Using Synthetic Data Sets to Train an End-to-End Convolutional Neural Network for 3D Seismic Fault Segmentation

Xinming Wu, Luming Liang, Yunzhi Shi, and Sergey Fomel | Published with permission: Geophysics, Vol 84 | May-June 2019


Delineating faults from seismic images is a key step for seismic structural interpretation, reservoir characterization, and well placement. In conventional methods, faults are considered as seismic reflection discontinuities and are detected by calculating attributes that estimate reflection continuities or discontinuities. We consider fault detection as a binary image segmentation problem of labeling a 3D seismic image with ones on faults and zeros elsewhere. We have performed an efficient image-to-image fault segmentation using a supervised fully convolutional neural network. To train the network, we automatically create 200 3D synthetic seismic images and corresponding binary fault labeling images, which are shown to be sufficient to train a good fault segmentation network. Because a binary fault image is highly imbalanced between zeros (nonfault) and ones (fault), we use a class-balanced binary cross-entropy loss function to adjust the imbalance so that the network is not trained or converged to predict only zeros. After training with only the synthetic data sets, the network automatically learns to calculate rich and proper features that are important for fault detection. Multiple field examples indicate that the neural network (trained by only synthetic data sets) can predict faults
from 3D seismic images much more accurately and efficiently than conventional methods. With a TITAN Xp GPU, the training processing takes approximately 2 h and predicting faults in a 128 × 128 × 128 seismic volume takes only milliseconds.


Faults are typically recognized as lateral reflection discontinuities in a 3D seismic image. Based on this observation, numerous methods have been proposed to detect faults by calculating attributes of measuring seismic reflection continuity such as semblance (Marfurt et al., 1998) and coherency (Marfurt et al., 1999; Li and Lu, 2014; Wu, 2017), or reflection discontinuity such as variance (Van Bemmel and Pepper, 2000; Randen et al., 2001) and gradient magnitude (Aqrawi and Boe, 2011). These seismic attributes, however, can be sensitive to noise and stratigraphic features, which also correspond to reflection discontinuities in a seismic image. This means that measuring seismic reflection continuity or discontinuity alone is insufficient to detect faults (Hale, 2013).

Faults are typically more vertically aligned, whereas stratigraphic features mostly extend laterally. Based on this observation, Gersztenkorn and Marfurt (1999) suggest using vertically elongated windows in computing seismic coherence to enhance faults while suppressing the stratigraphic features. Similarly, some other authors (Bakker, 2002; Hale, 2009; Wu, 2017) apply smoothing in directions perpendicular to seismic reflections in computing coherence or semblance by assuming that faults are typically normal to reflections. However, faults are seldom vertical or are not necessarily perpendicular to seismic reflections. Therefore, some authors (Hale, 2013; Wu and Hale, 2016) propose smoothing the numerator and denominator of the semblance along fault strikes and dips to compute the fault-oriented semblance or fault likelihood. However, calculating fault-oriented semblance is computationally more expensive than the previous attributes because it requires scanning over all possible combinations of fault strikes and dips to find the maximum fault likelihoods.

Some fault detection methods start with some initial fault attributes and further enhance them by smoothing the attributes along fault strikes and dips (Neff et al., 2000; Cohen et al., 2006; Wu and Zhu, 2017). These methods also need to smooth the fault attributes over all possible combinations of fault strikes and dips to obtain the best enhanced fault features. Similarly, some authors (Pedersen et al., 2002, 2003) propose to enhance fault features along paths of “artificial ants” by assuming that the paths follow faults. Wu and Fomel (2018) propose an efficient method to extract optimal surfaces following maximum fault attributes and use these optimal surfaces to vote for enhanced fault images of fault probabilities, strikes, and dips.

Recently, some convolutional-neural-network (CNN) methods have been introduced to detect faults by pixel-wise fault classification (fault or nonfault) with multiple seismic attributes (Huang et al., 2017; Di et al., 2018; Guitton, 2018; Guo et al., 2018; Zhao and Mukhopadhyay, 2018). Wu et al. (2018) use a CNN-based pixel-wise classification method to not only predict the fault probability but also estimate the fault orientations at the same time. These methods need to choose a local window or cube to make fault prediction at every image pixel, which is computationally highly expensive, especially in 3D fault detection. In this paper, we consider the fault detection as a more efficient end-to-end binary image segmentation problem by using CNNs. Image segmentation has been well-studied in computer science, and multiple powerful CNN architectures (e.g., Girshick et al., 2014; Ren et al., 2015; Ronneberger et al., 2015; Xie and Tu, 2015; Badrinarayanan et al., 2017; He et al., 2017) have been proposed to obtain superior segmentation results.

In this paper, we use an efficient end-to-end CNN, simplified from U-Net (Ronneberger et al., 2015), to perform our task of 3D binary fault segmentation. We simplify the original U-Net by reducing the number of convolutional layers and features at each layer, which significantly saves graphics processing unit (GPU) memory and computational time but still preserves high performance in our 3D fault detection tasks. Considering a fault binary image is highly biased with mostly zeros but only very limited ones on the faults, we use a balanced cross-entropy loss function for optimizing the parameters of our CNN model. To train and validate the neural network, we design a workflow to automatically generate 3D synthetic seismic and corresponding fault images. In this workflow, the seismic folding and faulting structures, wavelet peak frequencies, and noise are defined by a set of parameters, and each parameter can be chosen from some predefined range. By randomly choosing a combination of these parameters within the predefined ranges, we are able to generate numerous unique seismic images and corresponding fault labeling images. We train and validate, respectively, by using 200 and 20 pairs of synthetic seismic and fault images, which turned to be sufficient to train a good CNN model for our fault detection tasks. Although trained by using only synthetic seismic data sets, our CNN model can work much better and more efficiently than the conventional methods for 3D fault detection in field seismic data sets that are recorded at totally different surveys. By using a TITAN Xp GUP, our CNN model takes less than 5 min to predict faults in a large seismic volume with 450 × 1950 × 1200 samples.

Trainging Data Sets

Training and validating a CNN model often requires a large amount of images and corresponding labels. Manually labeling or interpreting faults in a 3D seismic image could be extremely time consuming and highly subjective. In addition, inaccurate manual interpretation, including mislabeled and unlabeled faults, may mislead the learning process. To avoid these problems, we propose an effective and efficient way to create synthetic seismic images and corresponding fault labels for training and validating our CNN model.

Synthetic seismic and fault images

Figure 1. The workflow of creating 3D synthetic training data sets. We first generate a horizontal reflectivity model (a) with a sequence of random numbers between −1 and 1. We then add some folding structures (b) to the model by vertically shearing the model and the shearing shifts are defined by a combination of several 2D Gaussian functions. We also add some planar shearing (c) to the model to increase the complexity of folding structures. We further add planar faulting to the model to obtain a folded and faulted reflectivity model (d). We finally convolve the reflectivity model with a Ricker wavelet to obtain a synthetic seismic image (e) and add some random noise to obtain a final image (f).

The workflow that we use to create synthetic seismic and fault images is similar to the one used for generating synthetic data sets by Wu and Hale (2016). In this workflow, we first generate a 1D horizontal reflectivity model r(x; y; z) (Figure 1a) with a sequence of random values that are in the range of [−1; 1]. We then create some folding structures in the reflectivity model by vertically shearing the model. We define the folding structures by using the following function:

which combines with multiple 2D Gaussian functions and a linear-scale function 1.5z/zmax. The combination of 2D Gaussian functions yields laterally varying folding structures, whereas the linear-scale function damps the folding vertically from below to above. In this equation, each combination of the parameters a0; bk; ck; dk, and σk yields some specific spatially varying folding structures in the model. By randomly choosing each of the parameters from the predefined ranges, we are able to create numerous models with unique structures. With the shift map s1 (x; y; z), we use a sinc interpolation to vertically shift the original reflectivity model r(x; y; z) to obtain a folded model r(x; y; z + s1(x; y; z)) as shown in Figure 1b. To further increase the complexity of the structures in the model, we also add some planar shearing defined as follows:

where the shearing shifts are laterally planar while being vertically invariant. The parameters e0; f, and g, again, are randomly chosen from some predefined ranges. By sequentially applying the planar shifts s2(x; y; z) to the previously folded model r(x; y; z + s1(x; y; z)), we obtain a new reflectivity model r(x; y; z + s1 + s2) as shown in Figure 1c.

After obtaining a folded reflectivity model, we then add planar faulting in the model. Although all the faults are planar, the fault orientations (dip and strike) and displacements of the faults are all different from each other. The fault displacements on each fault are allowed to be spatially varying along the directions of fault strike and dip. The common patterns of fault displacement distribution have been discussed by some authors (Muraoka and Kamata, 1983; Mansfield and Cartwright, 1996; Stewart, 2001). In generating faults in our synthetic models, we define the fault displacement distributions as a Gaussian function or linear function. In the case of Gaussian distribution, the fault displacements decrease from the fault center in all directions along the fault plane. In the other case of linear distribution, the fault displacements linearly increase (normal fault) or decrease (reverse fault) in the fault dip direction along the fault plane. The maximum fault displacement for each fault is randomly chosen in the range between 0 and 40 samples. From our experience, the images with more faults are more effective than those with fewer faults to train a CNN for fault segmentation. We therefore add more than five faults within a training image with the size of 128 × 128 × 128. However, these faults should not be too close to each other as shown in Figure 1d, in which we have added six planar faults.

