Geophysical Insights at EAGE 2017 in Paris, Booth #350
Rocky Roden and Ching Wen Chen in May edition of First Break - Interpretation of DHI Characteristics using Machine Learning
Deborah Sacrey wins the 2016 RMAG Best Speaker Award
Seismic interpretation and machine learning by Rocky Roden and Deborah Sacrey, GeoExPro, December 2016
Deborah Sacrey Speaking at the CAPA Technical Symposium, November 4, 2016

Interpreting Below Seismic Tuning using Multi-Attribute Seismic Analysis

An excerpt:

The SOM, or the self-organizing maps, employ the values from these different attributes and looks at the patterns and clusters that they form. We look at these on a sample-by-sample basis, which is very important, very significant in this overall process. What this allows us to do is to do a thin bed analysis and we're not hampered by the traditional frequency and amplitude of the wavelet resolution limitations.

So one of the principals in self-organizing maps that enable us to try to see thin beds is that it is unsupervised. SOM's are an artificial neural network employing unsupervised learning methods. That's very important that it is unsupervised. There is no previous training on anything ahead of time.

A SOM is nothing but a cluster or a pattern recognition approach. So what it does is it looks at all the information from these attributes and something we call attribute space. It identifies any sorts of natural patterns or clusters that develop. Again, we identify these different patterns or clusters with something we call neurons. The reason we do this is that these patterns or clusters, depending on the seismic attributes that you're employing, have geologic significance. We are seeing tremendous detail with these SOM analyses and this thin bed analysis. If we can look at this with numerous attributes, then we can start to see thin beds that are not limited by the conventional thinking of amplitudes and frequencies. more >

SOM Analysis to Further Visualize Thin Beds

A Webinar by Rocky Roden, Geophysical Insights

December 2016

Full Presentation Text...


 Well, good morning. This is Hal Green, I manage marketing for Geophysical Insights. We are pleased to welcome you all to the webinar on interpreting below seismic tuning using multi-attribute seismic analysis. Before we get started with Rocky's talk today, I'd like to just cover a couple of procedural items that will help us through this webinar. 

Please note in your webinar controls, which should be on the right hand side of your screen, you have a question box that looks like what is on the screen currently. This will be the best way to submit your questions which we do encourage. Unfortunately it is not practical to take the questions verbally, owing to the large number of participants at various locations across the US and Canada. To ensure the best audio experience for everyone involved, we would respectfully request that you use the dialogue box that is on the screen. Again, it is in your webinar controls. Just expand that and type in your questions. Rocky, at the end of the presentation, will read the question aloud and respond. Note the particular slide and we are happy to go back to that slide to respond to questions if needed. 

Now a note about our featured speaker, who many of you already know.  Rocky Roden has extensive knowledge of modern geoscience technical approaches and is the past chairman of The Leading Edge editorial board. As former chief geophysicist, and director of applied technology for Repsol YPF, his leadership role included advising corporate officers, geoscientists, and managers on interpretation, strategy, and technical analysis for exploration and development in the US, 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. Rocky's previous work experience includes exploration and development at Maxus Energy, Pogo Producing, Decca Survey, and Texaco. He holds a B.S. in Oceanographic Technology-Geology from Lamar University and an M.S. in Geological and Geophysical Oceanography from Texas A&M University. We want to welcome Rocky Roden to our webinar. At this point we are going to turn over controls to Rocky and we will begin. 


Good morning. I want to welcome everybody this morning to this presentation. Before I start I'd like to ask a question of you to think about during this presentation. Over the last fifteen years what significant advancements have we seen in seismic acquisition, processing, and interpretation? I would argue that in acquisition we have multi-azimuth acquisition, we have broad band acquisition, and we have pretty sophisticated source receiver arrays. In processing of course our depth migration algorithms have improved greatly over the last 15 years. Broad band processing. But in seismic interpretation, what significant advancements have we seen? 

