Using Self-Organizing Maps for Attribute Analysis
SIPES 2013 Annual Meeting
What we're doing, and what I want to talk about today, is self-organizing maps, and a SOM is a type of artificial neural network that trains without supervision on unclassified data to produce a map. Now people are not as familiar with unsupervised maps as they are supervised maps. An example of a supervised neural analysis might be, well, how many of you are familiar with some software called Strata Magic? What we're doing is unsupervised. What neural networks do for you in general in the public they've been around for a long time; cluster analyses, classification regression, forecasting dimensionality, reduction in data compression, and the data compression is kind of what we're into.
Unsupervised neural analysis using seismic attributes
By Deborah Sacrey, Auburn Energy
Full Presentation Text...
This afternoon we're going to talk about a new software I've been working on for the past couple years that has been started by Tom Smith.
We're doing unsupervised neural analysis using seismic attributes and trying to find anomalies in seismic data. Okay, a little bit about Geophysical Insights, it's a new company, like I said, Dr. Smith was a gentleman who started seismic microtechnology. Most of you know or use Kingdom software. He was a guy that started Kingdom and when the Capital Ventures bought him out about five years ago, he was too smart to kind of sit around and just retire.
So he started tinkering with unsupervised neural networks. He and Terry Tenor who is now deceased but he worked with Terry at Chevron. Started working on building an engine. He built his engine in FORTRAN. Now we put a car body on it in C++ but to look at how we could better use seismic data and find oil and gas. What we're doing now is we're putting the UI together. We're testing what we're doing. We're looking for test cases and we're taking up to 20 square miles of data or 2D Data.
If you want and you give us a problem we try to find a solution. Go to our website (geoinsights.com) it's all for free but we're looking for test data. So that's about Geophysical Insights right now. What we're doing, and what I want to talk about today, is self-organizing maps, and a SOM is a type of artificial neural network that trains without supervision on unclassified data to produce a map. Now people are not as familiar with unsupervised maps as they are supervised maps. An example of a supervised neural analysis might be, well, how many of you are familiar with some software called Strata Magic?
Okay, Strata Magic is a supervised neural analysis where you have known well data, so you have a good well and a bad well, and you want it to map the changes in the seismic waveform between the good well and the map well at the bad well. And it will actually give you a map channels that can be mapped, and that’s an example of supervised neural networks.
Unsupervised Neural Networks
What we're doing is unsupervised. What neural networks do for you in general in the public they've been around for a long time; cluster analyses, classification regression, forecasting dimensionality, reduction in data compression, and the data compression is kind of what we're into.
Okay, this first example is from Kahonen's book on unsupervised neural analysis and what he did is he put together 39 attributes, much like we would look at attributes in seismic data, he put together 39 attributes which represent the quality of life issues around each country in the world most countries in the world. He had 39, and he went through 126 different countries to do an analysis. He put the date in the spreadsheet and he looked at quality of live attributes in the country and ended up with a 9x13 hexagonal neuron topology and I’ll show you what the hexagonal neuron topology is later.
This is a list to the country's that he used in his study and by time he did the analysis, the neuron analysis, it was very interesting how the countries with very similar attributes tended to lineup. He put it in color so it's really easy for people to see but if you look at quality of life say, for Canada and US, very similar. Israel Australia and New Zealand and then you get over into the African countries and other parts of the world but they all line up according to the similarity that they had in the quality of life attributes. And this is kind of a concept that we're taking in, to using this in seismic and I'm not going to read all this but the key elements here are self-organizing map, neural map, what's an anomaly here.
The important thing about anomaly is that we're looking for probability of fit and when you have something as anomalous it has a very low probability. You have a neuron out there that's looking for something different in the seismic attributes from every other part of the seismic data.
Self-organizing Maps (SOM)
Classification. When we run a SOM analysis on seismic data we end up with two new volumes. One is called a classification volume, one is called a probability volume and that's where we see our anomalies and that's where we can highlight them and use them for different reasons and that classification and probability cut off are part of the things that we're showing.
Okay, this is the slide I have the movie hopefully it'll work. It's gonna work, I see it thinking! Okay, if you go to the Geophysical Insights website (geoinsights.com) we have a little example of how the SOM neural analysis works. What I'm doing here is I have put this together with three attributes and I'm using 200 data points.
