GEOLOGICAL/GEOPHYSICAL MODELING

Using gas chimneys as an exploration tool

Part 1 – The principles of gas chimney interpretation provide the rationale for using them for hydrocarbon exploration. Concluding installment will present actual field case histories

F. Aminzadeh, dGB-USA; P. de Groot, dGB BV; T. Berge, Forest Oil Co.; and G. Valenti, AGIP

his two-part series will highlight some applications for using gas chimneys in hydrocarbon exploration. Among them are: unraveling a basin’s hydrocarbon history; establishing the migration path; distinguishing between charged and noncharged prospects, or sealing versus nonsealing faults; identifying potential for over-pressure; and detecting geo-hazards.

This installment discusses concepts and procedures for interpreting gas chimney data, including principles such as attribute extraction, training neural networks and identifying chimneys.

Introduction

Before the discovery of hydrocarbons, surface oil and gas seeps were known in many hydrocarbon-rich areas of the world. Such seeps supported flames that ranged from ancient Persia to South America, playing a role in various cultures for thousands of years. Flames continue to burn today, Fig. 1. Those fires, for the most part, were fueled by seeps via subsurface "gas chimneys."

Fig. 1. Burning gas seeps near Baku, Azerbaijan. (Courtesy of Prof. Gulyev, Azerbaijan Academy of Science.)

Explorers such as Drake in the U.S., Darcy in Iran and Nobel in Azerbaijan used these surface seeps to aid in the discovery of oil fields. In fact, before the introduction of gravity, magnetic and seismic methods, natural hydrocarbon seepage and corresponding surface evidence was the most common exploration tool. As a result, most major oil fields discovered in the first 60 years of the oil industry were close to known seepage locations.

Surface geochemistry methods developed during the last 30 years analyze gas concentrations in the seawater or soil. Recent advances can reveal minor seepage with limited areal extent. Likewise, gas clouds or mud volcanoes on seismic sections are evidenced by the chaotic behavior of waveforms.

The newly introduced gas "chimney cube" detects subtle gas clouds and gas-migration effects. Chimney cubes can be used as a powerful tool, especially when combined with remote sensing data, structural / stratigraphic models and geologic features, such as pockmarks and evidence of mud volcanoes.

Concept and Procedure

The new concept – a seismic entity – is called a chimney cube. A chimney cube is a 3-D volume of seismic data that highlights vertical, chaotic seismic characters. These disturbances often are associated with gas chimneys. The cube facilitates the difficult task of manual interpretation of gas chimneys.

In practice, chimney cubes can reveal where hydrocarbons originated, how they migrated into a prospect and how they spilled from that prospect and/or created shallow gas, mud volcanoes or pock marks on the seafloor. As such, a chimney cube can be seen as a new, indirect hydrocarbon-indicator tool. Through this process, a seismic volume (and corresponding attributes) is provided as input to a neural network, and a chimney-cube model is generated as its output, Fig. 2. The procedure involves:

Fig. 2. A typical chimney-cube as an output of the trained network and input seismic volume (only a cross section of the input and output is shown). Click for enlarged view

• Calculating and identifying sets of single- and multi-trace seismic attributes that distinguish between chimneys and nonchimneys
• Designing and training a neural network with attributes extracted at interpreted chimneys and nonchimney locations
• Creating a chimney-cube volume from a multi-attribute transformation of the 3-D seismic volume, highlighting vertical disturbances as the output of the trained neural network
• Visualizing and interpreting the chimney volume.

Using the chimney cube in conjunction with acoustic impedance, AVO, fluid factor and other structural, stratigraphic and geophysical interpretations allows study of chimneys as the spatial link between source rock, reservoir trap, spill-point and shallow-gas anomalies.

Attribute extraction. Based on experience and area geology, an intelligent selection of attributes that have the potential to increase contrast between chimneys and nonchimneys are made. Attributes can be amplitude, energy, frequency, phase, dip, azimuth, similarity and coherence measures, among others.¹ Attributes can be extracted and merged from different input cubes, e.g., near- and far-offset stack, inverted acoustic impedance, etc. Many of these attributes are influenced by the shape and orientation of the extraction window. Fig. 3A shows the attribute-extraction window (actually an extraction volume).

