SPECIAL SECTION: i R s c k R o c k p h y s c o

p P h y s i c s

Constrained rock physics modeling

eismic reections are physically explained by contrasts in elastic properties in the subsurface, and rock physics is the science that links the seismic properties to geologic properties. Rock physics models attempt to quantify rock properties with physical parameters like clay content, saturation, and porosity, and use this to predict elastic rock properties, and further to extrapolate to scenarios that have not yet been observed. There are rock physics models made for friable sands, cemented sands, shaly sands, sandy shales, uncemented and cemented shales, to mention a few. Sometimes a large selection of available rock physics models can be more confusing than a blessing: Which model is best suited to predict the observations of a given data set, and how do we know that the model we choose is optimally calibrated to the data? This paper proposes a strategy to evaluate numerous rock physics models in a short time. The method nds the best prediction for each model, which makes it easy to see which models are suited for further modeling. The strategy is applied on a data set from the North Sea Brent group. A selection of the rock physics models with the obtained optimal input parameters is further tested to evaluate the ability to predict data outside the calibration data set. The modeling strategy presented ensures a minimum discrepancy between modeled and recorded data, without abandoning rock physics principles.

Methodology Empirical models meant to t a given data set aim to nd some coecients that give the best match, given a mathematical framework. A simple mathematical framework can be given as an estimation of P- and S-wave velocities (V P and VS) from reservoir parameters such as porosity ( ) and clay content (Vcl) like Han et al. (1986): VP,S =a+bVcl+c with different values for a, b, and c for V P and VS . The best empirical model will commonly be the one that has coecients that minimize the dierence between modeled and observed seismic properties within the chosen mathematical framework. Multivariate nonlinear regression is an eective and robust way of nding the optimal combination of coecients. Rock physics models commonly constitute more complicated mathematical frameworks than Hans relations, but there are nevertheless almost always some adjustable parameters like a, b, and c in Hans formula. These parameters are quantities that we have not measured (or cannot measure) like constituent shapes, geometric distribution of components, and eective microstructural properties. Essentially all rock physics models contain assumptions about the rock microstructure, but in most cases the values of the parameters that express the assumptions are unknown. Common examples on such parameters are number of grain contacts and aspect ratio of constituents and pores. Due to lack of exact values of these parameters, they are commonly set to physically realistic values that make the model t the data set. Having several adjustable input parameters, a trial and error strategy soon gets
76 The Leading Edge January 2009

very cumbersome. This is where the multivariate nonlinear regression can be used. The regression can be used to eciently estimate the optimal combination of input parameter values that are not exactly known. Upper and lower limits are used on every adjustable parameter, to ensure that the result is physically realistic. The multivariate term ensures that all parameters in the regression analysis are varied simultaneously to give the best set of parameters for a given model. A regression analysis needs to have an answer to which the modeled values can be compared. If we choose to use V P and VS , we get the parameter set that best replicates the sonic logs individually. By choosing acoustic impedance (AI), V P /VS ratio, and the ratio between bulk modulus and shear modulus (k/ ratio) as the answer, the rock physics model is focused on predicting AI, V P /VS ratio and (k/ ) ratio as accurately as possible, which should ensure a comprehensive t to the data set. The iteration performed in this paper is based on Matlab algorithms that apply least square nonlinear multivariate regression by using the reective Newton method, as described in Coleman and Li (1994, 1996). The results are slightly dependent on the starting values in the iteration process, but by using the mean of upper and lower constraints, the results are robust and do not improve signicantly by perturbing the values around the mean. The next section will discuss some parameters that can be considered adjustable in rock physics modeling. Rock physics input parameters Contact theories normally apply a certain number of grain contacts per grain, often called coordination number (C). This is traditionally either a static number or given as a function of porosity. The static coordination number is often used as a fudge factor that the rock physicist assigns to the rock to make the model t the data. The relation between number of grain contacts and porosity can also be used as a fudge factor, since the relation is likely to vary from rock to rock as a function of, e.g., rock texture, sorting, and crushing of grains. The same arguments can be used for parameters that describe idealized shapes, like aspect ratio of pores and grains. In addition, the coordination number or aspect ratio that gives the best result for one model may not give the best result for another model, given that all other parameters that are common are unchanged. Therefore such parameters, whether they are functions of other rock parameters or static numeric values, should be calibrated to the given data set. What is the bulk modulus of the solid part of a shale? There is no single answer because shales are composites of large varieties of minerals. And a given assemblage of minerals will change in composition and properties with burial due to diagenetic reactions. Even in the few cases where the shale mineralogy is known, literature values of the elasticity of each mineral have to be used to estimate the eective solid prop-