This is a free download that you can get on our website is called SOM-Lite. I'm showing you the different data points for the three different attributes that we're gonna use in this analogy. Or in this SOM map that we're going to make. We can look at it in 2D or we can look at in 3D. Part of this is built into the new software that we're working on and those little triangles out there represent the data points.
I'm setting this up to be an 8x8 neural analysis so I have 64 neurons. That are going to go in to these 200 data points and you can look at it in random or you can set it to 0. I'm gonna set this to 0, because I don't want any kind of bias out there for the neurons.
Then I'm going to tell it to start training and I'm going to run through ten different epics, and the epics are the training process, the neurons go through the data ten different times and they're gonna analyze the different data points. And you'll notice as we go through down here at the bottom I'm showing this, I had this timed really well. I'm going through Epic's, each epic they they train a little bit better so they start slowing down. They start out with jumping all around and then as they get through the training process they slow down. And they start figuring out which of the data points they're going to attach themselves to.
See they're not moving around near as much now and I'm at epic 7, 8, and you'll see that up in the corner there's a color topology map, similar to the country map that was in colors. These things are lining themselves up in colors according to the data points not stopped and you'll see the vectors that they've attached themselves to the data points.
I'll just rotate it a little bit for you so we can see how each of the neurons, 64 neurons, have analyzed and attached themselves to the 200 data points. We can look at it in the 3D or we can look at it in 2D view and that's kind of a top-down view. So what's anomalous in here? We want to look for the neurons which attach themselves to no data points. Those are the three little anomalies in the seismic data.
And that's just kind of a brief little show-and-tell of SOM-Lite, like I said it's free to download on our website, and it's fun to play with you know. We only have it going up to three attributes but you can put a thousand data points in there and you can have any, you can have a 16x16 or any kind number of neurons that you want to do just to watch it go through.
SOM and Interpretation
Okay, so the first kind of cases where I want to go to is, a client of mine came to me, they had shot a 3D, a 10 square mile 3D trying to extend the Eagle Ford into North East or, North-Central Texas.
The critical elements about looking at unconventional reservoirs is you've got to understand the reservoir geology, you got to understand the geochemistry, you got to understand the geomechanics and this is where the geomechanics and a lot of the acoustic impedance inversion. The inversion volumes really become important for looking for sweet spots in seismic data and certainly there's attributes that are involved in faults and fractures and stress regimes.
Coherency, curvature those things are also very important. With our client we ran the normal attributes that I could run and I start out with the project in Kingdom and I'll run through the stuff that's available in Kingdom and then I have Rock-solid so I'll do that and then Tom's new software also runs its own suite of attributes, we're working on new attributes just to bring to that.
But in addition to the normal attribute suite that we ran immediately because they had well control they had done some extra processing on the data and they brought us final density, landrow, muerow, Poissons’ brittleness, Poissons’ ratio, shear impedance and brittleness to the table.
So, we ran a SOM analysis, and this is kind of the outcome where we thought the sweet spot was for the Eagle Ford in this particular area. They had already drilled a couple wells in here, in fact they had drilled this guy right here which had very few shows and they had some mechanical issues with drilling the well and didn't end up producing anything out of it.
So the second well they drilled it's kind of hard to tell because it's, laterally going in here, they start getting some really good shows and again they had some mechanical issues and my client I have concurred that maybe their engineer is not the great engineer he thinks he is but that's another story.
And that's not disparaging to any, or all engineers, there's just a couple maybe are not quite up to snuff. Anyway so but they had excellent shows and excellent shows when they start getting into this sweet spot right here. They are drilling some other wells in this area that they haven't really started drilling in for the Eagle Ford yet but the second part the story which is more conventional is for the Buddha and they're looking at this area for the Buddha.
So if you looked at it with say a sweetness volume, the sweetness volume wouldn't have shown all of the sweet spots for the Eagle Ford, that we found when we ran a series of attributes into a self-organizing Map (SOM)
So here is the SOM analysis in the classification mine and it's kind of hard to see but the well is right here and it comes in here it got its first show at that little red spot and it started getting better and better shows as I got into this red area right here.