After creating a folded and faulted reflectivity model (Figure 1d), we finally convolve this model with a Ricker wavelet to obtain a 3D seismic image shown in Figure 1e. The peak frequency of the wavelet is also randomly chosen from a predefined range. Note that we convolve the reflectivity model with a wavelet after (not before) creating the folding and faulting in the model because the convolution will blur the sharp discontinuities near faults and therefore make the faults look more realistic. To further improve the realism of the synthetic seismic image, we also add some random noise to the image as shown in Figure 1f. From this noisy image (Figure 1f), we crop a final training seismic image (Figure 2a) with the size of 128 × 128 × 128 to avoid the artifacts near the boundaries. Figure 2b shows the corresponding binary fault labeling image, in which the faults are labeled by ones at two pixels adjacent to the faults from the hanging wall and footwall sides.

By using this workflow, we randomly choose parameters of folding, faulting, wavelet peak frequency, and noise to obtain 200 pairs of 3D unique seismic images and corresponding fault labeling images. We can actually generate much more unique training data sets, but we found that 200 pairs of images are sufficient to train a pretty good neural network for fault segmentation. Using the same workflow, we also automatically generated 20 pairs of seismic and fault labeling images for the validation.

Data augmentation

Creating unique training seismic and fault labeling images, as discussed above, is crucial to successfully train a fault segmentation neural network. Data augmentation during the training is also helpful to increase the diversity of the data sets and to prevent the neural network from learning irrelevant patterns. We apply simple data augmentations including vertical flip and rotation around the vertical time or depth axis. To avoid interpolation and artifacts near boundaries, we rotate the seismic and fault labeling volumes by only three options of 90°, 180°, and 270°. Because our input seismic and fault labeling volumes are 128 × 128 × 128 cubes, the flip and rotation will preserve the image size without needing interpolation or extrapolation. Note that we do not want to rotate the seismic and fault volumes around the inline or crossline axis because it will yield vertical seismic structures and flat faults, which are geologically unrealistic.

Fault Segmentation By CNN

We consider 3D fault detection as an image segmentation problem of labeling ones on faults, whereas zeros elsewhere in a 3D seismic image. We achieve such fault segmentation by using a simplified version of U-Net, an end-to-end fully CNN, which was first proposed by Ronneberger et al. (2015) for biomedical image segmentation, and then it was widely used for many other image segmentation problems. In 3D seismic images, the distribution of fault samples and nonfault samples is typically highly imbalanced; therefore, we use a balanced binary cross-entropy loss to optimize the parameters of the network as discussed by Xie and Tu (2015).

CNN architecture

We started our research on fault segmentation by using the original U-Net architecture (Ronneberger et al., 2015), which turned out to be more complicated than necessary for our problem of fault detection. We reduce the convolutional layers and features at each layer to save memory and computation but still preserve good performance in fault detection.

The finally simplified U-Net that we use for 3D fault detection is shown in Figure 3, in which an input 3D seismic image is fed to a network that contains a contracting path (left side) and an expansive path (right side) as in the original U-Net architecture. In the left contracting path, each step contains two 3 × 3 × 3 convolutional layers followed by a ReLU activation and a 2 × 2 × 2 max pooling operation with stride 2 for downsampling. We double the number of features after each step. Every step in the right expansion path contains a 2 × 2 × 2 upsampling operation, a concatenation with features from the left contracting path, and two 3 × 3 × 3 convolutional layers followed by a ReLU activation. Different from the original U-Net architecture, we do not include a 2 × 2 × 2 “up-convolution” layer after each upsampling as in the original expansion path. The upsampling operation is implemented by using the function UpSampling3D defined in Keras (Chollet, 2015). The final output layer is a 1 × 1 × 1 convolutional layer with a sigmoid activation to map each 16C feature vector to a probability value in the output fault probability map, which has the same size as the input seismic image. This simplified U-Net architecture totally consists of 15 convolutional layers, reduced from 23 convolutional layers in the original U-Net architecture. The number of features at these convolutional layers is also significantly reduced from the original architecture.

Balanced cross-entropy loss

The following binary cross-entropy loss function is widely used in the binary segmentation of a common image:

where N denotes the number of pixels in the input 3D seismic image. The term yi represents the true binary labels and pi represents the prediction probabilities (0 < pi < 1) computed from the sigmoid activation in the last convolutional layer. Because the true labels yi are binary values (0 or 1), the first term measures the prediction errors at the image pixels labeled by ones, whereas the second term measures the prediction errors at the pixels labeled by zeros.

This loss function works well for binary segmentation of common images, in which the distribution of zero/nonzero samples is more or less balanced. This loss function, however, is not suitable to measure the errors of fault segmentation, in which more than 90% of the samples are nonfault samples (labeled by zeros), whereas the fault samples (labeled by ones) are very limited. If we train the neural network using this loss function, the network can easily converge to the wrong direction and make zero predictions everywhere because zero prediction is a good solution to this loss function in the fault segmentation problem.

To solve this problem, we use the following balanced cross-entropy loss function as discussed by Xie and Tu (2015):

where represents the ratio between nonfault pixels and the total image pixels, whereas 1 − β denotes the ratio of fault pixels in the 3D seismic image.

Training and validation

We train the CNN by using 200 pairs of synthetic 3D seismic and fault images that are automatically created as in Figures 1 and 2. The validation data set contains another 20 pairs of such synthetic seismic and fault images, which are not used in the training data set. Considering the amplitude values of different real seismic images can be much different from each other, we normalize all the training seismic images, each image is subtracted by its mean value and divided by its standard deviation.

The size of each 3D seismic or fault image is 128 × 128 × 128. We choose this relatively small size because the memory of our GPU is limited to 12 GB. We would suggest to choose a larger size if the GPU memory is allowed. We feed the 3D seismic images to the neural network in batches and each batch contains four images, which consist of an original image and the same image rotated around the vertical time/depth axis by 90°, 180°, and 270°. We did not try a larger batch size, again, because of the GPU memory limitation. We use the Adam method (Kingma and Ba, 2014) to optimize the network parameters and set the learning rate to be 0.0001. We train the network with 25 epochs, and all the 200 training images are processed at each epoch. As shown in Figure 4, the training and validation accuracies gradually increase to 95%, whereas the training and validation loss converges to 0.01 after 25 epochs.

To verify the CNN model trained with 25 epochs, we apply this trained model together with another seven commonly used fault detection methods to the synthetic seismic volume (Figure 2a), which was not included in the training data sets. Figure 5a–5h shows all eight fault detection results that are, respectively, computed by using the methods of C3 (Gersztenkorn and Marfurt, 1999), C2 (Marfurt et al., 1999), planarity (Hale, 2009), structure-oriented linearity (Wu, 2017), structure-oriented semblance (Hale, 2009), fault likelihood (Hale, 2013; Wu and Hale, 2016), optimal surface voting (Wu and Fomel, 2018), and our CNN-based segmentation. The input for the optimal surface voting method is the planarity volume (Figure 5c), and the input for all the other methods is the amplitude volume (Figure 2a). Compared with the first five methods, the fault likelihood and optimal surface voting methods provide better fault detections in which the fault features are less noisy and can be more continuously tracked. Our CNN method achieves the best performance in computing an accurate, clean, and complete fault detection, which is most consistent with the true fault labeling shown in Figure 2b.

Figure 5. Fault detection on the synthetic validation volume (Figure 2a) by using (a-g) seven commonly used methods and (h) our CNN method. Compared with the methods of (a) C3 (Gersztenkorn and Marfurt, 1999), (b) C2 (Marfurt et al., 1999), (c) planarity (Hale, 2009), (d) structure-oriented planarity (Wu, 2017), (e) structure-oriented semblance (Hale, 2009), (f) the fault likelihood (Hale, 2013; Wu and Hale, 2016), and (g) optimal surface voting (Wu and Fomel, 2018) methods perform better fault detections. (h) Our CNN method achieves the best performance in obtaining an accurate, clean, and complete fault detection.

To quantitatively evaluate the fault detection methods, we further calculate the precision-recall (Martin et al., 2004) and receiver operating characteristic (ROC) (Provost et al., 1998) curves shown in Figure 6. From the precision-recall curves (Figure 6a), we can clearly observe that our CNN method (the red curve in Figure 6a) provides the highest precision for all choices of recall. The precisions of the fault likelihood (the orange curve in Figure 6a) and optimal surface voting (the magenta curve in Figure 6a) methods are relatively lower than the CNN method, but they are higher than the other five methods. The ROC curves in Figure 6b provide similar evaluations of the methods. In the next section, we will use the same CNN model (trained by only synthetic data sets) to four field seismic images that are acquired at different surveys. In this highly strict precision evaluation, the fault detections are expected to perfectly match the true fault labels with the thickness of only two samples. However, all of the methods should have higher precision if we consider that each fault is a thicker zone and all fault detections within the zone are good enough.

Figure 6. (a) Precision-recall and (b) ROC curves are used to evaluate the eight fault detections on the synthetic validation volume in Figure 5. Our CNN fault detection method performs significantly better than the other methods.