We certainly have more computer power. Our visualization is dramatically better than it has been. But we still pick peaks, troughs, and zero crossing.  Fundamentally, don’t we do the same thing from an interpretation perspective? I invite you to think about this as we go through this presentation.


Now when we talk about seismic interpretation and we're looking at tuning thickness, the conventional definition of tuning thickness, or vertical resolution as indicated by Sheriff in his dictionary, is a bed that is a quarter wavelength in thickness for which reflections from the top and the bottom, interfere.


This interference pattern is constructive where the contrasts of the two interfaces are of opposite polarity. Of course this produces a very strong amplitude. This is seen in this particular display here. We see a wedge model and this is all fundamental information of geophysics which we've known for years. We see the top and the bottom of a wedge. You take the absolute value of those amplitudes, sum them together and you see the plot of the normalized amplitude. Of course the time separation doesn't change what you get below tuning. 

Well, many years ago, Meckel and Nath, Neidell and Poggiagliolmi, and Schramm et al., will use this amplitude below tuning, this normalized amplitude below tuning, and scale that to come up with some sort of a thickness estimate based on scaling this information.


In fact, this has been used for many years. Now a little bit later, Brown et al in 1984 and 1985, they produces a series of cross plots of thickness vs. composite amplitude, again, determining an amplitude scalar to hopefully get a little bit more accurate calculation of net pay below tuning. 


Connolly in 2007, actually used band-limited impedance through colored inversion for thin bed determination. Again, they would, using this inversion data, generate a scalar to try to make some computations on below tuning or thin bed computations.


So there have been various techniques in the industry, but as Rob Simm indicated a few years ago, these approaches have limitations and require assumptions that may not be met all the time. For example, these assumptions assume that the thickness is the only thing that is happening below tuning and that no reservoir parameters or properties are changing. Which may or may not be accurate. 

So what we need is some sort of an approach that can image thin beds that goes beyond conventional amplitude, frequency, and impedance approaches and scaling approaches which we've used in the past. So what we are going to talk about is a multi attribute approach which uses numerous seismic attributes that exhibit below tuning effects.


So when we combine these in a machine learning approach, we can get a better idea, a more accurate depiction of thin beds. Not only thin beds but the stratigraphy that is related to those thin beds. 

We are going to be talking about instantaneous attributes. But years ago, Taner, Koehler, and Sheriff, in 1979, kind of wrote the definitive paper on attributes in specifically the complex seismic trace attributes. Where they were computing; the Hilbert Transform, which is basically a 90 degrees phase rotation quite often called the imaginary part or the quadrature, the instantaneous amplitude, envelope, or trace envelope, instantaneous phase and instantaneous frequency, and most of our work stations, regardless of what software you employ, usually can generate these attributes.


They are very fundamental, very basic attributes. They were actually derived from the electronics many years ago.

Now, Taner and Sheriff said in 1977, one way to look at these attributes, or a way to think about these attributes is that the amplitudes that we measure, we generate a source that transmits and reflects, we measure that energy at the surface, that measurement they indicate is really a measure of the kinetic energy. That is what we are measuring. As this wave front passes through the earth, the particle motion for p-waves rocks where the direction is in a propagation that it's going is that the elastic restoring forces, as this is being resisted, that energy can be thought of as potential energy. And that is what they indicate represents the elastic potential energy of this whole system. 

So one way to think about this is that the seismic amplitude trace, and then these complex trace attributes, as well as dozens of other attributes that have been derived through the years from these, represents instantaneous attributes that define the kinetic and potential energy of the seismic trace through space and time. Now some of these attributes are very useful and some of them may not be.


But, a different way of looking at it is looking at what they bring to the table in an interpretative perspective and combine them in this machine learning approach that we're going to go over today. 


So what's been done in the past on instantaneous attributes as far as identifying phenomena that occur at or below tuning? Well, Robertson and Nogami back in 1984 looked at a wedge model and asked what do you have below tuning that created an increase in frequency, and in fact they called it frequency tuning.


