data was obtained from actual seismic data d, but this actual seismic data d was not collected from the subsurface that is now being analyzed. The seismic data collected from the current subsurface is d’. In other words, the initial seismic data d is obtained from a first subsurface and this data is used to generate the training data 502, and the current seismic data d’ that is desired to be analyzed is obtained from a second subsurface, which might have a different geology from the first subsurface. Then, in step 1104, the modified model 500 (new loss function for generating actual amplitudes of the injectite zone) is trained with the synthetic training data 502 corresponding to the first subsurface. In step 1106, the actual seismic data d’, from the current subsurface (second subsurface) that is being analyzed, is fed to the modified model 500 for generating the predicted geologic formation 904. In step 1108, the predicted data 904 is evaluated and if a domain shift is determined in step 1110, then the method advances to step 1112, where new training datasets are generated from new clean regions 720’ selected from the current seismic data d’ (see FIG. 10A). In other words, the created new set of training datasets 502’ uses the new injectite-free data 720’ from the seismic data d’ 902 of the test area, as clean seismic data. This data is then used in step 1114 to retrain the modified model 500 in step 1104 and the current seismic data d’ is again fed to the trained modified model for generating new predicted injectite data 930, as shown in FIG. 10C. These steps minimize the domain shift and, consequently, significantly improved the results, as shown when comparing the data in FIGs. 10B and 10C (note that the false injectite 1000 have been removed in FIG. 10C). In one application, the transfer learning step 1114 is performed until there is no more (or limited) domain shifting, that is, until the modified model is generalized well and not picking false positive in the test dataset.

[0051] Next, the predicted injectite data section 930, which is obtained as 2D slices, is combined to generate a 3D seismic volume. More specifically, it is possible to combine the predicted injectites 910 on each 2D section 930 to generate a 3D seismic volume as shown in FIG. 12. According to a method illustrated in FIG. 13, it is possible to cluster the obtained 3D injectite bodies. The method includes a step 1302 of obtaining / receiving the 3D injectite prediction, followed by a step 1304 of generating binary masks of the injectites by thresholding the amplitude envelopes of the injectite signal based on [2], According to this amplitude envelope method, a waveform ^(t) that describes the injectite signal is Hilbert transformed to obtain a new waveform ^(t). Then, an analytic signal N(g(t)) is constructed so that W(^(t)) = g(t) + ig(t). The envelope of this analytic signal, which is illustrated in FIG. 14, is called the amplitude envelope and this envelope is used to calculate the injectite mask by thresholding.

[0052] Then, in step 1306, injectite body 3D clusters are created based on a pixel distance. In other words, a threshold distance is selected and any two pixels that are separated from each other by a distance smaller than the threshold distance are considered to be part of a same cluster and thus, they are assigned to that cluster. Multiple clusters may be formed into a given volume, as illustrated in FIG. 12. Then, in step 1308, a marching cube algorithm [3] is applied within each cluster to create a triangulated surface. These surfaces can be further visualized in step 1310 using any visualization libraries, as shown in FIG. 12.

[0053] The inventors found that the quality of seismic data d also plays an important role in model performance. FIGs. 15A to 15D compare the results of the pretrained modified CNN model 500 on vintage data 1510 versus more recent seismic data 1520. The model is not able to delineate injectites from the vintage seismic data 1510, as shown in FIG. 15C. However, the model performed well on recent seismic data 1520, being able to predict the injectites 1522 as shown in FIG. 15D. This shows that the quality of the initial seismic data d is an important factor. Therefore, the model performance could be impacted by the quality of seismic images, such as signal-to-noise ratio, frequency content and resolution, etc, and these could be taken into account during the model building.

[0054] An advantage of the modified CNN model 500 is its efficiency. The method is able to delineate the injectite body within few hours using GPU for inference while it used to take few weeks or months of manual interpretation. This means that the method can use this information in the middle of a seismic processing project. For instance, injectites are high-velocity sands trapped in thin layers, difficult to be detected by full waveform inversion at low frequency updates. With the methods discussed herein, it is possible now to use the CNN prediction 930 as a priori information in an early velocity model building stage.