Fig. 3. Window (volumes) to calculate attributes: A) for chimneys; B) bed termination; C) faults; D) general flexible bodies.

The vertically oriented extraction volume reflects the fact that detecting chimneys involves looking for vertically oriented bodies of considerable dimensions. Knowledge of chimney characteristics is used to calculate attributes such as energy and various types of trace-to-trace similarity in each extraction volume. Note that for detection of other bodies, such as faults and reflectors, other window orientations are used, Figs. 3B and 3C.

Finally, a window (actually, a volume) such as the one in Fig. 3D is used to highlight a more general geologic body. Such an extraction volume follows the desired object at every position. This implies that the volume has a flexible shape that follows the local dip and azimuth of the data. Local dip and azimuth can be calculated in many different ways. A modified version of the Wigner-Radon transformation scheme developed by Steeghs is used.² Calculated local dip and azimuth are not only used to steer attribute-extraction volumes, but also are a perfect vehicle to remove random noise prior to attribute-extraction processing.

Attributes that are inputs to the neural network (discussed in the next section) are calculated as follows:

• Reference time is the time (in ms) of the sample position at which attributes are extracted.
• Energy gate is the energy calculated within a time gate between A and B (in ms) relative to the reference time. Energy is calculated as the integration over the square of the amplitudes, divided by the number of samples.
• Dip angle variance is the variance of the dip angle for all samples inside a time gate with a block vertex of size 2 x n + 1 traces (n is the trace step-out). Dip is calculated using a sliding Fourier-Radon transform.³ The angle referred to is between the local seismic direction and a horizontal plane. It is calculated within an optional step-out.
• Similarity measure is the similarity between one trace segment and another trace segment, optionally calculated over multiple trace pairs and those same pairs after 90° rotation. The reference trace is called 0,0, while trace 1,1 is the neighboring trace, located in the upper right quadrant. The variables value = Average will output the average of the pair’s values, while value = Minimum will output the minimum of the pair’s values.

Similarity is defined as:

where S is the summation over the samples in the specified time gate. If trace samples are considered a vector in hyperspace, then similarity is one minus the length of the difference vector divided by the sum of the lengths. This value is always between 0 and 1.

• Time gate [A, B] is defined on the reference trace. To find the corresponding time gate [A', B'] on the traces at pos1, pos2 and the 90°-rotated trace pair, follow the local dip and azimuth from the reference position in the direction of the target traces. This process is called steering. Full-steering means following the dip from trace-to-trace until reaching the target trace.

The reference time on the target trace can be fine-tuned by searching for the same signal phase on the target trace, then comparing it to the starting position. It is used where noise levels are very low.

A key step in the procedure is that attributes are extracted in three separate time windows: one above, one at and one below the investigation point. This allows use of the fact that chimneys are vertical bodies with a certain dimension. Most time-windows in this network have 80-ms lengths, although some have 40-ms lengths. It should be emphasized that combining different attributes through the training process helps distinguish chimneys from other nonvertical disturbances. Fig. 4 demonstrates this fact by comparing chimneys – identified by a multitude of attributes (Fig. 4A) – with a single-attribute output of coherency (Fig. 4B) that does not separate chimneys from other disturbances. Also see Fig. 7A on the separation power of multitude of attributes against a single attribute.

Fig. 4. Comparison of chimney output (A) with "similarity" or "coherency" as the only attribute (B). Click for enlarged view

Training neural networks. Neural networks have recently gained popularity in various petroleum-industry applications. Typically, neural networks can be considered as a nonlinear transformation between its input parameters and the desired output. The main feature of these networks is the fact that one does not attempt to derive the explicit nonlinear relationship – unlike what is done in conventional regression analysis.

Instead, a collection of optimum weights is derived to relate the output of the nodes to their input using available data. These weights, combined with the intrinsic nonlinear (sigmoid) functions defining each node, create the implicit nonlinear functions between input and output of the entire network. Fig. 5 shows a schematic neural network where a nonlinear mapping of input variables – say, two selected seismic attributes, x1 and x2 – yields the output y through an implicitly derived, multidimensional nonlinear function.