That bottom display there shows you that when you go below tuning that there is an increase in frequency. This same phenomenon was seen by Tirado in 2004. In this particular case it's getting thinner as you go to the left of the tuning thickness.


It is the same sort of phenomenon. You see an increase in frequency when you get below tuning. 

So at least with instantaneous frequency, there are some things going on when you get below tuning. Now, Zeng, Radovich, Oliveros, and Hardage took advantage of what they call frequency spikes. What you see on the bottom display on the right is the instantaneous frequency. When you see these white arrows, those were identified as frequency spikes. What this is, is when the instantaneous frequency has a hard time computing because of very thin beds and/or some sort of constructive or interference patterns. On my display on the bottom right, the white arrows represent frequency spikes which are indicative of thin beds by the white arrows and the black arrows show some of the spikes associated with some interference patterns above tuning. 


Now, Taner took advantage of this and he developed what he called a thin bed indicator, which is just the difference between the instantaneous and the time averaged frequencies. Taking advantage of this, trying to see these frequency spikes, and relating those to thin beds.


Zeng identified when you did a 90 degrees rotated wavelet, which is quite similar to a Hilbert Transform, and you apply that to a wedge model, when you get to the tuning thickness or below, it's easier to understand the thickness of those beds below tuning as opposed to using a zero phase wavelet which you see in the upper display there. 


So people have been using different attributes. They have been trying to see these phenomenon that are occurring.


Another phenomenon that happens at or below tuning, and this is something we will have known for years when we're trying to understand the phase of our data, is if you take for example this wedge model and that bottom display, taking the amplitude envelope of that wedge model. So it is a very simple wedge model with amplitude envelope of the top and bottom of this wedge. Now if you take the peak of that envelope and compute what the instantaneous phase at the peak of those envelopes is, what you'll see is for thick beds you'll get plus or minus 180 degrees at the top. At the bottom it will be zero degrees. But when you get to tuning thickness, the phase will flip to minus 90. If that top reflection was a positive event it would be plus 90. So these are instantaneous phase effects that are occurring as we go from above tuning to below tuning. 


Davogustto et al in 2013 identified what he calls phase anomalies and spectral phase and magnitude. And if you'll notice on the right, the three sets of displays here at 15 hertz, 30 hertz, and 70 hertz, on this wedge model you get what he calls phase anomalies. Depending on where that particular phase of that data is.


And then he takes that one step farther and they identify what they call phase discontinuities and he calls phase residues. And you can see these effects, they are above tuning but they are also right at tuning.


So again, these are other sorts of phenomenon that are occurring that we see using different sorts of seismic attributes. 


So given that some of the research that's been done in the past, what other sorts of instantaneous attributes exhibit phenomena below tuning? Well, in an effort to try to determine that, I generated my own wedge model. That is a very simple low impedance wedge model and I generated 25 different instantaneous attributes on this wedge model to see what sort of phenomena occur at or below tuning.


I'm going to show a few of these results. Now the first display here is just the convolution of a 5-10-60-90 Hz wavelet with this wedge model. In the upper display is a wiggle trace variable area then the wiggle trace then a color raster display of this. 

Tuning thickness for this wedge model is 10 milliseconds for 30 feet.


And in fact you can verify that by the next display which is where I measured the top and the bottom and as you can see, the normalized amplitude or the absolute value of the top and the base of this wedge model, that's where you get the highest amplitude, the maximum constructive interference pattern. So that's the tuning thickness.


Over the next several displays, the dashed line represents the tuning thickness which is 30 feet for this model. 

So if you look at the upper display here, which is the envelope, as you can see in this particular example, at the tuning thickness you'll get an increase or a phenomenon happens just before and after you get to the tuning thickness with envelope. Running SOM is the next one which is literally a running Sum and is a pseudo inversion. Again, you can see things happening right before and right after the tuning thickness. The next display is relative acoustic impedance which is the high frequency trend. Again, you see very similar sorts of responses as you did with the running SOM. 