R o c k

p h y s i c s

erties. Literature values of clay mineral properties are highly varying and uncertain, and the models used to estimate eective solid properties rely on idealized assumptions about the mineral arrangement. Experience has shown that the solid parts of shales are normally signicantly softer than the more reliable elasticity values of quartz, which can be used as constraint. But how certain are we that the lowest values of a gamma-ray log correspond to clean quartz? Recently Homann and Batzle (2008) demonstrated that the eective grain bulk modulus in sandstones is variable due to even small concentrations of both clay and nonclay minerals. This supports using constraints rather than quartz mineral moduli also for clay-free sandstones. When other parameters such as critical porosity of sand and shale, or amount of precipitated cement are unknown, upper and lower bounds that are physically reason- Figure 1. The data set used for rock physics modeling in this paper. able can be used to optimize the rock physics modeling. There are clear lithological trends, with V P/VS ratio increasing with In general, literature, lab clay content. The lowest V P/VS ratios are for hydrocarbon-lled sands. experiments, seismic, and geological knowledge can all contribute to dene upper and lower constraints on various input parameters to rock physics modeling. There are also uncertainties in log parameters such as clay content, porosity, pore pressure, and velocities. These are not considered here, but can be added as an uncertainty to the estimated parameters from the rock physics Figure 2. Prediction accuracy ranged from left to right for a variety of rock physics models. Blue is the deviation from well log AI. Green is the deviation from the V P/VS ratio. Brown is the deviation from k/ model. Demonstration of strategy Figure 1 shows a well-log data set in the AI-V P /VS plane. The database consists of logs from 10 dierent wells. The logs are quality checked with internal lithology and uid prediction routines (LFP), which ensure comprehensive consistency between the dierent logs. Some wells did not have VS logs, and in these cases VS was estimated as a part of the LFP work. By using the porosity, saturation and clay content from well logs, we analyze which rock physics model is most suitable to describe the data set within upper and lower constraints of adjustable parameters. Table 1 shows some parameters which, in this modeling, are considered not absolutely certain. The parameters are constrained to make the models t rocks which range from clean quartz sandstones to pure shales. The solid clay moduli range from the values given by Tosaya et al. (1982) to the values given by Han et al. (1986). If there is any cement, it is set to have the same constraints as quartz. Hence, we use some apriori knowledge or assumptions to constrain the modeling. Figure 2 shows the prediction accuracy of some wellknown rock physics models. Parameters other than those in Table 1 are also used as iteration parameters in the modeling. Examples are the amount of cement in the cement models

ratio. All are given as percentages. The bars are ranged from left to right with increasing accumulated deviation from the data set. The rock physics models are described in Avseth et al. (2005). They are named (left to right) empirical approach, constant shale, Hertz-Mindlin-Hashin-Shtrikman (friable sand), constant cement, laminated sand shale, contact cement, shaly sand model, Hertz-Mindlin, silty shale, and grain-coating contact cement. The predictions are obtained using the narrow constraints of case 1 in Table 1. UC 1 Kcl cl Kcem cem Kq q a b 25 9 37 45 37 45 22, 5 36, 5 LC 1 21 7 35 43 35 43 17, 5 31, 5 UC 2 35 20 40 50 40 50 25 39 LC 2 15 5 25 20 25 20 15 29

c 19 9 21 7 Table 1. Values that constrain input parameters to rock physics modeling. cl, cem and q denote clay mineral, cement and quartz, respectively. a, b, and c are coecients that describe the number of grain contacts as a function of porosity: C = a b + c2. UC and LC are upper and lower constrain values for each parameter. The numbers 1 and 2 indicate case 1 and 2. The values in case 1 span short intervals and are referred to as narrow constraints, while the values in case 2 span a larger interval and are referred to as wide constraints.
January 2009 The Leading Edge 77