Fig. 5. Nonlinear mapping of all attributes through fully connected multi-layer perceptron. Click for enlarged view

In these object detection applications, after selected attributes have been extracted at a representative set of data points, they are recombined into a new set of attributes to facilitate the detection process. In this step, both supervised and unsupervised neural networks are used.

The main difference between the supervised and unsupervised training approaches lies in the amount of a-priori information supplied. Supervised learning requires a representative set of examples to train the neural network. For instance, networks can be trained to find the (possibly nonlinear) relationship between seismic response and the rock property of interest.^4,5 In this case, the training set is constructed from either real or simulated well data.

In unsupervised (or competitive) learning approaches, the aim is to find structure within the data and thus extract relevant properties or features. An example of this approach is the popular waveform-segmentation method, whereby waveforms along an interpreted horizon are segmented. The resulting data segments (patterns) are then interpreted in terms of changes in facies or fluids.

The same principles are used in the object-detection method. If unsupervised learning approaches are employed, attributes related to objects that one would like to detect are used. The supervised learning approach used here goes a step further. Not only are meaningful attributes used, but in addition, locations are identified in the seismic cube where examples of the object class to be detected are present. Seismic attributes are calculated at these positions as well as at control points outside the objects. The neural network is then trained to classify the input location as falling inside or outside the object. Application of the trained network yields a texture-enhanced volume in which the desired objects can be detected more easily.

The structure of a Multi-Layer-Perceptron (MLP) neural network is shown in Fig. 6B. The MLP’s input derives from various attributes calculated from seismic data at different time gates; output is a measure of the combined chimney-like behavior of these attributes. Fig. 7A shows the discrimination power of the attributes. At the training stage, appropriate weights for the input parameters and hidden layers (layers of neural network involving the nodes between input and output) are calculated.

Fig. 6. Structure of a neural network with different attributes calculated from seismic data. Click for enlarged view

As these weights stabilize (as manifested by the convergence in Fig. 7B), the training process is stopped. The optimal time to stop training is determined by monitoring the network’s performance on an independent test-data set, (the blue curve in Fig. 7B). Finally, with an acceptable percentage of correct classification, the network performs the volume transformation for the entire data set, (Fig. 7C). The following summarizes the procedural steps:

Fig. 7. Neural network performance: A) separation power for chimneys and nonchimneys, a single attribute (left) and multi-attribute neural network output node (right); B) convergence during the training; C) correct classification matrix. Click for enlarged view

1. A seed interpretation is made with locations inside manually interpreted chimneys and in a control set outside the chimneys, Fig. 6A (the picks in yellow and red).
2. Various energy and similarity attributes are extracted in three, vertically aligned extraction volumes around the seed locations.
3. Step 1 and 2 are repeated to create an independent test set.
4. A fully connected, MLP-type neural network is trained to classify the attributes as representing chimney or nonchimney (output vectors 1,0 or 0,1). Fig. 6B shows a typical network topology.
5. The trained network is applied to the entire data set yielding outputs at each sample location. Since the outputs are complementary, only the output on the chimney node is passed to produce the final result: a cube with values between roughly 0 (no chimney) and 1 (chimney), Fig. 2.

Identifying and interpreting chimneys. On seismic data, chimneys appear as vertical bodies of varying dimensions. Although shape and distribution may vary, cigar-shapes and distributions along faulted zones are common. On seismic images, internal texture shows chaotic reflection patterns of low energy. The exact outline of a chimney is very difficult to determine on conventional seismic displays, and only large chimneys can be recognized. To detect more-subtle disturbances, the data is transformed into a new cube that highlights vertical disturbances. High values in this cube indicate a high probability of chimneys.

Once chimneys are identified, they can be displayed in conjunction with other structural and reservoir-property information. This helps validate certain geological interpretations, such as the origination locales of hydrocarbons, spill points, reservoir accumulation and gas seepage to the surface.