[0055] The above discussed embodiments describe a novel workflow for injectite detection using U-Net. Other CNN models may be used if modified to generate the true amplitude of the injectite. Other geologic bodies, in addition to the injectites, may be detected with the methods discussed above. These methods address the challenges inherent in the limited number of training datasets, and produce a pretrained model which delineates injectite events on migrated seismic images. To further address the issue of domain shifting, the method includes a transfer learning approach, which avoids mispredicting faults and other diffraction events as injectites.

[0056] Based on the embodiments discussed above, a method for generating training seismic images for a convolutional neural network for detecting a geological feature is now discussed with regard to FIG. 16. The method includes a step 1600 of receiving seismic data d associated with a first subsurface, a step 1602 of applying an imaging process to the seismic data d for generating an image of the first subsurface, a step 1604 of selecting regions from the image, which are free of the geologic feature, a step 1606 of selecting regions from the image, which include the geologic feature, and a step 1608 of combining, in a random manner, one region of the regions free of the geologic feature with one region of the regions including the geologic feature to obtain the training seismic images. The geologic feature is injectite in one application. The regions free of the geologic feature are more numerous than the regions including the geologic feature.

[0057] The method may further include a step of training the CNN with the training seismic images to detect the geologic feature, and / or a step of receiving seismic data d’ corresponding to a second subsurface, and a step of applying the trained CNN to the seismic data d’ to find the geologic feature in a predicted image. The method may further include a step of analyzing the predicted image of the trained CNN to determine domain shifting, as the second subsurface includes a geology different from the first subsurface, and / or a step of applying transfer learning to the trained CNN to reduce prediction errors. The step of applying transfer learning includes extracting regions free of the geologic feature from the predicted image, which corresponds to the second subsurface, retraining the trained CNN with the regions free of the geologic feature from the second subsurface, and predicting an additional image of the second subsurface with the retrained CNN.

[0058] According to another embodiment, there is a method for detecting a geological feature in seismic data that uses a DNN or a modified CNN. The method includes a step 1700 of training the DNN or the modified CNN with synthetic data, corresponding to a first subsurface, a step 1702 of receiving seismic data d’ associated with a second subsurface, a step 1704 of generating predicted data by feeding the seismic data d’ to the trained, DNN or modified CNN, where the predicted data includes the geological feature, a step 1706 of extracting regions free of the geological feature from the seismic data d’, a step 1708 of retraining the DNN or modified CNN with the regions free of the geological feature, and a step 1710 of generating additional predicted data with the retrained, DNN or modified CNN. The additional predicted data exhibits the geological feature.

[0059] The method may further include analyzing the predicted image of the trained DNN or CNN to determine domain shifting, as the second subsurface includes a geology different from the first subsurface, and / or applying transfer learning to the trained DNN or CNN to reduce prediction errors, where the step of applying transfer learning includes: extracting regions free of the geologic feature from the predicted image, which corresponds to the second subsurface, retraining the trained DNN or CNN with the regions free of the geologic feature from the second subsurface, and predicting an additional image of the second subsurface with the retrained DNN or CNN. The DNN or CNN model is modified to use a mean square error loss function.

[0060] The method may further include applying an imaging process to the seismic data d for generating an image of the first subsurface, selecting regions from the image, which are free of the geologic feature, selecting regions from the image, which include the geologic feature, and combining in a random manner one region of the regions free of the geologic feature with one region of the regions including the geologic feature to obtain the training seismic images.

[0061] The methods discussed herein may be applied not only to the field of subsurface exploration, for example, hydrocarbon exploration and development, geothermal exploration and development, and carbon capture and sequestration, or other natural resource exploration and exploitation. They could also be employed for surveying and monitoring for windfarm applications, both onshore and offshore, and also for medical imaging applications.

[0062] The above-discussed procedures and methods may be implemented in a computing device as illustrated in FIG. 18. Hardware, firmware, software or a combination thereof may be used to perform the various steps and operations described herein. The computing device 1800 is suitable for performing the activities described in the above embodiments and may include a server 1801. Such a server 1801 may include a central processor (CPU) 1802 coupled to a random access memory (RAM) 1804 and to a read-only memory (ROM) 1806. ROM 1806 may also be other types of storage media to store programs, such as programmable ROM (PROM), erasable PROM (EPROM), etc. Processor 1802 may communicate with other internal and external components through input / output (I / O) circuitry 1808 and bussing 1810 to provide control signals and the like. Processor 1802 carries out a variety of functions as are known in the art, as dictated by software and / or firmware instructions.