It might not be surprising that the CNN model, trained by synthetic data sets, works well to detect faults in a synthetic seismic image (Figure 5h) that is also created by using the same workflow for creating the training data sets. We further test the same CNN model on four field seismic images that are acquired at different surveys. To be consistent with the synthetic training seismic images, each of the field seismic images is subtracted by its mean value and divided by its standard deviation to obtain a consistently normalized image. We compare our fault prediction results with the thinned fault likelihood (Hale, 2013; Wu and Hale, 2016), which is a superior attribute (better than most of the conventional attributes [Figures 5 and 6]) for fault detection.

Figure 7. (a) A 3D seismic image is displayed with faults that are detected by using (b) the trained CNN model, (c) fault likelihood, and (d) thinned fault likelihood.

The first 3D seismic volume in Figure 7a is a subset (128 [vertical] × 384 [inline] × 512 [crossline] samples) extracted from the Netherlands off-shore F3 block seismic data, which is graciously provided by the Dutch government through TNO and dGB Earth Sciences. Multioriented faults are apparent within this 3D seismic volume. Figure 7b shows the fault probability image predicted by using the trained CNN model. The color in this fault image represents the fault probability, which is computed by the sigmoid activation in the last convolutional layer. Although trained by only synthetic data sets, this CNN model works pretty well to provide a clean and accurate prediction of faults in this field seismic image. In this CNN fault probability image, most fault features have very high probabilities (close to 1) and only very limited noisy features are observed. Although we added only planar faults in the training data sets, the networks actually learn to detect curved faults in the field seismic image as shown on the time slice in Figure 7b. Figure 7c and 7d, respectively, shows the fault likelihood attribute (Hale, 2013; Wu and Hale, 2016) before and after thinning. The thinned fault likelihood (Figure 7d) works fine to highlight the faults within this seismic image. However, a lot of more noisy features are observed than in the CNN fault probability image (Figure 7b). In addition, as denoted by the yellow arrows on the inline slice (Figure 7d), the fault-oriented smoothing in calculating the fault likelihood actually extends the fault features beyond the top of true faults. In addition, the fault likelihood is computed from the semblance/coherence of seismic reflections, which can be sensitive to noisy reflections (the red features on the crossline in Figure 7d) but insensitive to the faults with small fault displacements (like those faults denoted by white arrows in Figure 7d). However, the trained CNN model is more robust to noise and can better measure the probability of faults with small displacements.

Figure 8. (a) A 3D seismic image is displayed with faults that are detected by using (b) the trained CNN model and (c) thinned fault likelihood.

The second 3D seismic image shown in Figure 8a is graciously provided by Clyde Petroleum Plc. through Paradigm. Different from the previous synthetic and field examples, the faults in this seismic image are not apparent as sharp reflection discontinuities. Instead, the faults are imaged like reflections in this 3D seismic image as shown in Figure 8a. However, the CNN model still works pretty well to detect the faults shown in Figure 8b, which means that the network wisely learned to predict faults by not detecting sharp discontinuities or edges. Figure 8c shows the thinned fault likelihoods that are noisier than the CNN fault probabilities as shown on the horizontal slice.

Figure 9. Faults are detected in a complicated 3D example by using (a-c) the trained CNN model and (d-f) thinned fault likelihoods.

The third 3D seismic image shown in Figure 9 is a subset (210 [vertical] × 600 [inline] × 825 [crossline] samples) extracted from a larger seismic reflection volume that is acquired across the Costa Rica margin, northwest of the Osa Peninsula to image the fault properties in the subduction zone. Multiple sets of closely spaced faults are apparent in this 3D seismic volume as discussed by Bangs et al. (2015). The fault detection in this example is more challenging than the previous ones because the faults are very close to each other, the reflection structures are not well-imaged, and the image is pretty noisy. Figure 9a–9c shows the CNN fault probabilities at different slices. We observe that most faults are clearly labeled in this CNN fault probability images, and these faults can be continuously tracked by following the probability features. Multiple sets of faults striking in different directions can be clearly observed on the horizontal slice in these CNN fault probability images. Figure 9d–9f shows the thinned fault likelihoods at the same slices, which can detect most faults, but the fault features are much noisier than the CNN fault probabilities. In addition, many of the faults are mislabeled, especially in areas where the seismic structures are noisy.

Figure 10. A 3D seismic image overlaid with the CNN fault probabilities at different slices ((a) and (b)), where most of the faults are clearly and accurately labelled.

Figure 10 shows the fourth larger seismic volume (450 [vertical] × 1950 [inline] × 1200 [crossline] samples) that is acquired at the Campos Basin, offshore Brazil. This image shows that the sediments are heavily faulted due to the salt bodies at the bottom of the volume. The CNN fault probabilities shown in Figure 10a and 10b clearly and accurately label numerous closely spaced faults in this seismic volume. The faulting patterns are clearly visible on the time slices of the CNN fault probability image. To be able to better visualize the fault detection in this example, we further display two subvolumes of seismic amplitude and CNN fault probabilities (Figure 11b and 11d), in which most of the faults are clearly and accurately labeled except some subtle faults. The horizontal slices in Figure 11b and 11d, respectively, display clear patterns of polygonal and radial faults that may be associated with salt diapirs (Rowan et al., 1999; Carruthers, 2012).

Figure 11. Two subvolumes of (a and c) the seismic amplitude and (b and d) CNN fault probability are extracted from the full volumes in Figure 10.

In addition to the above field examples, we also applied the same trained CNN model to two other 3D seismic images Kerry-3D and Opunake-3D, which are provided on the SEG Wiki website. The fault segmentation results are clean and accurate as shown in the SEG Wiki website (Wu, 2018a, 2018b, 2019).

In summary, although the CNN model is trained by using only 200 synthetic seismic images, it works pretty well to detect faults in 3D field seismic volumes that are recorded at totally different surveys. In addition, the 3D fault prediction using the trained CNN model is highly efficient. By using one TITAN Xp GPU, computing the large CNN fault probability volume (450 [vertical] × 1950 [inline] × 1200 [crossline] samples) in Figure 10 takes less than 3 min. Computing fault likelihoods for the same volume, however, required approximately 1.5 h when using a workstation with 32 cores.

In our training and validating data sets, we avoid including thrust and listric faults with low dip angles. These faults often appear strong reflection features in a seismic image other than reflection discontinuities as the faults discussed in this paper. Therefore, all the conventional fault detection methods, based on measuring reflection discontinuity or continuity, often fail to detect the thrust and listric faults. However, our CNN-based method has potential to successfully detect these faults by training another specific model, which is actually what we will focus on in future research.


We have discussed an end-to-end CNN to efficiently detect faults from 3D seismic images, in which the fault detection is considered as a binary segmentation problem. This neural network is simplified from the originally more complicated U-Net to save GPU memory and computational time (for training and prediction) but still reserve the high performance for fault detection. Because the distribution of fault and nonfault samples is heavily biased, we use a balanced loss function to optimize the CNN model parameters. We train the neural network by using only 200 pairs of 3D synthetic seismic and fault volumes, which are all automatically generated by randomly adding folding, faulting, and noise in the volumes. Although trained by using only synthetic data sets, the neural network can accurately detect faults from 3D field seismic volumes that are acquired at totally different surveys. Although we add only planar faults in the training data sets, the network actually learns to detect curved faults in the field seismic images.


This research is financially supported by the sponsors of the Texas Consortium for Computation Seismology (TCCS). The 1st, 3rd and 4th authors gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan Xp GPU used for this research.

Aqrawi, A. A., and T. H. Boe, 2011, Improved fault segmentation using a dip guided and modified 3D Sobel filter: 81st Annual International Meeting, SEG, Expanded Abstracts, 999–1003, doi: 10.1190/1.3628241.

Badrinarayanan, V., A. Kendall, and R. Cipolla, 2017, Segnet: A deep convolutional encoder-decoder architecture for image segmentation: IEEE Transactions on Pattern Analysis and Machine Intelligence, 39, 2481–2495, doi: 10.1109/TPAMI.2016.2644615.

Bakker, P., 2002, Image structure analysis for seismic interpretation: Ph.D. thesis, Delft University of Technology.

Bangs, N. L., K. D. McIntosh, E. A. Silver, J. W. Kluesner, and C. R. Ranero, 2015, Fluid accumulation along the Costa Rica subduction thrust and development of the seismogenic zone: Journal of Geophysical Research: Solid Earth, 120, 67–86, doi: 10.1002/2014JB011265.

Carruthers, T., 2012, Interaction of polygonal fault systems with salt diapirs: Ph.D. thesis, Cardiff University.

Chollet, F., 2015, Keras,, accessed September 2018.

Cohen, I., N. Coult, and A. A. Vassiliou, 2006, Detection and extraction of fault surfaces in 3D seismic data: Geophysics, 71, no. 4, P21–P27, doi: 10.1190/1.2215357.

Di, H., M. Shafiq, and G. AlRegib, 2018, Patch-level MLP classification for improved fault detection: 88th Annual International Meeting, SEG, Expanded Abstracts, 2211–2215, doi: 10.1190/segam2018-2996921.1.

Gersztenkorn, A., and K. J. Marfurt, 1999, Eigenstructure-based coherence computations as an aid to 3-D structural and stratigraphic mapping: Geophysics, 64, 1468–1479, doi: 10.1190/1.1444651.

Girshick, R., J. Donahue, T. Darrell, and J. Malik, 2014, Rich feature hierarchies for accurate object detection and semantic segmentation: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 580–587.

Guitton, A., 2018, 3D convolutional neural networks for fault interpretation: 80th Annual International Conference and Exhibition, EAGE, Extended Abstracts, doi: 10.3997/2214-4609.201800732.

