[英文] 整合数据作更好的地质模型的一些原则

Joseph

原题：Somestrategies to integrate all your data, and I mean ALL your data, into bettergeological models that you can actually use to learn about your reservoir
作者：Duncan Findlay

Thispost is based on a couple of conference talks I gave last year (so please givemuch of the credit to my co-authors) focusing on building a better geologicalmodels of the McMurray Formation, but there are lessons here which may beapplied to just about every heterogeneous reservoir. Do you know one that isn’theterogeneous? Me neither. Because of this I started Applied RealisticGeoscience to focus on integrating all the data we have to build something thatwill be proactively useful and that can meaningfully “see” below seismicresolution in many cases, rather than just something which is thrown togetherfor simulation purposes late in the game. You can find some of the referencesand other information at www.realisticgeoscience.com.

Thereare a few things we should be aware of up front. The McMurray Formation is bygeneral consensus a fluvial estuarine environment. As such, we see fluvialfeatures such as point bars, and yet we see the effects of sea water in boththe sedimentary features (tidal couplets) and the extent of the mud deposition.It is important to keep in mind what these systems actually look like when weare building a detailed and useful model. This has ramifications for everythingfrom cell size to stratigraphic architecture, so it is important to give it itsfull consideration in order to get it right, right at the beginning. Check outfigure 1.


Figure 1: The El Sira River ~300km from Lima, Peru


This is a video taken from Google Earth Engine whichis an animated globe incorporating almost 30 years of satellite imagery thatGoogle has acquired. This is a fantastic dataset, and very useful for gettingan idea of what the real world looks like, and how it changes over a humantimescale. Figure 1 is focused about 300 km away from Lima in Peru, and reallygives a great view of the amount of change you can see in a relatively shortamount of time in a fluvial system. Take a second and watch it go around a fewtimes, look for how the river erodes and causes truncations. Look also at theshapes you see, and what a point bar actually looks like in plan view, in “thewild”. These are the kind of shapes we should be using to build our models, andif we aren’t, then were not doing a good enough job. We know they are there, wecan see them on seismic in many cases (see figure 2), why don’t we incorporatethem?





Figure 2: from LABRECQUE et al.2011: point barsrevealed in seismic of the McMurray


So now we know what point bars look like in plan view,how about in cross section? In a great analogue from Dinosaur Provincial Park(figure 3), we see the kind of structures we should be seeing in the subsurfaceof the McMurray Formation. We have inclined fluvial strata deposited vialateral accretion as a point bar develops, and an abandonment facies where thechannel gets filled in by mud when the point bar gets cut off by the riverfinding a new course.





Figure 3: from Findlay et al. 2014, annotated fluvialstratigraphy from Dinosaur Provincial Park


This is great as far as it goes, but the scale of thesediments preserved at Dinosaur Provincial Park is much smaller than thosepreserved in the McMurray Formation, so it’s time to go to the horse’s mouth,and actually have a look at it at the mine face (Figure 4). Note the inclinednature and variable mud lengths within the bedding. The geometries,sedimentology and stratigraphy we see are important, and are features thatshould be incorporated into our model.





Figure 4: from Findlay et al. 2014, annotated lidarimages of the McMurray Formation taken at the Mildred Lake mine face

Seehow far we have come already? We know how the model should look in both planview and cross section, and we haven’t even looked at any specific core,petrophysical or seismic data from the McMurray yet. The situation isimproving, but you do still see McMurray models which are essentiallystatistical fuzz, and do not even attempt to incorporate the features that canbe observed in the figures we have considered thus far.

Let’stake a look at the data we can glean from wells, and see how we can incorporatethis meaningfully into our model. You can get a lot out of well data, and italways drives me crazy when an operator on a misguided mission to cut costsdecides to cut back on the core and wireline log data collected. These days,the actual cost of the coring and logging is so low compared to the overalldrilling costs that is makes very little sense to cheap out on the data. Afterall, the mission isn’t simply to drill a successful hole (despite what some rigmanagers seem to think), it is to collect as much data as possible from thesubsurface. For me this includes core, FMI (Formation Micro Imager), sonic, andat least a triple combo.