SLIDE: 24         

The next display is instantaneous phase, the band width, and instantaneous Q. You can start to see some of these phase effects that we were describing earlier, that are happening before you get to the tuning thickness, and you can see what happens as you get closer to the tuning thickness and some of the changes in the responses depending on whatever the scale these particular attributes are on. They are all different scales. 


The top one here is the instantaneous frequency and we can see in that display, as we saw by the previous researchers, that you do get an increase of frequency when you get below tuning. The next display is the average energy. You can see what is happening at or below the tuning thickness. The second one is the envelope second derivative, you can see some of the effects that are happening there. So there are phenomena happening near, at, or below tuning thickness.


Now the next 3 are quite dramatic and their responses are below tuning. One is Taner's thin bed indicator we discussed previously. The next one is the dominant frequency and the next one is the instantaneous frequency envelope weighted. So there are some very obvious phenomenon or effects that are happening below tuning on these 3 attributes. 


So, through these several sorts of attributes, there are things happening. There are features happening at or below tuning and in the process of looking at these different sorts of attributes, which exhibit these features below tuning, we've looked at applying a self organized map. We call it SOM, which is a non-linear, unsupervised, neural network approach to see if we can understand and unravel how these bed thicknesses relate to these phenomena that are happening below tuning. And we do this in this multi-attribute analysis. 

And the way that we do this is the SOM, or the self organizing maps, employ the values from these different attributes and looks at the patterns and clusters that they form. We look at these on a sample by sample basis, which is very important, very significant in this overall process.


What this allows us to do is to do a thin bed analysis and we're not hampered by the traditional frequency and amplitude of the wavelet resolution limitations. And you'll see this as we get into this.


So one of the principals in self organizing maps that enables us to try to see thin beds is that it is unsupervised. SOM's are an artificial neural network employing unsupervised learning methods. That's very important that it is unsupervised. There is no previous training on anything ahead of time. 

A SOM is nothing but a cluster or a pattern recognition approach. So what it does is it looks at all the information from these attributes and something we call attribute space, and I'll explain that to you in just a minute. It identifies any sorts of natural patterns or clusters that develop. Again, we identify these different patterns or clusters with something we call neurons. These neurons are nothing more than a point that identifies a pattern or a cluster. I think this will become clearer here in just a minute.


The reason we do this is that these patterns or clusters, depending on the seismic attributes that you're employing, have geologic significance. There is actually a much broader application of this, it's not just thin beds. There are many geological features from stratigraphy to facies to DHI's. It really depends on the attributes that you're using in the particular SOM analysis. In this case, of course, we're looking for thin beds. 


A fundamental understanding is that we're looking at every single sample in whatever set of seismic attributes we are employing. In this case I've got 8 seismic attributes. The first one is the amplitude, in the area those were derived or computed and calculated basically from that amplitude. But, all of those other attributes are at very different scales and they are measuring something very different in the over all energy in this particular data that we are looking at from the acoustic response. Every single sample from every one of these attributes is employed. Now the thing to remember is, when we take all this information and put it in attribute space, we normalize or standardize this information. So they are all on the same scale. So the amplitude does not dominate and has no more impact than average energy. Envelope has no more impact than instantaneous frequency. They are all normalized or standardized to the same scale. What we do is we look for any sorts of natural patterns that evolve out of combining all this information. 


So the best way, probably, to understand this graphically is in this example here. What you're looking at is a very simple reflection coefficient model. If you look to the left, I have 6 reflection coefficients. One is the top of a high impedance. You can see a positive reflection. The next one is the top of a very low impedance section. We have a trough and a peak but the thing to remember in this particular model is that this is 2 milliseconds thick. This is a sample interval thick here. It's a very thin bed. The next one down is a thin bed with a positive acoustic impedance where you have a positive and a negative response. Again, that is only 2 milliseconds thick. Then at the bottom is a positive reflection coefficient. So if I take a 30 hertz wavelet and apply it to the reflection coefficient model with some noise, you get what we call is the real. 