Guo, B., L. Li, and Y. Luo, 2018, A new method for automatic seismic fault detection using convolutional neural network: 88th Annual International Meeting, SEG, Expanded Abstracts, 1951–1955, doi: 10.1190/segam2018-2995894.1.

Hale, D., 2009, Structure-oriented smoothing and semblance: CWP Report 635.

Hale, D., 2013, Methods to compute fault images, extract fault surfaces, and estimate fault throws from 3D seismic images: Geophysics, 78, no. 2, O33-O43, doi: 10.1190/geo2012-0331.1.

He, K., G. Gkioxari, P. Dollár, and R. Girshick, 2017, Mask R-CNN: Proceedings of the IEEE International Conference on Computer Vision, 2980–2988

Huang, L., X. Dong, and T. E. Clee, 2017, A scalable deep learning platform for identifying geologic features from seismic attributes: The Leading Edge, 36, 249–256, doi: 10.1190/tle36030249.1.

Kingma, D. P., and J. Ba, 2014, Adam: A method for stochastic optimization: CoRR, abs/1412.6980.

Li, F., and W. Lu, 2014, Coherence attribute at different spectral scales: Interpretation, 2, no. 1, SA99–SA106, doi: 10.1190/INT-2013-0089.1.

Mansfield, C., and J. Cartwright, 1996, High resolution fault displacement mapping from three-dimensional seismic data: Evidence for dip linkage during fault growth: Journal of Structural Geology, 18, 249–263, doi: 10.1016/S0191-8141(96)80048-4.

Marfurt, K. J., R. L. Kirlin, S. L. Farmer, and M. S. Bahorich, 1998, 3-D seismic attributes using a semblance-based coherency algorithm: Geophysics, 63, 1150–1165, doi: 10.1190/1.1444415.

Marfurt, K. J., V. Sudhaker, A. Gersztenkorn, K. D. Crawford, and S. E. Nissen, 1999, Coherency calculations in the presence of structural dip: Geophysics, 64, 104–111, doi: 10.1190/1.1444508.

Martin, D. R., C. C. Fowlkes, and J. Malik, 2004, Learning to detect natural image boundaries using local brightness, color, and texture cues: IEEE Transactions on Pattern Analysis and Machine Intelligence, 26, 530–549, doi: 10.1109/TPAMI.2004.1273918.

Muraoka, H., and H. Kamata, 1983, Displacement distribution along minor fault traces: Journal of Structural Geology, 5, 483–495, doi: 10.1016/0191 8141(83)90054-8.

Neff, D. B., J. R. Grismore, and W. A. Lucas, 2000, Automated seismic fault detection and picking: U.S. Patent 6, 018, 498.

Pedersen, S. I., T. Randen, L. Sonneland, and Ø. Steen, 2002, Automatic fault extraction using artificial ants: 72nd Annual International Meeting, SEG, Expanded Abstracts, 512–515, doi: 10.1190/1.1817297.

Pedersen, S. I., T. Skov, A. Hetlelid, P. Fayemendy, T. Randen, and L. Sønneland, 2003, New paradigm of fault interpretation: 73rd Annual International Meeting, SEG, Expanded Abstracts, 350–353, doi: 10.1190/1.1817918.

Provost, F. J., T. Fawcett, and R. Kohavi, 1998, The case against accuracy estimation for comparing induction algorithms: Proceedings of the 15th International Conference on Machine Learning, 445–453.

Randen, T., S. I. Pedersen, and L. Sønneland, 2001, Automatic extraction of fault surfaces from three-dimensional seismic data: 81st Annual International Meeting, SEG, Expanded Abstracts, 551–554, doi: 10.1190/1.1816675.

Ren, S., K. He, R. Girshick, and J. Sun, 2015, Faster R-CNN: Towards real-time object detection with region proposal networks: Advances in Neural Information Processing Systems, 91–99.

Ronneberger, O., P. Fischer, and T. Brox, 2015, U-Net: Convolutional networks for biomedical image segmentation: International Conference on Medical Image Computing and Computer-Assisted Intervention, 234–241.

Rowan, M. G., M. P. Jackson, and B. D. Trudgill, 1999, Salt-related fault families and fault welds in the Northern Gulf of Mexico: AAPG Bulletin, 83, 1454–1484.

Stewart, S., 2001, Displacement distributions on extensional faults: Implications for fault stretch, linkage, and seal: AAPG Bulletin, 85, 587-599.

Van Bemmel, P. P., and R. E. Pepper, 2000, Seismic signal processing method and apparatus for generating a cube of variance values: U.S. Patent 6,151,555.

Wu, X., 2017, Directional structure-tensor based coherence to detect seismic faults and channels: Geophysics, 82, no. 2, A13–A17, doi: 10.1190/geo2016-0473.1.

Wu, X., 2018a, Kerry-3D,, accessed March 2019.

Wu, X., 2018b, Opunake-3D,, accessed March 2019

Wu, X., 2019, GitHub repository,, accessed March 2019

Wu, X., and S. Fomel, 2018, Automatic fault interpretation with optimal surface voting: Geophysics, 83, no. 5, O67–O82, doi: 10.1190/geo2018-0115.1.

Wu, X., and D. Hale, 2016, 3D seismic image processing for faults: Geophysics, 81, no. 2, IM1–IM11, doi: 10.1190/geo2015-0380.1.

Wu, X., Y. Shi, S. Fomel, and L. Liang, 2018, Convolutional neural networks for fault interpretation in seismic images: 88th Annual International Meeting, SEG, Expanded Abstracts, 1946–1950, doi: 10.1190/segam2018-2995341.1.

Wu, X., and Z. Zhu, 2017, Methods to enhance seismic faults and construct fault surfaces: Computers and Geosciences, 107, 37–48, doi: 10.1016/j.cageo.2017.06.015.

Xie, S., and Z. Tu, 2015, Holistically-nested edge detection: Proceedings of the IEEE International Conference on Computer Vision, 1395–1403.

Zhao, T., and P. Mukhopadhyay, 2018, A fault-detection workflow using deep learning and image processing: 88th Annual International Meeting, SEG, Expanded Abstracts, 1966–1970, doi: 10.1190/segam2018-2997005.1.


Request access by filling the form below to download full PDF.
Most Popular Papers
Case Study: An Integrated Machine Learning-Based Fault Classification Workflow
We build a fully automated machine learning-based fault detection workflow to compute the parameterized fault classification results ...
Case Study with Petrobras: Applying Unsupervised Multi-Attribute Machine Learning for 3D Stratigraphic Facies Classification in a Carbonate Field, Offshore Brazil
This study with Petrobras presents the results of a multi-attribute, machine learning study over a pre-salt carbonate field in the ...
Applying Machine Learning Technologies in the Niobrara Formation, DJ Basin, to Quickly Produce an Integrated Structural and Stratigraphic Seismic Classification Volume Calibrated to Wells
Carolan Laudon, Jie Qi, Yin-Kai Wang, Geophysical Research, LLC (d/b/a Geophysical Insights), University of Houston | Published with permission: Unconventional Resources ...
Shopping Cart
  • Registration confirmation will be emailed to you.

  • We're committed to your privacy. Geophysical Insights uses the information you provide to us to contact you about our relevant content, events, and products. You may unsubscribe from these communications at any time. For more information, check out our Privacy Policy

    Jan Van De MortelGeophysicist

    Jan Van De Mortel

    Jan is a geophysicist with a 30+ year international track record, including 20 years with Schlumberger, 4 years with Weatherford, and recent years actively involved in Machine Learning for both oilfield and non-oilfield applications. His work includes developing solutions and applications around transformer networks, probabilistic Machine Learning, etc. Jan currently works as a technical consultant at Geophysical Insights for Continental Europe, the Middle East, and Asia.

    Mike PowneyGeologist | Perceptum Ltd

    Mike Powney

    Mike began his career at SRC a consultancy formed from ECL where he worked extensively on seismic data offshore West Africa and the North Sea. Mike subsequently joined Geoex MCG where he provides global G&G technical expertise across their data portfolio. He also heads up the technical expertise within Geoex MCG on CCUS and natural hydrogen. Within his role at Perceptum, Mike leads the Machine Learning project investigating seismic and well data, offshore Equatorial Guinea.

    Tim GibbonsSales Representative

    Tim Gibbons

    Tim has a BA in Physics from the University of Oxford and an MSc in Exploration Geophysics from Imperial College, London. He started work as a geophysicist for BP in 1988 in London before moving to Aberdeen. There he also worked for Elf Exploration before his love of technology brought a move into the service sector in 1997. Since then, he has worked for Landmark, Paradigm, and TGS in a variety of managerial, sales, and business development roles. Since 2018, he has worked for Geophysical Insights, promoting Paradise throughout the European region.

    Dr. Carrie LaudonSenior Geophysical Consultant

    Applying Unsupervised Multi-Attribute Machine Learning for 3D Stratigraphic Facies Classification in a Carbonate Field, Offshore Brazil

    We present results of a multi-attribute, machine learning study over a pre-salt carbonate field in the Santos Basin, offshore Brazil. These results test the accuracy and potential of Self-organizing maps (SOM) for stratigraphic facies delineation. The study area has an existing detailed geological facies model containing predominantly reef facies in an elongated structure.