Ifyou don’t collect core, you don’t know your oil saturation (yea yea, Archieetc. but I have definitely seen some strange resistivity responses that wouldmake you think you had drilled a duster, when upon examining the core, youactually have 20m of beautiful oil charged sand), you have very limitedinformation on your particle size distributions, you have no ability toevaluate permeability, you don’t have any information on any small mud bedsthat are too thin to be resolved on your well logs yet effect yourpermeability……. I could go on.

FMIgives you a high resolution, wrap around image of the wellbore, and thereforethe rock that has been drilled through. This is useful not just because you canlook at the image and identify sedimentary features and surfaces, but alsobecause analysis of this image gives you precise information regarding theorientation of the features observed. This has proved useful for understandingthe stratigraphic architecture of your reservoir (by identifying theorientation of bedding etc.), and gives you information on your stress regime(by identifying the orientation of fractures). I have found it invaluable toidentify sand on sand contacts by seeing a dramatic change in dip orientationin the middle of a sand body, and identifying an incised channel in anotherwise sandy reservoir in the same way. Important observations that wouldotherwise be missed. In a pinch, a regular, non-FMI dipmeter can useful, but asthis uses fewer sensors, it doesn’t give you the image of the FMI, and doesn’tallow the direct discrimination between dunes, bedding and fractures, amongstother things. More on this later.

Ifyou don’t collect sonic data, you can’t tie your wells to seismic. Which meansyou don’t know for sure where you are in depth on your seismic, which is prettyimportant. Imagine if you are making drilling decisions based on an interpretedseismic survey, only to find that you were looking in a shallower or deeperzone than your target.

Withouta triple combo (which is a gamma ray tool, porosity tool and resistivity tool)you seriously limit your ability to build a stratigraphic framework and correlatebetween wells, in addition to losing out on the other benefits of understandingporosity, resistivity and gamma ray signals.

ButI digress. Knowing what we do about the general style of the sedimentaryenvironment, we can inform this conceptual model with data from the wells. Wecan build a stratigraphic framework based on the correlations we make betweenwells (based on all the data of course) and gain an understanding of thevertical trends in porosity, grain size, oil saturation….. i.e. everything thatis important to your method of production. You might be surprised by howconsistent individual stratigraphic units can be, if you can identifyindividual bars etc. This kind of information will prove invaluable when itcomes time to populate your model with the statistics describing your data. Fornow, the key item is to identify the main surfaces so we can begin to makemeaningful seismic interpretations, and provide accurate stratigraphic surfacesto build the structure of the model.

Butwell data, however powerful, is only giving you a 2 dimensional view of thesubsurface. It becomes more powerful when combined with seismic data, and wecan meaningfully join our stratigraphic framework picks, made using well data,in three dimensions. It’s like a 3d “join the dots” puzzle, and can be greatfun. This might also be where you go back and re-evaluate your well data. If,based on the core, you thought you were dealing with an incised valley system,but the seismic looks like a delta, you need to think again about what thesedata sets are telling you. For this reason, you might end up iterating betweenthe well data, core and seismic to hone your interpretations.

I’llbe honest, I am not a geophysicist, but I will trust the insights of theskilled professionals I have worked with that you can correlate various aspectsof the seismic data with porosity, density, and perhaps even oil saturation. Ihave however interpreted a lot of 3D data (known in the biz as cubes, eventhough the surveys are never cubic), and the various other seismically derivedvolumes just mentioned. I just need them depth converted etc. first. Being ageologist is advantageous when it comes to interpreting seismic, simply becausegeologists tend to have a better mental image of what the reservoir looks like.I just want to emphasize that integrating seismic should be a team effort, withevery team member bringing what they do best to the forefront.