This is the convolution of that wavelet with those reflection coefficients with a little bit of noise added to it. Right next to it is the Hilbert Transform, which is nothing more than a 90 degrees phase rotation. So what we have taken is the real seismic trace at 2 milliseconds sample rate, the Hilbert Transform at 2 millisecond sample rate, take every one of those samples and duplicate each one of those a hundred times so I have a lot more data points and then cross plot them. This is only two attributes and here you can see the real vs. the Hilbert Transform. This is what we call attribute space. Of course for two attributes we can make a 2 dimensional cross block. 

You may be able to see, if you had 3 attributes, a 3 dimensional cross plot. But, you go beyond that and we have a hard time seeing beyond 3 dimensions. You start getting into 5 and 10 dimensions, and I just can't see that. Mathematically that's not a difficult issue. So if we look at this cross plot of the real and the Hilbert Transform of these two traces duplicated a hundred times each, that's the pattern that it forms. If you take those data points in attribute space and run a self organizing map on those data points, what you see is what's on the right display. The data points never move in attribute space. The only things that move around the neurons is they move through the data trying to identify the different patterns or clusters in the data. 

What you're looking at on the right is the same data now identified by a self organizing map analysis. In this particular case there were 64 neurons that went through this data looking for the patterns. Each pattern is identified by a neuron that is associated by a very specific color, as you can see. And those colors are associated with each of these neurons in this 2D color map. Now, how does that relate back to our model?


Well the next display is that same cross plot with the patterns recognized and identified by the SOM analysis. And now what you're looking at, if you look in the upper part of this particular display, you'll see a cluster of points circled with the number 8. That 8 represents the number 8 neuron in the 2D color map to the left. And that represents the top of that low impedance thin bed on the model on the right. 

If you look down in the bottom right portion of this 2D cross plot, you see this cluster of points circled, labeled 55. That is associated with the number 55 neuron in this 2D color map. And if you look at the model, that's the base of that thin bed. You're seeing the top and the base of a 2 millisecond bed in this particular data set. And in fact, if you look at it again, on number 28 on this 2D cross plot, that's the 28th neuron in the 2D color map that relates to the top of the positive impedance thin bed. The 33 in the 2D color map is the base and you can see the other reflection coefficients, 57 and 59, they are identified the same way. 

So, if this is with just 2 attributes, this is the methodology that we go through to identify geological features. In this particular case we are looking at thin beds.


Now, with real data and all of the complications and the ways, and issues we have to deal with, probably 2 attributes are going to be very difficult to see these things. But, what we do is go through and run a series of these attributes that produce and exhibit information at and below tuning. So it is very similar to what I just showed you in a very simple model. Say for example we have 10 attributes now. I've got 10 attributes that I want to apply to see if they help me identify thin beds. So if I have a 3D survey, which you're looking at on the left, and I've generated 10 attributes, every single sample point has 10 values because I have 10 attributes. I've generated 10 attributes and if I want to run it over this whole volume, each sample has 10 values, because I have 10 attributes that I've generated from it. So I take all those sample points and I put those in what we call attribute space. 

So these 10 attributes with all these samples, are in 10 dimensional space now. So I take all that information and that's when we normalize or standardize the values of all these data points. So all these points now are all in the same scale sitting in attribute space and that is when we run self organizing maps. In this particular case, we're looking for 64 patterns or clusters in the data. Now when you go into this, you may not know how many patterns or clusters there are. We know from experience though, that typically we will run from 3x3, which is about 9 neurons to as many as 12x12 which is 144 neurons. We typically run a whole series of different numbers of neuron in these different matrices because we don't know exactly. It depends on what we're looking at. It's the scale of the feature we're looking at. And of course in this case we are looking for thin beds. So we are probably going to use a lot of neurons to try to identify these thin beds. 