An example of such an interpretation by Heggland⁶ is given in Fig. 8. The deeper cloud of high amplitudes (red) corresponds to the outline of a salt dome, while the shallow cloud of high amplitudes is interpreted to represent a hydrocarbon-charged reservoir. Chimneys (yellow) surrounding the salt dome indicate upward fluid migration from a deeper reservoir. The high density of shallower chimneys indicates charging of the shallow reservoir. The sub-seafloor surface exhibits a radial fault pattern caused by upward movement of the salt.

Fig. 8. Visualization of chimneys with the structure. (Courtesy of Roar Heggland, Statoil.)

Chimneys are visible up to the seabed, and a small mound is present at the seabed, close to the top of the shallowest chimney on the right-hand side. This may be a small mud volcano generated by transport of sediments, fluid and/or gas to the seafloor. The presence and distribution of chimneys that have been mapped in this area make the presence of a deep and a shallow hydrocarbon-charged reservoir more likely.

Acknowledgment

The authors appreciate the contributions of Statoil, specifically those of Roar Heggland and Paul Meldahl to the development of this technology and extensive insight and collaboration on earlier projects. Forest Oil and AGIP are thanked for granting permission to present the data examples. The authors also acknowledge Herald Ligtenberg of dGB for his contributions to gas chimney processing.

Literature Cited

¹ Aminzadeh, F., Pattern Recognition and Image Processing, Geophysical Press, 1990.

² Steegs, T. P. H., "Local Power Spectra and Seismic Interpretation," PhD thesis, Section for Applied Geophysics, faculty of Applied Earth Sciences, Delft University of Technology, 1997. ISBN 90-90108 12-2.

³ Tingdahl, K. M., "Improving Seismic Detectability Using Intrinsic Directionality," MSc thesis, Earth Science Centre, Göteborg University, rep. B194 1999.

⁴ de Groot, P. F. M., "Volume transformation by way of neural network mapping," 61th. EAGE conference, Helsinki, 7 – 11 June 1999.

⁵ de Groot, P. F. M., "Seismic Reservoir Characterisation Using Artificial Neural Networks," Muenster, 19th. Mintrop Seminar, 16 – 18 May 1999.

⁶ Heggland, R., Meldahl, P., de Groot, P., and Aminzadeh, F., "Seismic chimney interpretation examples from the North Sea and the Gulf of Mexico," American Oil and Gas Reporter, 2000.

Coming Next Month

Part 2 – South African and Gulf of Mexico case studies illustrate use of gas chimney technology for hydrocarbon exploration.

The authors
	Fred Aminzadeh, President and CEO of dGB-USA (fazedgbusa.com, worked for Unocal for 17 years. He has authored many books and articles on different aspects of geophysical technology, including modeling, pattern recognition, seismic attributes and seismic data processing. He also holds three patents on AVO modeling, seismic-while-drilling and hybrid reservoir characterization. Mr. Aminzadeh earned his PhD from the University of Southern California, and has served as chairman of SEG’s research committee for two years.
	Paul de Groot is co-founder and director of dGB (de Groot-Bril Earth Sciences B.V.), which specializes in quantitative seismic interpretation, stratigraphic analysis, seismic inversion, neural networks-based reservoir characterization and gas chimney / fracture detection. He began his career in 1981 with Shell, then became director of Quest Geophysical Services and subsequently, served as senior geophysicist for TNO Institute of Applied Geoscience before co-founding dGB in 1995. Mr. De Groot holds MSc and PhD degrees in geophysics from the University of Delft.
	Timothy Berge, chief geophysicist at Forest Oil International, has a Masters degree in geology from the University of Texas at Austin. He has worked for Exxon for 17 years in Denver, Midland, Bogota Colombia and Houston; for Corpoven in Puerto La Cruz, Venezuela, and with Apache in China. His technical expertise includes the areas of structural geology, sequence and seismic stratigraphy, workstation applications and 3-D interpretation. Mr. Berge is the author of numerous publications, and is SEG’s Global Affair’s committee’s regional coordinator for Africa.
	Giuseppe Valenti holds an MS degree in geology from the Universitá of Parma, Italy, and joined Eni, AGIP Division, in 1988. After few years spent in geophysical research and seismic stratigraphy application, he moved to West Africa (Angola) and then to U.S. (Gulf of Mexico) to concentrate on deep water E&P.

Part 2