Howeverit is also important to know the weaknesses of seismic. I’m speakingspecifically about resolution. The lateral resolution varies, but is commonlyon the order of 10m, and the vertical resolution is perhaps 8m, and may varywith depth. I won’t get into the technical details of tuning thickness, foldetc. but, while every survey is different, the resolution is often insufficientto identify individual beds in a complicated reservoir. It is anunderappreciated aspect of a good model that you can combine all your data toessentially “see” below seismic resolution. You can see from figure 2 that goodquality seismic imaging can tell you a lot about the reservoir, but thecollection of seismic data is effected by many factors such as surfacecondition (if your survey was collected on land), charge size, geophonedensity, reservoir thickness and reservoir composition (amongst much else). Tosee the kind of detail revealed in figure 2 you often need an expensive seismicshoot and a little gas in the top of the reservoir. The density contrast reallylights the data up. However, this density contrast often disappears when you gobelow the gas zone, so it gets difficult to see the detail of the reservoirbelow this resolution. In addition, the non-optimal seismic imaging of areservoir containing virtually no gas can look like figure 5, which is useful ofcorrelating major stratigraphic surfaces, but it is a major challenge to getmuch else out of it.

Figure5: An example of seismic where detailed stratigraphic information is difficultto extract. From Findlay et al. 2014

However,all is not lost, we still have the surfaces to build the structure of ourmodel, and we can still use what we have to build a realistic model of thesubsurface. We just have to work a little harder. So use the surfaces you haveinterpreted to build your structural framework, and let’s move on to figuringout how to build the layering scheme between these surfaces.

Let’snow consider the FMI/dipmeter for a moment. As previously discussed, it can providevaluable information on the orientation of the bedding in the subsurface. Howcan we use this to build a better model? Consider Figure 6, and pay closeattention to the two rightmost tracks.

Figure6: Sample log complete with dipmeter data. From Findlay et al.2014

Thesecond from the right track is commonly called a tadpole plot, and this is thestarting point for examining dipmeter data. The direction of the point on thedots indicates the dip direction (relative to north) and the lateral positionindicates the severity of the dip. In this way it is possible to identifydunes, IHS (inclined heterolithic stratification) and major changes in dipdirection that may indicate a major stratigraphic surface which might bedifficult to identify otherwise.

Joseph

The well in Figure 6 was selected because it represented a fairly simply, unidirectional case, but even so, the high magnitude dips at the base of the sandy section represent the dune system that was migrating through the thalweg of the channel, and the lower magnitude dips represent the IHS deposited via lateral accretion above this, with a reactivation part way up the section. The right most track on figure 6 is a rose diagram displaying all the data from the dipmeter track over a given interval. It is clear here that the sandy fluvial section of this well has a very strong northeast orientation, with an average dip of ~7 degrees. This is useful information on the orientation of these beds at this location. Map this dominant dip direction spatially as in figure 7, and you can see how we can build a surface that honours this data, and can provide a layering scheme for part of the model.

Figure 7: Map of sand thickness with dip orientations from a sample area. From Findlay et al. 2014

With a little know how, you can build a surface using this kind of data which honours not only the dip direction, but also the average dip magnitude as in Figure 8. This is what you would use to build your layering scheme.

Figure 8: Surface built using the dipmeter discussed in the text. From Findlay et al 2014

But say you don’t have a lot, or perhaps any, dipmeter data? Here is another strategy you can employ to get similar results. In this example, we have one well on which someone ran a dipmeter, but we also have some horizontal well data which we can use to establish a layering surface. Consider figure 9 which contains a log of a different well, but laid out in a similar fashion to figure 6.