R o c k

p h y s i c s

(00.3) and critical porosity for shales and sands (0.60.8 and 0.35-0.45, respectively). A reduced tangential shear factor (slip factor) is included to obtain optimal match with observed V P /VS ratios in the sandstones, since contact theory is known to overpredict shear-wave velocities by neglecting rotational freedom and slip at grain contacts. An empirical model is also included. It is seen in Figure 2 that the empirical model will probably Figure 3. Prediction accuracy with wider constraints. The rock physics models are described in be the most suitable for predictions if Figure 2. The predictions are obtained using the wide constraints of case 2 in Table 1. Blue is the no extrapolation is needed, since it has deviation from well-log AI. Green is the deviation from the V P/VS ratio. Brown is the deviation the lowest prediction deviation. The from k/ ratio. All are given as percentages. empirical model is based on the four most relevant and available parameters for the studied data set: porosity, clay content, dierential pressure (Pd), and uid bulk modulus (Kf ). The equation is based on regression with nine coefcients each for AI, V P /VS ratio, and k/ ratio: AI,VP/VS,k/=a2+b+cVcl 2 +dVcl+ePd2+fPd+gKf2+hKf+i. The uid bulk modulus is derived from saturation. This is a highly adaptive model with a high number of coecients. In cases with insucient knowl- Figure 4. Blind test of the ve best models in Figure 2 and Figure 3. (a) Deviation from welledge about rock composition, upper log data with narrow constraints as given in case 1 in Table 1. (b) Deviation from well-log data and lower constraints can be extended with wide constraints as given in case 2 in Table 1. The coecients in the empirical model do not for some parameters to include all have constraints, and the model gives the same result in both (a) and (b). Blue is the deviation contingencies. If the constraints on all from well-log AI. Green is the deviation from the V P/VS ratio. Brown is the deviation from k/ ratio. All are given as percentages. The models range from best (to the left) to worst predictions (to solid properties are wide, as case 2 in the right). Table 1, the model predictions genModel Slip Kcl cl Kcem cem Ksst sst erally become more accurate, since the solution space increases. But if the constraints are wide on HMHS 0.67 21 7 37 43 all of the parameters, there is a danger of compenCCT 0 25 9 35 43 37 45 sating for one unlikely value with another, and ConstCem 0.5 25 9 35 43 36 45 yet still get a good t to the data. In theory, the ConstShale 0.63 25 7 37 43 constraints in case 2 allow for unrealistic material properties where the sand can be softer than the HMHS 0.77 23 6 37 38 clay, and clay shear modulus can be higher than CCT 0.63 23 8 25 23 39 50 the clay bulk modulus. Case 2 uses the rock physConstCem 0.83 28 12 39 43 37 41 ics model in a more empirical way, with less emConstShale 0.75 24 7 37 41 phasis on physical principles. Therefore, it is also less reliable when extrapolating outside the range Table 2. Comparison of parameter output from four models in the regression of the dataset. Figure 3 shows that most models analysis. The upper part of the table corresponds to case 1 in Table 1, while the predict AI, V P /VS , and k/ ratios with smaller lower part corresponds to case 2. The modeling results can be seen in Figure 2 and Figure 3. deviations when the constraints are wider. The reason why V P /VS ratio deviations are consistently lowest is that the regression method estimates the actual gives the best t with narrow constraints. Table 2 compares deviation and not the percentage. Another weighting of the the actual parameter values that were used to obtain the redeviation would have levelled out the dierences. One pitfall sults in Figures 2 and 3. When the constraints are wide, some with wide constraints is, however, that the model can con- unrealistic elasticity values can be observedlike sandstone verge to a local minimum instead of a global minimum when grain bulk and shear moduli of 39 GPa and 50 GPa for the estimating the minimum deviation between modeled and re- CCT model. The moduli are in the expected range when the corded data. This is the case for the shaly sand model, which constraints are narrower. Table 2 and the results in Figure 2
78 The Leading Edge January 2009

R o c k

p h y s i c s

Figure 5. Predicted data (white) superimposed on the well-log data. Well-log data are colored by clay content. Only sands with clay content less than 0.2 and shales with clay content higher than 0.8 are shown for visualization purposes.

and Figure 3 further show that even though changing some input parameters for a given model signicantly, the accuracy of the prediction does not necessarily change much. This is due to the counteracting eect of various parameters; for example, a change in clay properties can be partly compensated by a change in nonclay properties, but the improvement in the prediction may not be signicant. The database that was used to calibrate the rock physics models did only contain porosities between 0.10 and 0.35. The well logs also contain data with porosities below and above this interval, and these data were used in a blind test of the ve best models in Figure 2 and Figure 3. The empirical model does not have constraints, while the four rock physics models are evaluated for both narrow and wide constraints on input parameters. Figure 4 shows the results from the blind test. The CCT and constant cement model predict considerably better results when the constraints are narrow, while the two other models are practically equal for narrow and wide constraints. The highly adapted empirical model is no longer the best model to describe the data, but is passed by the two best rock physics models. This is expected, since the empirical model does not obey any physical laws when it extrapolates. All ve models in the blind test predict data that are highly correlated to the well data. The correlations between recorded and predicted AI, V P /VS ratio, and k/ ratio vary between 0.75 and 0.95 for the various models, both with narrow and wide constraints. This indicates that the models manage to mimic trends in the data outside the calibration database; predicted AI increases when observed AI increases and so on. In general very wide constraints allow for physically unrealistic parameter combinations. Therefore case 1 is likely to give the most reliable extrapolation results, since it honors physical properties expected in the rock to a higher degree than case 2. Figures 24 show that two rock physics models stand out as especially well suited for prediction of the data, and they deviate little from each otherthe HMHS model and the constant shale model. Both models assume that the