So after we've run a self organizing map and it's gone through all these data points in attribute space, we non-linearly map this back to that 2D color map like I showed you in the previous example. So each one of these neurons represents a pattern in the data. It represents a cluster in the data. Some of these neurons and some of these patterns may be noise. Some of these patterns may be showing you geologic features of interest. In this particular case, we're selecting a group of attributes we think can help us see thin beds. 


So now I'm going to show you 3 examples that my colleagues have been working on and where they've used multiple attributes using Paradise from Geophysical Insights, to identify these thin bed responses.


This first one is in the Eagle Ford shale. My colleague, Patricia Santogrossi, has done a tremendous amount of work in detail with a tremendous amount of information. It's not normal in a little over a 200 square mile 3D survey that's sampled at 2 milliseconds to have over 300 wells, various pilot holes, horizontal borehole images, and even more importantly, there are 4 cores with x-ray diffraction and saturation information. 


So we have a tremendous amount of information to corroborate our results in this particular SOM analysis. Now, in this particular portion of the Eagle Ford shale, the frequency in the Eagle Ford shale, depending on where you are, is about 20 to 27 hertz and the velocities vary a little bit. But the range in the conventional, vertical resolution in this area is about 100 to 150 feet thick, depending on where you are. The sample interval, which is 2 milliseconds is about 10 to 12 feet thick. So that's what we're looking at in this particular data set. 


Now what Patricia did is she took a whole series of instantaneous attributes as you see listed on the left there, and she ran a principal component analysis to try to get an understanding of what are the most prominent, the attributes that show the biggest variation of the data. Potentially the most prominent attributes that potentially will show you some thin bed responses.


And what you see here on the right is all the 10 attributes that she ultimately picked. That we ran in the self organizing maps. Now this first display shows you a 100 millisecond window and you can see the top of the Eagle Ford Marl, which is the top of the whole Eagle Ford shale interval. Then the top of the Buda which is basically the base of the Eagle Ford interval. You can see there is a whole lot of variation in these intervals as well as above that in the Austin Chalk.


There are a lot of facies changes. There are a lot of variations going on that potentially mean something, especially from a facies and a stratigraphic variation perspective. 

So then what you're looking at now is what the conventional seismic data looks like compared to the results from the SOM analysis. In this particular case she used 64 neurons. Now, the Eagle Ford shale as you can see on the left, in the conventional seismic data, is tied up in this big trough and big peaks. There's not a lot of variation to it. There's not a lot of subtlety to it because all of that is a lot of information and detail going on in there that we just can't see with conventional amplitude data. But look on the right at all the different variations that the neurons are identifying in this interval using those 10 attributes in the SOM analysis. 


And as you'll see in the next display, we have these very specific intervals tied into the SOM analysis results. In fact, we're looking at one of the wells here. You can see the top of the low resistivity Eagle Ford shale unit. There's a very thin transitional unit between the low resistivity shale and the higher resistivity Eagle Ford shale below. And then there' s a basal clay rich section below that. And above all this is the Eagle Ford marl and then you get into the Austin Chalk. But look at the details and I'm going to blow this up in just a minute, where these specific events are being identified. And you can't see this in conventional amplitude data. Remember, we are not training this on wells and hoping that whatever I'm training on works in this area. These are the natural patterns, the natural clusters that are in the data based on these attributes in the SOM analysis.


So the next display you're looking at is the same well log, actually, the same SOM analysis vertical seismic display but it's blown up. And I want you to look, there's 2 very important things to see on here. One is this very thin transitional unit between the upper and lower Eagle Ford, which is identified by neuron number 55. That is one sample interval thick. Remember that the Eagle Ford shale and every sample in here varies from about 10 to 12 feet thick. That's what we're seeing right there. We're seeing that number 55, at least at this well location, that little thin unit. And what's even more important, I think, personally is if you look at this dashed white line at the top, what you're looking at is the unconformity between the Eagle Ford Marl and the Austin Chalk. I have worked on numerous projects in the Eagle Ford shale and I have never seen that. I have never seen that unconformity. You can't see that it's embedded somewhere in a zero crossing or a trough in a conventional amplitude data. The details in here are tremendous. 