    Carrie LaudonSenior Geophysical Consultant - Geophysical Insights

    Automatic Fault Detection and Applying Machine Learning to Detect Thin Beds

    Rapid advances in Machine Learning (ML) are transforming seismic analysis. Using these new tools, geoscientists can accomplish the following quickly and effectively:

    • Run fault detection analysis in a few hours, not weeks
    • Identify thin beds down to a single seismic sample
    • Generate seismic volumes that capture structural and stratigraphic details

    Join us for a ‘Lunch & Learn’ sessions daily at 11:00 where Dr. Carolan (“Carrie”) Laudon will review the theory and results of applying a combination of machine learning tools to obtain the above results.  A detailed agenda follows.


    Automated Fault Detection using 3D CNN Deep Learning

    • Deep learning fault detection
    • Synthetic models
    • Fault image enhancement
    • Semi-supervised learning for visualization
    • Application results
      • Normal faults
      • Fault/fracture trends in complex reservoirs

    Demo of Paradise Fault Detection Thoughtflow®

    Stratigraphic analysis using machine learning with fault detection

    • Attribute Selection using Principal Component Analysis (PCA)
    • Multi-Attribute Classification using Self-Organizing Maps (SOM)
    • Case studies – stratigraphic analysis and fault detection
      • Fault-karst and fracture examples, China
      • Niobrara – Stratigraphic analysis and thin beds, faults
    Thomas ChaparroSenior Geophysicist - Geophysical Insights

    Paradise: A Day in The Life of the Geoscientist

    Over the last several years, the industry has invested heavily in Machine Learning (ML) for better predictions and automation. Dramatic results have been realized in exploration, field development, and production optimization. However, many of these applications have been single use ‘point’ solutions. There is a growing body of evidence that seismic analysis is best served using a combination of ML tools for a specific objective, referred to as ML Orchestration. This talk demonstrates how the Paradise AI workbench applications are used in an integrated workflow to achieve superior results than traditional interpretation methods or single-purpose ML products. Using examples from combining ML-based Fault Detection and Stratigraphic Analysis, the talk will show how ML orchestration produces value for exploration and field development by the interpreter leveraging ML orchestration.

    Aldrin RondonSenior Geophysical Engineer - Dragon Oil

    Machine Learning Fault Detection: A Case Study

    An innovative Fault Pattern Detection Methodology has been carried out using a combination of Machine Learning Techniques to produce a seismic volume suitable for fault interpretation in a structurally and stratigraphic complex field. Through theory and results, the main objective was to demonstrate that a combination of ML tools can generate superior results in comparison with traditional attribute extraction and data manipulation through conventional algorithms. The ML technologies applied are a supervised, deep learning, fault classification followed by an unsupervised, multi-attribute classification combining fault probability and instantaneous attributes.

    Thomas ChaparroSenior Geophysicist - Geophysical Insights

    Thomas Chaparro is a Senior Geophysicist who specializes in training and preparing AI-based workflows. Thomas also has experience as a processing geophysicist and 2D and 3D seismic data processing. He has participated in projects in the Gulf of Mexico, offshore Africa, the North Sea, Australia, Alaska, and Brazil.

    Thomas holds a bachelor’s degree in Geology from Northern Arizona University and a Master’s in Geophysics from the University of California, San Diego. His research focus was computational geophysics and seismic anisotropy.

    Aldrin RondonSenior Geophysical Engineer - Dragon Oil

    Bachelor’s Degree in Geophysical Engineering from Central University in Venezuela with a specialization in Reservoir Characterization from Simon Bolivar University.

    Over 20 years exploration and development geophysical experience with extensive 2D and 3D seismic interpretation including acquisition and processing.

    Aldrin spent his formative years working on exploration activity in PDVSA Venezuela followed by a period working for a major international consultant company in the Gulf of Mexico (Landmark, Halliburton) as a G&G consultant. Latterly he was working at Helix in Scotland, UK on producing assets in the Central and South North Sea.  From 2007 to 2021, he has been working as a Senior Seismic Interpreter in Dubai involved in different dedicated development projects in the Caspian Sea.

    Deborah SacreyOwner - Auburn Energy

    How to Use Paradise to Interpret Clastic Reservoirs

    The key to understanding Clastic reservoirs in Paradise starts with good synthetic ties to the wavelet data. If one is not tied correctly, then it will be easy to mis-interpret the neurons as reservoir, whin they are not. Secondly, the workflow should utilize Principal Component Analysis to better understand the zone of interest and the attributes to use in the SOM analysis. An important part to interpretation is understanding “Halo” and “Trailing” neurons as part of the stack around a reservoir or potential reservoir. Deep, high-pressured reservoirs often “leak” or have vertical percolation into the seal. This changes the rock properties enough in the seal to create a “halo” effect in SOM. Likewise, the frequency changes of the seismic can cause a subtle “dim-out”, not necessarily observable in the wavelet data, but enough to create a different pattern in the Earth in terms of these rock property changes. Case histories for Halo and trailing neural information include deep, pressured, Chris R reservoir in Southern Louisiana, Frio pay in Southeast Texas and AVO properties in the Yegua of Wharton County. Additional case histories to highlight interpretation include thin-bed pays in Brazoria County, including updated information using CNN fault skeletonization. Continuing the process of interpretation is showing a case history in Wharton County on using Low Probability to help explore Wilcox reservoirs. Lastly, a look at using Paradise to help find sweet spots in unconventional reservoirs like the Eagle Ford, a case study provided by Patricia Santigrossi.

    Mike DunnSr. Vice President of Business Development

    Machine Learning in the Cloud

    Machine Learning in the Cloud will address the capabilities of the Paradise AI Workbench, featuring on-demand access enabled by the flexible hardware and storage facilities available on Amazon Web Services (AWS) and other commercial cloud services. Like the on-premise instance, Paradise On-Demand provides guided workflows to address many geologic challenges and investigations. The presentation will show how geoscientists can accomplish the following workflows quickly and effectively using guided ThoughtFlows® in Paradise:

    • Identify and calibrate detailed stratigraphy using seismic and well logs
    • Classify seismic facies
    • Detect faults automatically
    • Distinguish thin beds below conventional tuning
    • Interpret Direct Hydrocarbon Indicators
    • Estimate reserves/resources

    Attend the talk to see how ML applications are combined through a process called "Machine Learning Orchestration," proven to extract more from seismic and well data than traditional means.

    Sarah Stanley
    Senior Geoscientist

    Stratton Field Case Study – New Solutions to Old Problems

    The Oligocene Frio gas-producing Stratton Field in south Texas is a well-known field. Like many onshore fields, the productive sand channels are difficult to identify using conventional seismic data. However, the productive channels can be easily defined by employing several Paradise modules, including unsupervised machine learning, Principal Component Analysis, Self-Organizing Maps, 3D visualization, and the new Well Log Cross Section and Well Log Crossplot tools. The Well Log Cross Section tool generates extracted seismic data, including SOMs, along the Cross Section boreholes and logs. This extraction process enables the interpreter to accurately identify the SOM neurons associated with pay versus neurons associated with non-pay intervals. The reservoir neurons can be visualized throughout the field in the Paradise 3D Viewer, with Geobodies generated from the neurons. With this ThoughtFlow®, pay intervals previously difficult to see in conventional seismic can finally be visualized and tied back to the well data.

    Laura Cuttill
    Practice Lead, Advertas

    Young Professionals – Managing Your Personal Brand to Level-up Your Career

    No matter where you are in your career, your online “personal brand” has a huge impact on providing opportunity for prospective jobs and garnering the respect and visibility needed for advancement. While geoscientists tackle ambitious projects, publish in technical papers, and work hard to advance their careers, often, the value of these isn’t realized beyond their immediate professional circle. Learn how to…

    • - Communicate who you are to high-level executives in exploration and development
    • - Avoid common social media pitfalls
    • - Optimize your online presence to best garner attention from recruiters
    • - Stay relevant
    • - Create content of interest
    • - Establish yourself as a thought leader in your given area of specialization
    Laura Cuttill
    Practice Lead, Advertas

    As a 20-year marketing veteran marketing in oil and gas and serial entrepreneur, Laura has deep experience in bringing technology products to market and growing sales pipeline. Armed with a marketing degree from Texas A&M, she began her career doing technical writing for Schlumberger and ExxonMobil in 2001. She started Advertas as a co-founder in 2004 and began to leverage her upstream experience in marketing. In 2006, she co-founded the cyber-security software company, 2FA Technology. After growing 2FA from a startup to 75% market share in target industries, and the subsequent sale of the company, she returned to Advertas to continue working toward the success of her clients, such as Geophysical Insights. Today, she guides strategy for large-scale marketing programs, manages project execution, cultivates relationships with industry media, and advocates for data-driven, account-based marketing practices.

    Fabian Rada
    Sr. Geophysicist, Petroleum Oil & Gas Services

    Statistical Calibration of SOM results with Well Log Data (Case Study)

    The first stage of the proposed statistical method has proven to be very useful in testing whether or not there is a relationship between two qualitative variables (nominal or ordinal) or categorical quantitative variables, in the fields of health and social sciences. Its application in the oil industry allows geoscientists not only to test dependence between discrete variables, but to measure their degree of correlation (weak, moderate or strong). This article shows its application to reveal the relationship between a SOM classification volume of a set of nine seismic attributes (whose vertical sampling interval is three meters) and different well data (sedimentary facies, Net Reservoir, and effective porosity grouped by ranges). The data were prepared to construct the contingency tables, where the dependent (response) variable and independent (explanatory) variable were defined, the observed frequencies were obtained, and the frequencies that would be expected if the variables were independent were calculated and then the difference between the two magnitudes was studied using the contrast statistic called Chi-Square. The second stage implies the calibration of the SOM volume extracted along the wellbore path through statistical analysis of the petrophysical properties VCL and PHIE, and SW for each neuron, which allowed to identify the neurons with the best petrophysical values in a carbonate reservoir.