Figure 9: Sample log of a different well from a different area. Note dipmeter data in the 2 rightmost tracks. From Findlay et al. 2014

Note that, like figure 6, this well has a very strong orientation in the sand prone section. However, without other dipmeter wells nearby, what can we use to build a reliable layering surface? In this instance, we have some additional MWD (measure while drilling) data from 10 horizontal well pairs which proved useful. Consider figure 10.

Figure 10: MWD Resistivity profiles from 10 horizontal wells with annotated dip direction. From Findlay et al. 2014

Here we see the deep resistivity profiles from the upper wells on the pad along with the dipmeter well near the centre of the image. The colour scheme and thickness of the resistivity data have been manipulated to show the subtle trend of a bed which intersected 6 of the wells in a smooth arc. The magnitude of dip observed in figure 9 lines up beautifully with this trend and allows the construction of a layering surface which honours all of this data, which is shown in figure 11.

Figure 11: Dip surface build by combining dipmeter and horizontal well data. From Findlay et al. 2014

Once we have established a dipping surface, we can use it to guide the layering in the model. The next step is to consider the intuition behind some of the techniques which are commonly used to build the geostatistical part of a model.

I won’t go into using data analysis to figure out what vertical and lateral trends you have in your reservoir because it is going to be different for every model you build. Suffice to say that you need to honour any trends you identify from your interrogation of the data you have available. Know that your grain size becomes finer as you move up vertically? Then you need to build that trend into your geostatistics. What does this change in grain size imply for your permeability? Perhaps we should be honouring that as well.

However, the first model you need to build is a facies model. By this point you should have identified a number of lithofacies, electrofacies or some other kind of facies from your log and core data which you think accurately represents the rock types you observe there. This is often done based on mud content. 0-5% mud being facies 1, 5-15% facies 2 etc, which is fine and easy as you can often just load the core description provided by the core handling company or the geologists hired to log the core. A better way would be to identify lithofacies associations based on sedimentalogical processes, for reasons that will become clear shortly. Nardin et al. 2013 provide an excellent scheme for the McMurray Formaton, and this is very close to the scheme that is used in the example to come.

An important step in building your facies model is considering the size of the variogram you want to use. Variograms are often poorly described and misunderstood, but an easy way to visualise what they are is to think about a rugby ball around one of your data points in space. The larger this rugby ball, the further the algorithm will look to try and join up two points of the same value. In effect, the larger your variogram, the larger your facies will be in three dimensions. But this tells us nothing about how big we should make them. There are variogram analysis tools available to analyse the lateral and vertical variability of your data set, but I suspect they are not going to be a lot of practical use, except for analysing the vertical, particularly in the McMurray (or any other fluvial system), and here is why: Facies in the McMurray are generally too small to be intersected by more than one well. Consider figure 12

Figure 12: McMurray Lidar image with 2 hypothetical wells, from Findlay et al. 2014

This is another annotated McMurray lidar image from Findlay et al 2014 with two hypothetical wells drawn on for emphasis. We know that the McMurray contains dipping beds, and we know that they dip at 5-10 degrees. Now assume you cut that so that apparent dip is the same as actual dip, and that you have a 35m thick reservoir which is dipping at 5 degrees. The geometry of this means that if you want to see the same bed twice, you need to have a well every 400m. Do you have that kind of well density? Not many do. Things get worse though if you have steeper dips, you will need finer spaced wells. Also, even if you have 400m well density, you have other problems. How do you know which bed in well b is the same as the bed you are looking at in well a? You don’t, and even if you did, you would have only 2 data points for that plane. Will it have the same facies in both wells? Unlikely, but even if it did, is 2 data points adequate to assess the variability along that plain within the reservoir? Almost certainly not.

I took you a long way to make the small point that you cannot use well based variogram analysis to define your lateral facies variogram dimensions, unless your facies lateral extent is considerable larger than your well spacing. This is certainly not the case in the McMurray.

So if we can’t use the variogram tools to define our lateral variogram dimensions, what should we do? I suggest we use actual measurements from the formation in question. Consider figure 13, modified from Nardin et al, 2013.