This is the top the Eagle Ford right here this is a zone I mapped kind of the oil window and then this is the top of the Buddha so this area would be good for the Eagle Ford and the Buddha in this particular instance. They had shown that this little area in this area right here before they had mechanical problems and stopped.
Well again, if you look at it in the conventional attribute this being sweetness you wouldn't necessarily know where they were getting their shows. Now the Buddha still shows up a sweet but I think that's a frequency based issue with the Buddha being a limestone in here to see that the attribute analysis brings out information in the data that you wouldn't necessarily see if you just looked at each attribute on its own.
This is a summation of the SOM analysis. I did a 12x16 grid. These are a list of the attributes I put in there. Some of which are conventional attributes like, envelope slope or the PSTN volume but a lot of them are the specialized impedance volumes that they gave to us to use for the analysis. So part 2 this particular clients problem was trying to identify something in the Buddha which is more of a conventional reservoir in here even though they're doing laterals in the Buddha there have been some excellent Buddha wells completed nearby.
The first two wells which were Eagle Ford test did not hit the Buddha but obviously there's a little bit of a Buddha sweet spot where the where the Eagle Ford sweet spot was but this is the area that we determine was the Buddha sweet spot. And indeed they were drilling a Buddha well right in here.
And they got to a certain point they were having excellent shows two to four thousand units of gas, they were heavy on their mud and they were still getting those kind of gas shows and the pressures were so high, I can't remember the total story but again, they had some mechanical problems and they had to stop. They realize that they need to set a liner and they were going to set the liner then drill out and keep on going, which they ultimately did. They drilled about another 150 feet past this point right here but ended up completing the well, or the last thing I heard they had completed the well naturally for about three hundred barrels a day.
They haven't fracked to my knowledge but I think that there's more running room for them here in in the Buddha according to our SOM analysis and again the sweet spot in Buddha and I think this is probably an amplitude driven or a frequency driven issue with the sweetness, but obviously they’re in a sweet spot, no pun intended.
But you wouldn't maybe have noticed the rest of the area of the sweet spot in the Buddha if you hadn't combined and done the SOM analysis. Here's the well. The well is coming down here and its turning and coming down here there's another fault block there and there that look like they're also good. But this is where the first well went. And this is where they stop.
The Buddha on a sweetness volume, see the sweetness volume here looks well in there anyway this is the result the SOM analysis, now because this is more of a conventional reservoir. I ended up using more conventional type attributes I used attenuation, I used average energy I did some dip variants looking at some of the fractures stuff. I did use some of their inversion, poissance, brittleness and stuff but I put up a different mix of attributes in here for a more conventional reservoir than I did for the unconventional reservoir.
Okay case number two and, this is an area that Brian Calhoun and I have been working on in Lavaca County and I was fortunate in our little piece of 3D to have gathers, so I created some very specialized AVO type volumes, far minus nears, and far minus nears times fars, in this area.
Data quality was really good we had really good resolution in the gathers out to about 45 degrees and it took getting out past 35 or 40 degrees before you saw the results of this type 2 AVO, so you probably wouldn't have seen it on the stack data not as obvious anyway.
These are these are the volumes I created and this was my original mapping, before the well was there, and I looked at it and said so well you know I like that amplitude but how many of us have drilled amplitudes that turned up to be fizz water or something less than desirable. You're looking for an analysis that can probably help reduce the risk in type 2 and type 3 reservoirs and I told Brian at the time I said, I give that about seventy acres And he says, we'll see.
Okay so this is what it looked like on the PSTN raw volume trough over P you know nice little thing but you know it's not anything to write home to mother about. And there's still a lot of risk with these kinds of reservoirs.
So when I did the SOM analysis I use the AVO volumes. I use the PSTN Raw amplitude. I used average energy, I used sweetness in the pot and the results are kinda interesting because my seventy acres now turns into about 180 to 200 acres because these are depletion drive reservoirs in the Yegua with a finpay probably about 18 feet somewhere between 24 and 28 percent porosity but all these little guys appear to be linked.
And sure enough, I have to give Brian credit on this he told me a long time ago when we first started producing the well he says this is a bigger reservoir because of the pressures and then the seventy acres and I have to say he's right. Hurts to say that but I have to say it.