    Heather Bedle
    Assistant Professor, University of Oklahoma

    Heather Bedle received a B.S. (1999) in physics from Wake Forest University, and then worked as a systems engineer in the defense industry. She later received a M.S. (2005) and a Ph. D. (2008) degree from Northwestern University. After graduate school, she joined Chevron and worked as both a development geologist and geophysicist in the Gulf of Mexico before joining Chevron’s Energy Technology Company Unit in Houston, TX. In this position, she worked with the Rock Physics from Seismic team analyzing global assets in Chevron’s portfolio. Dr. Bedle is currently an assistant professor of applied geophysics at the University of Oklahoma’s School of Geosciences. She joined OU in 2018, after instructing at the University of Houston for two years. Dr. Bedle and her student research team at OU primarily work with seismic reflection data, using advanced techniques such as machine learning, attribute analysis, and rock physics to reveal additional structural, stratigraphic and tectonic insights of the subsurface.

    Jie Qi
    Research Geophysicist

    An Integrated Fault Detection Workflow

    Seismic fault detection is one of the top critical procedures in seismic interpretation. Identifying faults are significant for characterizing and finding the potential oil and gas reservoirs. Seismic amplitude data exhibiting good resolution and a high signal-to-noise ratio are key to identifying structural discontinuities using seismic attributes or machine learning techniques, which in turn serve as input for automatic fault extraction. Deep learning Convolutional Neural Networks (CNN) performs well on fault detection without any human-computer interactive work. This study shows an integrated CNN-based fault detection workflow to construct fault images that are sufficiently smooth for subsequent fault automatic extraction. The objectives were to suppress noise or stratigraphic anomalies subparallel to reflector dip, and sharpen fault and other discontinuities that cut reflectors, preconditioning the fault images for subsequent automatic extraction. A 2D continuous wavelet transform-based acquisition footprint suppression method was applied time slice by time slice to suppress wavenumber components to avoid interpreting the acquisition footprint as artifacts by the CNN fault detection method. To further suppress cross-cutting noise as well as sharpen fault edges, a principal component edge-preserving structure-oriented filter is also applied. The conditioned amplitude volume is then fed to a pre-trained CNN model to compute fault probability. Finally, a Laplacian of Gaussian filter is applied to the original CNN fault probability to enhance fault images. The resulting fault probability volume is favorable with respect to traditional human-interpreter generated on vertical slices through the seismic amplitude volume.

    Dr. Jie Qi
    Research Geophysicist

    An integrated machine learning-based fault classification workflow

    We introduce an integrated machine learning-based fault classification workflow that creates fault component classification volumes that greatly reduces the burden on the human interpreter. We first compute a 3D fault probability volume from pre-conditioned seismic amplitude data using a 3D convolutional neural network (CNN). However, the resulting “fault probability” volume delineates other non-fault edges such as angular unconformities, the base of mass transport complexes, and noise such as acquisition footprint. We find that image processing-based fault discontinuity enhancement and skeletonization methods can enhance the fault discontinuities and suppress many of the non-fault discontinuities. Although each fault is characterized by its dip and azimuth, these two properties are discontinuous at azimuths of φ=±180° and for near vertical faults for azimuths φ and φ+180° requiring them to be parameterized as four continuous geodetic fault components. These four fault components as well as the fault probability can then be fed into a self-organizing map (SOM) to generate fault component classification. We find that the final classification result can segment fault sets trending in interpreter-defined orientations and minimize the impact of stratigraphy and noise by selecting different neurons from the SOM 2D neuron color map.

    Ivan Marroquin
    Senior Research Geophysicist

    Connecting Multi-attribute Classification to Reservoir Properties

    Interpreters rely on seismic pattern changes to identify and map geologic features of importance. The ability to recognize such features depends on the seismic resolution and characteristics of seismic waveforms. With the advancement of machine learning algorithms, new methods for interpreting seismic data are being developed. Among these algorithms, self-organizing maps (SOM) provides a different approach to extract geological information from a set of seismic attributes.

    SOM approximates the input patterns by a finite set of processing neurons arranged in a regular 2D grid of map nodes. Such that, it classifies multi-attribute seismic samples into natural clusters following an unsupervised approach. Since machine learning is unbiased, so the classifications can contain both geological information and coherent noise. Thus, seismic interpretation evolves into broader geologic perspectives. Additionally, SOM partitions multi-attribute samples without a priori information to guide the process (e.g., well data).

    The SOM output is a new seismic attribute volume, in which geologic information is captured from the classification into winning neurons. Implicit and useful geological information are uncovered through an interactive visual inspection of winning neuron classifications. By doing so, interpreters build a classification model that aids them to gain insight into complex relationships between attribute patterns and geological features.

    Despite all these benefits, there are interpretation challenges regarding whether there is an association between winning neurons and geological features. To address these issues, a bivariate statistical approach is proposed. To evaluate this analysis, three cases scenarios are presented. In each case, the association between winning neurons and net reservoir (determined from petrophysical or well log properties) at well locations is analyzed. The results show that the statistical analysis not only aid in the identification of classification patterns; but more importantly, reservoir/not reservoir classification by classical petrophysical analysis strongly correlates with selected SOM winning neurons. Confidence in interpreted classification features is gained at the borehole and interpretation is readily extended as geobodies away from the well.

    Heather Bedle
    Assistant Professor, University of Oklahoma

    Gas Hydrates, Reefs, Channel Architecture, and Fizz Gas: SOM Applications in a Variety of Geologic Settings

    Students at the University of Oklahoma have been exploring the uses of SOM techniques for the last year. This presentation will review learnings and results from a few of these research projects. Two projects have investigated the ability of SOMs to aid in identification of pore space materials – both trying to qualitatively identify gas hydrates and under-saturated gas reservoirs. A third study investigated individual attributes and SOMs in recognizing various carbonate facies in a pinnacle reef in the Michigan Basin. The fourth study took a deep dive of various machine learning algorithms, of which SOMs will be discussed, to understand how much machine learning can aid in the identification of deepwater channel architectures.

    Fabian Rada
    Sr. Geophysicist, Petroleum Oil & Gas Servicest

    Fabian Rada joined Petroleum Oil and Gas Services, Inc (POGS) in January 2015 as Business Development Manager and Consultant to PEMEX. In Mexico, he has participated in several integrated oil and gas reservoir studies. He has consulted with PEMEX Activos and the G&G Technology group to apply the Paradise AI workbench and other tools. Since January 2015, he has been working with Geophysical Insights staff to provide and implement the multi-attribute analysis software Paradise in Petróleos Mexicanos (PEMEX), running a successful pilot test in Litoral Tabasco Tsimin Xux Asset. Mr. Rada began his career in the Venezuelan National Foundation for Seismological Research, where he participated in several geophysical projects, including seismic and gravity data for micro zonation surveys. He then joined China National Petroleum Corporation (CNPC) as QC Geophysicist until he became the Chief Geophysicist in the QA/QC Department. Then, he transitioned to a subsidiary of Petróleos de Venezuela (PDVSA), as a member of the QA/QC and Chief of Potential Field Methods section. Mr. Rada has also participated in processing land seismic data and marine seismic/gravity acquisition surveys. Mr. Rada earned a B.S. in Geophysics from the Central University of Venezuela.

    Hal GreenDirector, Marketing & Business Development - Geophysical Insights

    Introduction to Automatic Fault Detection and Applying Machine Learning to Detect Thin Beds

    Rapid advances in Machine Learning (ML) are transforming seismic analysis. Using these new tools, geoscientists can accomplish the following quickly and effectively: a combination of machine learning (ML) and deep learning applications, geoscientists apply Paradise to extract greater insights from seismic and well data for these and other objectives:

    • Run fault detection analysis in a few hours, not weeks
    • Identify thin beds down to a single seismic sample
    • Overlay fault images on stratigraphic analysis

    The brief introduction will orient you with the technology and examples of how machine learning is being applied to automate interpretation while generating new insights in the data.

    Sarah Stanley
    Senior Geoscientist and Lead Trainer

    Sarah Stanley joined Geophysical Insights in October, 2017 as a geoscience consultant, and became a full-time employee July 2018. Prior to Geophysical Insights, Sarah was employed by IHS Markit in various leadership positions from 2011 to her retirement in August 2017, including Director US Operations Training and Certification, the Operational Governance Team, and, prior to February 2013, Director of IHS Kingdom Training. Sarah joined SMT in May, 2002, and was the Director of Training for SMT until IHS Markit’s acquisition in 2011.

    Prior to joining SMT Sarah was employed by GeoQuest, a subdivision of Schlumberger, from 1998 to 2002. Sarah was also Director of the Geoscience Technology Training Center, North Harris College from 1995 to 1998, and served as a voluntary advisor on geoscience training centers to various geological societies. Sarah has over 37 years of industry experience and has worked as a petroleum geoscientist in various domestic and international plays since August of 1981. Her interpretation experience includes tight gas sands, coalbed methane, international exploration, and unconventional resources.