Figure 13: Facies lateral extent statistics measured from the McMurray Formation, modified from Nardin et al. 2013

Nardin et al (2013) painstakingly built a facies scheme, and measured the length of these facies in the McMurray Formation along a mine face, often in dip and strike direction (amongst a lot else, check the paper out. It’s a goodie). In this figure we can see that the mode of lithofacies association c in the strike direction is about 5m and in the dip direction about 8m. This means that the most common bed of this lithofacies association would be 5m wide, and 8m long, orientated down the dipping bed. Let’s think for a minute what that means for our model. What cell size did we use? Did we leave it on the default 50m x 50m square? Better rethink that. If we don’t use a small enough cell size, we won’t see any of this rock type in the model. There are essentially no exposures of this lithofacies association that are 50m x 50m in lateral extent. It also means that we can perhaps extract from this data what our variogram dimensions should be (i.e. 5m x 8m). We can do something similar for the rest of the lithofacies we are going to use, and specify our variogram dimensions accordingly.

Now we have all the pieces in place, we can build our facies model. The kind of methodologies described here built the model exhibited in figure 14. The model isn’t supposed to look exactly like the Lidar image, they are hundreds of kilometres apart in reality, but I think you will agree that the stratigraphic architectural style has been preserved. This fine detail matters when we need to simulate as subtle permeability changes can have a huge effect on steam propagation, and we can be confident that our model is seeing below seismic resolution (compare figure 14 with figure 5, see what I mean?). With this facies model built, we can start to populate other properties in a realistic manner, but that is a post for another day.

Figure 14: McMurray lidar image on top compared to the model built using the outlines methodologies below. From Findlay et al. 2014

Thanks for reading this post, we hope you have found it insightful. You can email me at duncan@realisticgeoscience.com, visit http://www.realisticgeoscience.com , or follow the company on LinkedIn and Twitter. Perhaps Realistic can be of use to you? Look forward to hearing from you.

Duncan

References (the bottom two are available from Research Gate)

Findlay, DJ, Nardin, T, Couch, A. and Wright, A , 2014 Modeling Lateral Accretion in the McMurray Formation at Grizzly Oil Sands Algar Lake SAGD Project, Canadian Heavy Oil Conference, Calgary, November 2014.

Findlay, DJ, Nardin, T, Wright, A and Mojarad, RS, 2014, Modeling Lateral Accretion in McMurray Formation Fluvial-Estuarine Channel Systems: Grizzly Oil Sands’ May River SAGD Project, Athabasca, Geoconvention 2014, Calgary, May 2014.

Labreque, PA, Hubbard, SM, Jenson, JL, and Nielson H, 2011, Sedimentology and stratigraphic architecture of a point bar deposit, Lower Cretaceous McMurray Formation, Alberta, Canada, Bulletin of Canadian Petroleum Geology 59, No. 2, p. 147–171.

Nardin, TR, Feldman, HR, and Carter, BJ, 2013, Stratigraphic Architecture of a Large-Scale Point Bar Complex in the McMurray Formation: Syncrude’s Mildred Lake Mine, Alberta, Canada. in FJ Hein et al (Eds.). Heavy-oil and Oil-sand Petroleum Systems in Alberta and Beyond. AAPG Studies in Geology 64, p. 273-311.

Joseph

超过字数限制了，所以截成两段后图有些窜位。抱歉！

Word版本可以下载：

游客，如果您要查看本帖隐藏内容请回复

~仲卿寅瑾梓墨~

谢谢郝哥的精彩分享！~

yuejie215

谢谢了啊

x-7751158

感谢郝老师分享！

qundao2000

谢谢郝哥的精彩分享！~

joylin

good paper

请问郝哥

图中井上显示的是什么数据，怎么弄成这种效果的啊

leochen

谢谢分享

pubangyi

感谢分享