[0063] Server 1801 may also include one or more data storage devices, including hard drives 1812, CD-ROM drives 1814 and other hardware capable of reading and / or storing information, such as DVD, etc. In one embodiment, software for carrying out the above-discussed steps may be stored and distributed on a CD-ROM or DVD 1816, a USB storage device 1818 or other form of media capable of portably storing information. These storage media may be inserted into, and read by, devices such as CD-ROM drive 1814, disk drive 1812, etc. Server 1801 maybe coupled to a display 1820, which may be any type of known display or presentation screen, such as LCD, plasma display, cathode ray tube (CRT), etc. A user input interface 1822 is provided, including one or more user interface mechanisms such as a mouse, keyboard, microphone, touchpad, touch screen, voice-recognition system, etc.

[0064] Server 1801 may be coupled to other devices, such as sources, sensors, or any other seismic data system. The server may be part of a larger network configuration as in a global area network (GAN) such as the Internet 1828, which allows ultimate connection to various landline and / or mobile computing devices.

[0065] As described above, the apparatus 1800 may be embodied by a computing device. However, in some embodiments, the apparatus may be embodied as a chip or chip set. In other words, the apparatus may comprise one or more physical packages (e.g., chips) including materials, components and / or wires on a structural assembly (e.g., a baseboard). The structural assembly may provide physical strength, conservation of size, and / or limitation of electrical interaction for component circuitry included thereon. The apparatus may therefore, in some cases, be configured to implement an embodiment of the present invention on a single chip or as a single “system on a chip.” As such, in some cases, a chip or chipset may constitute means for performing one or more operations for providing the functionalities described herein.

[0066] In an example embodiment, the processor 1802 may be configured to execute instructions stored in the memory device 1804 or otherwise accessible to the processor. Alternatively, or additionally, the processor may be configured to execute hard coded functionality. As such, whether configured by hardware or software methods, or by a combination thereof, the processor may represent an entity (e.g., physically embodied in circuitry) capable of performing operations according to an embodiment of the present invention while configured accordingly. Thus, for example, when the processor is embodied as an ASIC, FPGA or the like, the processor may be specifically configured hardware for conducting the operations described herein. Alternatively, as another example, when the processor is embodied as an executor of software instructions, the instructions may specifically configure the processor to perform the algorithms and / or operations described herein when the instructions are executed. However, in some cases, the processor may be a processor of a specific device (e.g., a pass-through display or a mobile terminal) configured to employ an embodiment of the present invention by further configuration of the processor by instructions for performing the algorithms and / or operations described herein. The processor may include, among other things, a clock, an arithmetic logic unit (ALU) and logic gates configured to support operation of the processor.

[0067] The term “about” is used in this application to mean a variation of up to 20% of the parameter characterized by this term.

[0068] It will be understood that, although the terms first, second, etc. may be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first object or step could be termed a second object or step, and, similarly, a second object or step could be termed a first object or step, without departing from the scope of the present disclosure. The first object or step, and the second object or step, are both, objects or steps, respectively, but they are not to be considered the same object or step.

[0069] The terminology used in the description herein is for the purpose of describing particular embodiments and is not intended to be limiting. As used in this description and the appended claims, the singular forms "a," "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term "and / or" as used herein refers to and encompasses any possible combinations of one or more of the associated listed items. It will be further understood that the terms "includes," "including," "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. Further, as used herein, the term "if" may be construed to mean "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context.

[0070] The disclosed embodiments provide a modified CNN for determining a geologic body or feature in seismic data associated with a subsurface. CNN is trained based on synthetic data generated by randomly pairing regions free of the geologic feature from an image, and regions that include the geologic feature for the same image. It should be understood that this description is not intended to limit the invention. On the contrary, the embodiments are intended to cover alternatives, modifications and equivalents, which are included in the spirit and scope of the invention as defined by the appended claims. Further, in the detailed description of the embodiments, numerous specific details are set forth in order to provide a comprehensive understanding of the claimed invention. However, one skilled in the art would understand that various embodiments may be practiced without such specific details.