    Sarah holds a Bachelor’s of Science degree with majors in Biology and General Science and minor in Earth Science, a Master’s of Arts in Education and Master’s of Science in Geology from Ball State University, Muncie, Indiana. Sarah is both a Certified Petroleum Geologist, and a Registered Geologist with the State of Texas. Sarah holds teaching credentials in both Indiana and Texas.

    Sarah is a member of the Houston Geological Society and the American Association of Petroleum Geologists, where she currently serves in the AAPG House of Delegates. Sarah is a recipient of the AAPG Special Award, the AAPG House of Delegates Long Service Award, and the HGS President’s award for her work in advancing training for petroleum geoscientists. She has served on the AAPG Continuing Education Committee and was Chairman of the AAPG Technical Training Center Committee. Sarah has also served as Secretary of the HGS, and Served two years as Editor for the AAPG Division of Professional Affairs Correlator.

    Dr. Tom Smith
    President & CEO

    Dr. Tom Smith received a BS and MS degree in Geology from Iowa State University. His graduate research focused on a shallow refraction investigation of the Manson astrobleme. In 1971, he joined Chevron Geophysical as a processing geophysicist but resigned in 1980 to complete his doctoral studies in 3D modeling and migration at the Seismic Acoustics Lab at the University of Houston. Upon graduation with the Ph.D. in Geophysics in 1981, he started a geophysical consulting practice and taught seminars in seismic interpretation, seismic acquisition and seismic processing. Dr. Smith founded Seismic Micro-Technology in 1984 to develop PC software to support training workshops which subsequently led to development of the KINGDOM Software Suite for integrated geoscience interpretation with world-wide success.

    The Society of Exploration Geologists (SEG) recognized Dr. Smith’s work with the SEG Enterprise Award in 2000, and in 2010, the Geophysical Society of Houston (GSH) awarded him an Honorary Membership. Iowa State University (ISU) has recognized Dr. Smith throughout his career with the Distinguished Alumnus Lecturer Award in 1996, the Citation of Merit for National and International Recognition in 2002, and the highest alumni honor in 2015, the Distinguished Alumni Award. The University of Houston College of Natural Sciences and Mathematics recognized Dr. Smith with the 2017 Distinguished Alumni Award.

    In 2009, Dr. Smith founded Geophysical Insights, where he leads a team of geophysicists, geologists and computer scientists in developing advanced technologies for fundamental geophysical problems. The company launched the Paradise® multi-attribute analysis software in 2013, which uses Machine Learning and pattern recognition to extract greater information from seismic data.

    Dr. Smith has been a member of the SEG since 1967 and is a professional member of SEG, GSH, HGS, EAGE, SIPES, AAPG, Sigma XI, SSA and AGU. Dr. Smith served as Chairman of the SEG Foundation from 2010 to 2013. On January 25, 2016, he was recognized by the Houston Geological Society (HGS) as a geophysicist who has made significant contributions to the field of geology. He currently serves on the SEG President-Elect’s Strategy and Planning Committee and the ISU Foundation Campaign Committee for Forever True, For Iowa State.

    Carrie LaudonSenior Geophysical Consultant - Geophysical Insights

    Applying Machine Learning Technologies in the Niobrara Formation, DJ Basin, to Quickly Produce an Integrated Structural and Stratigraphic Seismic Classification Volume Calibrated to Wells

    This study will demonstrate an automated machine learning approach for fault detection in a 3D seismic volume. The result combines Deep Learning Convolution Neural Networks (CNN) with a conventional data pre-processing step and an image processing-based post processing approach to produce high quality fault attribute volumes of fault probability, fault dip magnitude and fault dip azimuth. These volumes are then combined with instantaneous attributes in an unsupervised machine learning classification, allowing the isolation of both structural and stratigraphic features into a single 3D volume. The workflow is illustrated on a 3D seismic volume from the Denver Julesburg Basin and a statistical analysis is used to calibrate results to well data.

    Ivan Marroquin
    Senior Research Geophysicist

    Iván Dimitri Marroquín is a 20-year veteran of data science research, consistently publishing in peer-reviewed journals and speaking at international conference meetings. Dr. Marroquín received a Ph.D. in geophysics from McGill University, where he conducted and participated in 3D seismic research projects. These projects focused on the development of interpretation techniques based on seismic attributes and seismic trace shape information to identify significant geological features or reservoir physical properties. Examples of his research work are attribute-based modeling to predict coalbed thickness and permeability zones, combining spectral analysis with coherency imagery technique to enhance interpretation of subtle geologic features, and implementing a visual-based data mining technique on clustering to match seismic trace shape variability to changes in reservoir properties.

    Dr. Marroquín has also conducted some ground-breaking research on seismic facies classification and volume visualization. This lead to his development of a visual-based framework that determines the optimal number of seismic facies to best reveal meaningful geologic trends in the seismic data. He proposed seismic facies classification as an alternative to data integration analysis to capture geologic information in the form of seismic facies groups. He has investigated the usefulness of mobile devices to locate, isolate, and understand the spatial relationships of important geologic features in a context-rich 3D environment. In this work, he demonstrated mobile devices are capable of performing seismic volume visualization, facilitating the interpretation of imaged geologic features.  He has definitively shown that mobile devices eventually will allow the visual examination of seismic data anywhere and at any time.

    In 2016, Dr. Marroquín joined Geophysical Insights as a senior researcher, where his efforts have been focused on developing machine learning solutions for the oil and gas industry. For his first project, he developed a novel procedure for lithofacies classification that combines a neural network with automated machine methods. In parallel, he implemented a machine learning pipeline to derive cluster centers from a trained neural network. The next step in the project is to correlate lithofacies classification to the outcome of seismic facies analysis.  Other research interests include the application of diverse machine learning technologies for analyzing and discerning trends and patterns in data related to oil and gas industry.

    Dr. Jie Qi
    Research Geophysicist

    Dr. Jie Qi is a Research Geophysicist at Geophysical Insights, where he works closely with product development and geoscience consultants. His research interests include machine learning-based fault detection, seismic interpretation, pattern recognition, image processing, seismic attribute development and interpretation, and seismic facies analysis. Dr. Qi received a BS (2011) in Geoscience from the China University of Petroleum in Beijing, and an MS (2013) in Geophysics from the University of Houston. He earned a Ph.D. (2017) in Geophysics from the University of Oklahoma, Norman. His industry experience includes work as a Research Assistant (2011-2013) at the University of Houston and the University of Oklahoma (2013-2017). Dr. Qi was with Petroleum Geo-Services (PGS), Inc. in 2014 as a summer intern, where he worked on a semi-supervised seismic facies analysis. In 2017, he served as a postdoctoral Research Associate in the Attributed Assisted-Seismic Processing and Interpretation (AASPI) consortium at the University of Oklahoma from 2017 to 2020.

    Rocky R. Roden
    Senior Consulting Geophysicist

    The Relationship of Self-Organization, Geology, and Machine Learning

    Self-organization is the nonlinear formation of spatial and temporal structures, patterns or functions in complex systems (Aschwanden et al., 2018). Simple examples of self-organization include flocks of birds, schools of fish, crystal development, formation of snowflakes, and fractals. What these examples have in common is the appearance of structure or patterns without centralized control. Self-organizing systems are typically governed by power laws, such as the Gutenberg-Richter law of earthquake frequency and magnitude. In addition, the time frames of such systems display a characteristic self-similar (fractal) response, where earthquakes or avalanches for example, occur over all possible time scales (Baas, 2002).

    The existence of nonlinear dynamic systems and ordered structures in the earth are well known and have been studied for centuries and can appear as sedimentary features, layered and folded structures, stratigraphic formations, diapirs, eolian dune systems, channelized fluvial and deltaic systems, and many more (Budd, et al., 2014; Dietrich and Jacob, 2018). Each of these geologic processes and features exhibit patterns through the action of undirected local dynamics and is generally termed “self-organization” (Paola, 2014).

    Artificial intelligence and specifically neural networks exhibit and reveal self-organization characteristics. The reason for the interest in applying neural networks stems from the fact that they are universal approximators for various kinds of nonlinear dynamical systems of arbitrary complexity (Pessa, 2008). A special class of artificial neural networks is aptly named self-organizing map (SOM) (Kohonen, 1982). It has been found that SOM can identify significant organizational structure in the form of clusters from seismic attributes that relate to geologic features (Strecker and Uden, 2002; Coleou et al., 2003; de Matos, 2006; Roy et al., 2013; Roden et al., 2015; Zhao et al., 2016; Roden et al., 2017; Zhao et al., 2017; Roden and Chen, 2017; Sacrey and Roden, 2018; Leal et al, 2019; Hussein et al., 2020; Hardage et al., 2020; Manauchehri et al., 2020). As a consequence, SOM is an excellent machine learning neural network approach utilizing seismic attributes to help identify self-organization features and define natural geologic patterns not easily seen or seen at all in the data.

    Rocky R. Roden
    Senior Consulting Geophysicist

    Rocky R. Roden started his own consulting company, Rocky Ridge Resources Inc. in 2003 and works with several oil companies on technical and prospect evaluation issues. He is also a principal in the Rose and Associates DHI Risk Analysis Consortium and was Chief Consulting Geophysicist with Seismic Micro-technology. Rocky is a proven oil finder with 37 years in the industry, gaining extensive knowledge of modern geoscience technical approaches.