rock is uncemented. The constant cement model predicts that small amounts of cement are present. This model is a combination of the CCT model and the HMHS model. Testing the various models in this study indicates that the rock contains little or no cement, and that porosity variations are primarily due to clay content and sorting eects. This is also consistent with what we know about the majority of the Brent Group sands in this area. The HMHS model with narrow constraints is chosen as the rock physics model to proceed with, and Figure 5 shows the prediction accuracy in the AI-V P /VS plane. Only clay content outside the interval 0.20.8 is used to improve visibility. The model seems to capture shale trends, sand trends, and uid trends quite satisfactorily. The eect of oil saturation in clean sands is a decrease in both AI and V P /VS ratio, which is what we observe both in the data and for the model. Gassmann theory is combined with the HMHS model to account for saturation eects. After obtaining a reliable rock physics model, it can be used to extrapolate to new combinations of porosities, clay contents, saturations, and dierential pressure, to see what seismic properties can be expected. Figure 6 shows some rock physics templates created from the HMHS model with narrow constraints on input parameters. In Figure 6a, the template is created with constant dierential pressure of 20 MPa. The data plots nicely upon matching colors in most cases, and in some cases the template extrapolates to parameter values not found in the dataset. Much of the data has dierential pressure lower or higher than 20 MPa, which were used in Figure 6a. Therefore dierential pressure is included as a third dimension in 6b and 6c, which make the template even more consistent with the data. Figure 6b shows that we can expect smaller contrast in V P /VS ratio between clean sands and pure shales when dierential pressure decreases. Further, the AI range of shale decreases with increasing pressure, while the AI range of sandstones is less aected. In 6c, the model shows that uid sensitivity decreases with increasing pressure, which is expected since the rock framework stiens with increasing pressure. The rock physics modeling performed here applies one model for all porosities, clay contents, and saturations. An even better result could have been achieved by splitting the database up in various lithology classes like shale, hard sand, and soft sand, and nding the best model for each class. When using the rock physics models to predict rock microstructure, a division into lithology classes gives the best results. A few of the samples from Brent Group sands in this study actually contain quartz or calcite cement. These are not captured when applying all data at once. Conclusions A method for fast evaluation of rock physics models for a given data set has been presented. This method nds the model and corresponding set of input parameters that result in the lowest discrepancy to the data. Wide constraints on input parameters commonly result in a better t to a given data set, which is also shown in this study. A blind test showed
January 2009 The Leading Edge 79

R o c k

p h y s i c s

that the model parameters obtained resulted in robust models for prediction of data outside the calibration data set. The predictive power of the models increases when the constraints are within physically realistic bonds; hence, the best model parameters for a given data set are not necessarily the best for prediction about new data sets. If a model successfully manages to predict the data, it can be used to extrapolate to nd seismic properties for new parameter combinations not found in the original data set. Rock physics templates can be made to visualize trends in the data. Modeling performed in this paper shows that varying some adjustable parameters within narrow sensible bounds, based on assumptions or knowledge about the rock parameters and local geology, will give the best possible t to the data without abandoning the physical principles of the model. The method presented can be used further to predict details in the rock microstructure, since practically all rock physics models are based on an idea of the rock microstructure. The presented strategy is highly exible, since all kinds of rock physics models can be evaluated for prediction abilities of all kinds of seismic properties.
Suggested reading. Quantitative Seismic InterpretationApplying Rock Physics Tools to Reduce Interpretation Risk by Avseth et al. (Cambridge, 2005). On the convergence of reective Newton methods for large-scale nonlinear minimization subject to bounds by Coleman and Li (Mathematical Programming, 1994). An interior, trust region approach for nonlinear minimi- Figure 6. (a) A two-dimensional rock physics template (pressure is constant) created using the zation subject to bounds by Coleman HMHS model with input parameters obtained from the narrow constraints in Table 1. Oilsaturated sandstones are gray and get increasingly blue with increasing brine saturation. Clean and Li (SIAM Journal on Optimiza- brine-saturated sandstones are yellow, and get increasingly green with increasing clay content. tion, 1996). Eects of Porosity and Clay (b) Three-dimensional rock physics template with well-log data, showing brine-saturated Content on Acoustic Properties of Sand- sandstones (yellow) and shales (green). This template includes pressure and therefore depth stones and Unconsolidated sediments by trends. (c) Three-dimensional rock physics template with well-log data, showing sandstones with Han (PhD dissertation, Stanford, 1986). saturation ranging from oil-lled (gray) to brine-lled (blue). Eects of porosity and clay content on wave velocities in sandstones by Han et al. (Geophysics, Acknowledgments: I thank Per Avseth and Harald Flesche for valu1986). Measurements of the eective rock grain modulus by able discussions and comments, and StatoilHydro and the Gullfaks Homann and Batzle (SEG 2008 Summer Research Work- partner Petoro for permission to use the data. shop). Acoustical Properties of Clay-Bearing Rocks by Tosaya (PhD dissertation, Stanford, 1982). Corresponding author: andd@statoilhydro.com
80 The Leading Edge January 2009