[0071] Although the features and elements of the present embodiments are described in the embodiments in particular combinations, each feature or element can be used alone without the other features and elements of the embodiments or in various combinations with or without other features and elements disclosed herein.

[0072] This written description uses examples of the subject matter disclosed to enable any person skilled in the art to practice the same, including making and using any devices or systems and performing any incorporated methods. The patentable scope of the subject matter is defined by the claims, and may include other examples that occur to those skilled in the art. Such other examples are intended to be within the scope of the claims. References The entire content of all the publications listed herein is incorporated by reference in this patent application. [1] Olaf Ronneberger, Philipp Fischer, and Thomas Brox, Ronneberger, 0., Fischer, P., &Brox, T.

[2015] U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, 234-241. [2] Ulrich, T. J.

[2006] , Envelope calculation from the Hilbert transform. Los Alamos Nat. Lab., Los Alamos, NM, USA, Tech. Rep:, [3] Lorensen, W. E., &Cline, H. E.

[1987] Marching cubes: A high resolution 3D surface construction algorithm. ACM siggraph computer graphics, 21(4), 163-169.

Claims

1. A method for generating training seismic images (502) for a convolutional neural network, CNN, (500) for detecting a geological feature (910), the method comprising:receiving (1600) seismic data d associated with a first subsurface;applying (1602) an imaging process to the seismic data d for generating an image (700) of the first subsurface;selecting (1604) first regions (710) from the image (700), which are free of the geologic feature (910);selecting (1606) second regions (720) from the image (700), which include the geologic feature (910); andcombining (1608) in a random manner one region of the first regions (710) free of the geologic feature with one region of the second regions (720) including the geologic feature to obtain the training seismic images (502).

2. The method of Claim 1, wherein the geologic feature is injectite.

3. The method of Claim 1, wherein the first regions free of the geologic feature are more numerous than the second regions including the geologic feature.

4. The method of Claim 1, further comprising:training the CNN with the training seismic images to detect the geologic feature.

5. The method of Claim 4, further comprising:receiving seismic data d’ corresponding to a second subsurface; andapplying the trained CNN to the seismic data d’ to find the geologic feature in a predicted image.

6. The method of Claim 5, further comprising:analyzing the predicted image of the trained CNN to determine domain shifting, as the second subsurface includes a geology different from the first subsurface.

7. The method of Claim 6, further comprising:applying transfer learning to the trained CNN to reduce prediction errors.

8. The method of Claim 7, wherein the step of applying transfer learning includes:extracting regions free of the geologic feature from the predicted image, which corresponds to the second subsurface;retraining the trained CNN with the regions free of the geologic feature from the second subsurface; andpredicting an additional image of the second subsurface with the retrained CNN.

9. A device (1600) for generating training seismic images (502) for a convolutional neural network, CNN, (500) for detecting a geological feature (910), the device comprising:an interface (1810) configured to receive (1600) seismic data d associated with a first subsurface; anda processor (1802) connected to the interface (8610) and configured to, apply (1602) an imaging process to the seismic data d for generating an image (700) of the first subsurface;select (1604) first regions (710) from the image (700), which are free of the geologic feature (910);select (1606) second regions (720) from the image (700), which include the geologic feature (910); andcombine (1608) in a random manner one region of the first regions (710) free of the geologic feature with one region of the second regions (720) including the geologic feature to obtain the training seismic images (502).

10. The device of Claim 9, wherein the geologic feature is injectite.

11. A method for detecting a geological feature (910) in seismic data d’, the method comprising:training (1700) a deep neural network, DNN, with synthetic data (502), corresponding to a first subsurface;receiving (1702) the seismic data d’ associated with a second subsurface;generating (1704) predicted data (930) by feeding the seismic data d’ to thetrained DNN (500), wherein the predicted data (930) includes the geological feature(910);extracting (1706) new regions (720’) free of the geological feature (910) from the seismic data d’;retraining (1708) the DNN (500) with the new regions (720’) free of thegeological feature (910); andgenerating (1710) additional predicted data (930) with the retrained DNN (500),wherein the additional predicted data (930) exhibits the geological feature (910).

12. The method of Claim 11, further comprising:analyzing the predicted image of the trained DNN to determine domain shifting, as the second subsurface includes a geology different from the first subsurface.