So, this entire area, Patricia has identified 13 different neurons that represent the various facies in the very specific Eagle Ford shale unit which is about 14 milliseconds. 23 different neurons in the whole Eagle Ford group including the basal clay rich shale and the Eagle Ford Marl at the top, which is a roughly 28 millisecond window. And as I showed you previously, we see numerous individual units that are as thin as a sample interval thick. 10-12 feet.


Remember, conventional resolution here is about 100 to 150 feet thick. But, think of the implications of what I just showed you. If I am truly identifying very specific facies below tuning, very thin beds, now we can actually map these thin beds. We can actually see the facies distribution. We are not scaling amplitudes now, we are not scaling inversion, and we are actually seeing the real distribution of facies. Now what you're looking at here on the left is an offset about 12 milliseconds below the Eagle Ford shale and then on the right, you're 16 milliseconds below the top of the Eagle Ford shale. And you can see how this fairway is changing and how the facies are changing and narrowing from the left display to the right display. 


So we are not looking at scale data anymore. We are looking at the real facies distribution of these thin beds that we are seeing below tuning. The next example is from my colleague, Deborah Sacrey, who worked in the Pettus Sand in Onshore Texas. She has a 3D pre stack depth migration of volume, however, this is processed to 1 millisecond sample rate. She is looking for some strand line deposits on this little thin Pettus Sand. Conventional tuning thickness here is 45 feet and the sample interval which is 1 millisecond, is about 5 feet thick.


Deborah, when she did this analysis, and trying to determine what might be the important attributes to use to see these thin beds for the SOM analysis, she ran principal component analysis also. She took the top attributes in the first 3 principal components to come up with this list of 7 attributes to run in the analysis. 


So what you're looking at here is just the amplitude section in color with a variable area density display overlay on top of it. And what you're looking at in the down dip is a very thin little gas sand that was discovered and produced and a very thin oil sand up dip. Now if you look at that interval on this conventional amplitude data, there's not a whole lot going on in that interval of interest.


Now what Deborah did is she took those 7 attributes, ran a SOM analysis, and used 100 neurons. So she was looking for 100 patterns or natural clusters in the data. And now you can see the extreme detail that we are seeing in this particular geologic setting. And if you look on the right, that well that produced gas, that's where that interval was. This very thin brown neuron has identified this interval with that very thin pay. But the real play here actually was the up dip oil that was in the strand line play and that's exactly what you're seeing here in this up dip event. It's sort of a gold color. It's a very thin oil. You can see the whole strand line distribution there, at least in this vertical section. But there is another separate one just down dip of that. 


And remember, when you looked at this on the regular amplitude data, you really don't see anything there. And these are very thin, very subtle facies variations that we are seeing in this particular example.


The next example that I'm going to show you is some more work by Deborah Sacrey looking at the Middle Frio in Texas. There are some very thin sands. 3D pre stack depth migrated data, processed at 2 milliseconds in this particular area. She is looking for bar development.


That is what produces out here in porosity on what's called Alibel Sand. And conventional tuning here is about 200 feet. 2 milliseconds is about 10 foot thickness.


Again, she ran through principal component analysis with this data. She came up with 6 attributes that she determined might be good to show us in the SOM analysis, some thin bed responses. 

Now what you're looking at here is an arbitrary line that is kind of in a strike direction. And this is the regular conventional amplitude data. Again, the color raster is the amplitude and then you have a variable area display overlay on top of that. And the Alibel sand that we are looking for is located between this green and black horizon. This very thin interval. Because I have a well here that's perforated that Alibel sand.