This is what it looks like in the SOM analysis, it’s,you know, really isolated this is the map again and this is a line through the well now there was another anomaly that we did see in the area if you look up here the time slice that shows this anomaly really well is just coming through the bottom of this one because it's a slightly different elevation.
This is down thrown and this is another channel coming through here but you can see obviously that this one is linked in a channel system. This, and this guy just he hit the fault and kind of spread out. The two wells show up quite well. We did drill this one and it was successful and the nice thing about the software is we can go in and we have a 2D and 1D color bar and we can isolate these things in the 3D viewer by just picking on certain neurons and bring out and what we're learning you know the more data we work with more tests we do the more we're learning there's anomalies in the low probability. When you click on and put a combination of neurons together sometimes you can bring out channel sequences, sometimes you can bring out stratigraphic information that may not be hydrocarbon bearing but may impact what you're looking for in the area.
Principal Component Analysis (PCA) and Interpretation
The next thing we're working on is principal component analysis (PCA) and that's the next thing past SOM. What Principal Component Analysis is is if you have a lot of attributes and you want to find out the ones that give you the most bang for your buck or make the most impact in your interpretation and that's what principal component analysis does.
In this particular example I used 25 attributes, and I did a PCA and I did on a line by line basis in the dataset that Apache gave us. Apache gave me 150 curvature volumes and 150 spectral decomposition volumes and they specifically wanted me to run a PCA on those volumes to find out which curvature volumes and which spectral decomp volumes had the most impact. So what this kind of does is this little bar right here, this first line and the second line and so on and so forth is a combination of these twenty five attributes that make the most impact. The highest, what they call, eigenvalue.
The software they we're working on and it's changed its look even since I made these slides now we have a graph below and we're showing the percent contribution of each of the attributes in each one of these bars you can cursor over each of these bars and these numbers will all change and we can go and different lines and say look at the different line in a whole different set of attributes comprised of different values in these lines and, again this will change.
The idea is to take this green line and raise or lower it to include the top 10 eigenvalues and then run a SOM on those so everything that's very similar is down below the curves. If you've got 150 attributes and you only want to run a SOM to see what the top 10 do, you want to run a principal component analysis (PCA) to get rid of all the riff-raff and only run it on the good stuff.
In conclusion unsupervised neural networks analysis or SOM maps can solve a lot of problems. They can get in and look at attributes to reduce risk and identify specific solutions and I guess another example I have of this a recent problem we had with some Repsol data offshore Guiana.
They hit a zone of high-pressure when they were drilling a hundred million dollar well and the pressure got so great they had to stop the well and they could not see the pressure in the seismic data. So they gave us 10 square miles offshore Guiana. They gave us the well information. They gave us the DSP all the time depth data. They gave us the pressure profiles and they said run a SOM analysis to see if you can see the pressure.
We used a specific set of analysis employing curvature and similarity and some specialized volumes that we run in the software and sure enough when you put in the 3D volume and you isolate, you make transparent you know certain neurons, we knew exactly where they hit pressure. They stopped the well three hundred feet short of going out of pressure and so now they have us doing an analysis in the Gulf of Mexico where they're having some problems and we're going to try to see if the same sweet of attributes also worked for pressure identification.
But there was something that their geophysicists and their engineers could not see in the regular data that we were able to pull out using the SOM analysis and the more information you have, wells, production, the better the solution could be to specific attribute selections. Neural analysis can be done on 2D or 3D data and a principal component analysis, if you have lots of attributes going in, will save you time and possibly give you better results.
And finally, we're going to the SEG for the first time so if you're going to SEG come visit us and thank you guys very much.
[sound not picked up in video]
Well in the case of the Eagle Ford I ran it from 1.2 to 1.6 seconds on the Repsol data we ran it over two time windows trying to get an upper or lower boundary for the pressure. A lot of volumes and I'll start out running the whole darn volume and then see if I need to narrow down the window.
In the case of the Yegua AVO's, we had so much interference from the shallow Frio production that I had to narrow that window down as well to isolate the Yegua better.
No, a volume-based analysis and not necessarily a timed now, as we improve the software we're going to start being able to bring horizons in and then do the analysis and time windows above and below horizons or between horizons and things like that but we're not quite there yet.