13. The method of Claim 12, further comprising:applying transfer learning to the trained DNN to reduce prediction errors.

14. The method of Claim 13, wherein the step of applying transfer learning includes:extracting new regions free of the geologic feature from the seismic data d’, which corresponds to the second subsurface;retraining the trained DNN with the new regions free of the geologic feature from the second subsurface; andpredicting an additional image of the second subsurface with the retrained DNN.

15. The method of Claim 11, wherein the DNN model is modified to use a mean square error loss function in the last layer.

16. The method of Claim 11, further comprising:applying an imaging process to the seismic data d for generating an image (700) of the first subsurface;selecting first regions (710) from the image (700), which are free of the geologic feature (910);selecting second regions (720) from the image (700), which include the geologic feature (910);combining in a random manner one region of the first regions (710) free of the geologic feature with one region of the second regions including the geologic feature to obtain the training seismic images (502).

17. The method of Claim 11, wherein the geologic feature is injectite.

18. The method of Claim 11, wherein the firs regions free of the geologic feature are more numerous than the second regions including the geologic feature.

19. A device (1800) for detecting a geological feature (910) in seismic data, the device comprising:a processor (1802) implementing a convolutional neural network, CNN, (500), which is trained with synthetic data (502), corresponding to a first subsurface; andan interface (1810) connected to the processor (1802) and configured to receive (1702) seismic data d’ associated with a second subsurface,wherein the processor (1802) is configured to,generate (1704) predicted data (930) by feeding the seismic data d’ to the trained CNN (500), wherein the predicted data (930) includes the geological feature (910);extract (1706) new regions free of the geological feature (910) from the seismic data d’;retrain (1708) the trained CNN (500) with the new regions free of the geological feature (910); andgenerate (1710) additional predicted data (930) with the retrained CNN (500), wherein the additional predicted data (930) exhibits the geological feature (910).

20. The device of Claim 19, wherein the processor is further configured to:analyze the predicted image output by the trained CNN to determine domain shifting, as the second subsurface includes a geology different from the first subsurface; andapply transfer learning to the trained CNN to reduce prediction errors.Application No: GB2400419.4Examiner:Contract Unit ExaminerClaims searched: 1-20Date of search: 7 January 2025Patents Act 1977: Search Report under Section 17Documents considered to be relevant:Category Relevant to claims Identity of document and passage or figure of particular relevance X Y X,Y X Y Y X: 1, 3-9, 11-16, 18-20 Y: 2, 10, 17 X: 1, 3-9, 11-16, 18-20 Y: 2, 10, 17 X: 1, 3-7, 9 Y: 2, 10 2, 10, 17 WO2021 / 258096 Al (EXXONMOBIL UPSTREAM RES CO) paragraphs [0003], [0005], [0020] - paragraph [0133]; claims 1-19; figures 1-10 WO2020 / 146863 Al (SCHLUMBERGER TECHNOLOGY CORP; SCHLUMBERGER CA LTD ET AL) paragraphs [0002], [0028] - paragraph [0154]; claims 1-20; figures 1-21 EP4030198 Al (EMERSON PARADIGM HOLDING LLC) paragraphs [0002], [0013], [0020] - paragraph [0107]; claims 1-12; figures 1-12 GEOPHYSICAL PROSPECTING, vol 68, 2019, VALENTIN TSCHANNEN ET AL, "Detection of point scatterers using diffraction imaging and deep learning", pages 830-844 page 830 - page 842; figures 1-11Categories:X Document indicating lack of novelty or inventive step A Document indicating technological background and / or state of the art. Y Document indicating lack of inventive step if P Document published on or after the declared priority date but combined with one or more other documents of same category. before the filing date of this invention. & Member of the same patent family E Patent document published on or after, but with priority date earlier than, the filing date of this application.Field of Search:Search of GB, EP. WO &US patent documents classified in the following areas of the UKCX :Worldwide search of patent documents classified in the following areas of the IPC____________G01V_________________________________________________The following online and other databases have been used in the preparation of this search reportInternational Classification:Subclass Subgroup Valid From G01V 0001 / 30 01 / 01 / 2006 G01V 0001 / 32 01 / 01 / 2006

Images