Again, not a lot of character in that particular unit. Here she used 64 neurons and now you're looking at the same green and black horizon with Alibel sand in the middle there. You can see the detail in the facies that are coming out now. And this perforation down here is kind of associated with a light blue and I'm going to show you this in more detail in a minute.


A light blue neuron that came out. This is a dip line to that same well. You can see this circle here. This is what it looks like on the conventional amplitude data, this is what it looks like in the SOM analysis with those 64 neurons. And you can see that blue neuron seems to be associated with this perforation interval.


I'm going to show you an expanded version of that. So what you're looking at, right over here, is this very thin, it's a 6 foot perforated interval, in a 10 foot sand. Which fundamentally is one sample thick. You can see how that's distributed across this particular feature.


So now if you look at that and start to map that area to see the distribution of this, that particular well is located right here. Add those two lines that ran through it. That's 6 foot perforated interval that 10 foot sand is located right here. Now you can see the aerial distribution of that bar. These are the bar locations here and there were a few wells down in the south west portion here that were determined to have mechanical failures. Well now in reality now that you see the distribution of this bar, they were either outside the bar or on the very fringes of the bar. And you're seeing so much detail here that if you notice there is probably a little title erosional inlet sitting right there that could potentially be separating this portion of the bar to the southern portion. 


So we are seeing tremendous detail with these SOM analysis and this thin bed analysis. So, regardless of the phenomena below tuning on these different attributes, it's a very complicated constructive and destructive wavelet interference patterns that are always going to be evident and present in our data when you start to look at thin beds. These things occur because of the geology, whatever is happening is producing these effects or acoustic responses. But if we can look at this with numerous attributes, then we can start to see thin beds that are not limited by the conventional thinking of amplitudes and frequencies. And this typical scaling of amplitude and inversion data. We can actually see thin beds and look at these very high resolution results from the SOM analyses. 

Again, these self organizing maps using these numerous attributes, take advantage of looking at these natural patterns in the data that are produced by these specific seismic attributes. And again, we are not restricted to the frequency of the data for resolution. Our main restriction actually is having decent quality data in the sample interval. The sample interval is what is restricting us here, not the frequency of the data, and not the typical wavelet of the data. It prevents us from seeing a lot of detail in the conventional sense.

SLIDE: 58 [Last slide]

So at the very end here I'd like to repeat what Albert Einstein said. "The definition of insanity is doing the same thing year after year and expecting different results." And I think we are. I think we are seeing something that we are pretty excited about and we've done about 10 to a dozen of these things and we keep getting very encouraging results and we are excited about what we are seeing. 

 [Questions and answers begin.]

[Hal] Alright I'm afraid we are about 20 minutes after the hour and a couple of questions are still coming in but let me offer those folks that we will respond to them individually after our call. I do want to again thank everyone for participating today and offer contact details should you want to have a copy of the presentation. We would be delighted to send that to you.  

Please send me an email at That is also in the invitation to our event today and here is my phone number for your reference, which is likewise in the signature of my email. For those of you who may be interested in test-driving Paradise for 60 days, we offer a free evaluation program. It is a complete commercial license to Paradise. We would be delighted to offer that up for your evaluation on your data and there is some training and consulting that goes with that. There's a wealth of information of course on our website at and a link to that page will be sent to everyone that's registered. There will be a blog post on the website containing all the good questions that folks have asked without naming the individuals and a brief response from Rocky Roden. That blog post will also allow you to add your own comments. Likewise, please send me a note if you would like a copy of the presentation and the slide deck today. We will be posting those to the Paradise customer portal and I would be delighted to forward a link to the portal with your log in credentials to that. With that, I want to again thank everyone for participating. Lots of great questions and interest across the U.S. and Canada. We will be holding these again in the future in the New Year. I want to wish all of you wonderful holidays and an outstanding New Year. Thanks again.