    Rocky holds a BS in Oceanographic Technology-Geology from Lamar University and a MS in Geological and Geophysical Oceanography from Texas A&M University. As Chief Geophysicist and Director of Applied Technology for Repsol-YPF, his role comprised of advising corporate officers, geoscientists, and managers on interpretation, strategy and technical analysis for exploration and development in offices in the U.S., Argentina, Spain, Egypt, Bolivia, Ecuador, Peru, Brazil, Venezuela, Malaysia, and Indonesia. He has been involved in the technical and economic evaluation of Gulf of Mexico lease sales, farmouts worldwide, and bid rounds in South America, Europe, and the Far East. Previous work experience includes exploration and development at Maxus Energy, Pogo Producing, Decca Survey, and Texaco. Rocky is a member of SEG, AAPG, HGS, GSH, EAGE, and SIPES; he is also a past Chairman of The Leading Edge Editorial Board.

    Bob A. Hardage

    Bob A. Hardage received a PhD in physics from Oklahoma State University. His thesis work focused on high-velocity micro-meteoroid impact on space vehicles, which required trips to Goddard Space Flight Center to do finite-difference modeling on dedicated computers. Upon completing his university studies, he worked at Phillips Petroleum Company for 23 years and was Exploration Manager for Asia and Latin America when he left Phillips. He moved to WesternAtlas and worked 3 years as Vice President of Geophysical Development and Marketing. He then established a multicomponent seismic research laboratory at the Bureau of Economic Geology and served The University of Texas at Austin as a Senior Research Scientist for 28 years. He has published books on VSP, cross-well profiling, seismic stratigraphy, and multicomponent seismic technology. He was the first person to serve 6 years on the Board of Directors of the Society of Exploration Geophysicists (SEG). His Board service was as SEG Editor (2 years), followed by 1-year terms as First VP, President Elect, President, and Past President. SEG has awarded him a Special Commendation, Life Membership, and Honorary Membership. He wrote the AAPG Explorer column on geophysics for 6 years. AAPG honored him with a Distinguished Service award for promoting geophysics among the geological community.

    Bob A. Hardage

    Investigating the Internal Fabric of VSP data with Attribute Analysis and Unsupervised Machine Learning

    Examination of vertical seismic profile (VSP) data with unsupervised machine learning technology is a rigorous way to compare the fabric of down-going, illuminating, P and S wavefields with the fabric of up-going reflections and interbed multiples created by these wavefields. This concept is introduced in this paper by applying unsupervised learning to VSP data to better understand the physics of P and S reflection seismology. The zero-offset VSP data used in this investigation were acquired in a hard-rock, fast-velocity, environment that caused the shallowest 2 or 3 geophones to be inside the near-field radiation zone of a vertical-vibrator baseplate. This study shows how to use instantaneous attributes to backtrack down-going direct-P and direct-S illuminating wavelets to the vibrator baseplate inside the near-field zone. This backtracking confirms that the points-of-origin of direct-P and direct-S are identical. The investigation then applies principal component (PCA) analysis to VSP data and shows that direct-S and direct-P wavefields that are created simultaneously at a vertical-vibrator baseplate have the same dominant principal components. A self-organizing map (SOM) approach is then taken to illustrate how unsupervised machine learning describes the fabric of down-going and up-going events embedded in vertical-geophone VSP data. These SOM results show that a small number of specific neurons build the down-going direct-P illuminating wavefield, and another small group of neurons build up-going P primary reflections and early-arriving down-going P multiples. The internal attribute fabric of these key down-going and up-going neurons are then compared to expose their similarities and differences. This initial study indicates that unsupervised machine learning, when applied to VSP data, is a powerful tool for understanding the physics of seismic reflectivity at a prospect. This research strategy of analyzing VSP data with unsupervised machine learning will now expand to horizontal-geophone VSP data.

    Tom Smith
    President and CEO, Geophysical Insights

    Machine Learning for Incomplete Geoscientists

    This presentation covers big-picture machine learning buzz words with humor and unassailable frankness. The goal of the material is for every geoscientist to gain confidence in these important concepts and how they add to our well-established practices, particularly seismic interpretation. Presentation topics include a machine learning historical perspective, what makes it different, a fish factory, Shazam, comparison of supervised and unsupervised machine learning methods with examples, tuning thickness, deep learning, hard/soft attribute spaces, multi-attribute samples, and several interpretation examples. After the presentation, you may not know how to run machine learning algorithms, but you should be able to appreciate their value and avoid some of their limitations.

    Deborah Sacrey
    Owner, Auburn Energy

    Deborah is a geologist/geophysicist with 44 years of oil and gas exploration experience in Texas, Louisiana Gulf Coast and Mid-Continent areas of the US. She received her degree in Geology from the University of Oklahoma in 1976 and immediately started working for Gulf Oil in their Oklahoma City offices.

    She started her own company, Auburn Energy, in 1990 and built her first geophysical workstation using Kingdom software in 1996. She helped SMT/IHS for 18 years in developing and testing the Kingdom Software. She specializes in 2D and 3D interpretation for clients in the US and internationally. For the past nine years she has been part of a team to study and bring the power of multi-attribute neural analysis of seismic data to the geoscience public, guided by Dr. Tom Smith, founder of SMT. She has become an expert in the use of Paradise software and has seven discoveries for clients using multi-attribute neural analysis.

    Deborah has been very active in the geological community. She is past national President of SIPES (Society of Independent Professional Earth Scientists), past President of the Division of Professional Affairs of AAPG (American Association of Petroleum Geologists), Past Treasurer of AAPG and Past President of the Houston Geological Society. She is also Past President of the Gulf Coast Association of Geological Societies and just ended a term as one of the GCAGS representatives on the AAPG Advisory Council. Deborah is also a DPA Certified Petroleum Geologist #4014 and DPA Certified Petroleum Geophysicist #2. She belongs to AAPG, SIPES, Houston Geological Society, South Texas Geological Society and the Oklahoma City Geological Society (OCGS).

    Mike Dunn
    Senior Vice President Business Development

    Michael A. Dunn is an exploration executive with extensive global experience including the Gulf of Mexico, Central America, Australia, China and North Africa. Mr. Dunn has a proven a track record of successfully executing exploration strategies built on a foundation of new and innovative technologies. Currently, Michael serves as Senior Vice President of Business Development for Geophysical Insights.

    He joined Shell in 1979 as an exploration geophysicist and party chief and held increasing levels or responsibility including Manager of Interpretation Research. In 1997, he participated in the launch of Geokinetics, which completed an IPO on the AMEX in 2007. His extensive experience with oil companies (Shell and Woodside) and the service sector (Geokinetics and Halliburton) has provided him with a unique perspective on technology and applications in oil and gas. Michael received a B.S. in Geology from Rutgers University and an M.S. in Geophysics from the University of Chicago.

    Hal GreenDirector, Marketing & Business Development - Geophysical Insights

    Hal H. Green is a marketing executive and entrepreneur in the energy industry with more than 25 years of experience in starting and managing technology companies. He holds a B.S. in Electrical Engineering from Texas A&M University and an MBA from the University of Houston. He has invested his career at the intersection of marketing and technology, with a focus on business strategy, marketing, and effective selling practices. Mr. Green has a diverse portfolio of experience in marketing technology to the hydrocarbon supply chain – from upstream exploration through downstream refining & petrochemical. Throughout his career, Mr. Green has been a proven thought-leader and entrepreneur, while supporting several tech start-ups.

    He started his career as a process engineer in the semiconductor manufacturing industry in Dallas, Texas and later launched an engineering consulting and systems integration business. Following the sale of that business in the late 80’s, he joined Setpoint in Houston, Texas where he eventually led that company’s Manufacturing Systems business. Aspen Technology acquired Setpoint in January 1996 and Mr. Green continued as Director of Business Development for the Information Management and Polymer Business Units.

    In 2004, Mr. Green founded Advertas, a full-service marketing and public relations firm serving clients in energy and technology. In 2010, Geophysical Insights retained Advertas as their marketing firm. Dr. Tom Smith, President/CEO of Geophysical Insights, soon appointed Mr. Green as Director of Marketing and Business Development for Geophysical Insights, in which capacity he still serves today.

    Hana Kabazi
    Product Manager

    Hana Kabazi joined Geophysical Insights in October of 201, and is now one of our Product Managers for Paradise. Mrs. Kabazi has over 7 years of oil and gas experience, including 5 years and Halliburton – Landmark. During her time at Landmark she held positions as a consultant to many E&P companies, technical advisor to the QA organization, and as product manager of Subsurface Mapping in DecsionSpace. Mrs. Kabazi has a B.S. in Geology from the University of Texas Austin, and an M.S. in Geology from the University of Houston.

    Dr. Carrie LaudonSenior Geophysical Consultant - Geophysical Insights

    Carolan (Carrie) Laudon holds a PhD in geophysics from the University of Minnesota and a BS in geology from the University of Wisconsin Eau Claire. She has been Senior Geophysical Consultant with Geophysical Insights since 2017 working with Paradise®, their machine learning platform. Prior roles include Vice President of Consulting Services and Microseismic Technology for Global Geophysical Services and 17 years with Schlumberger in technical, management and sales, starting in Alaska and including Aberdeen, Scotland, Houston, TX, Denver, CO and Reading, England. She spent five years early in her career with ARCO Alaska as a seismic interpreter for the Central North Slope exploration team.