You are on page 1of 8

Calibration and ranking of pore-pressure prediction models

MARIO A. GUTIERREZ, NEIL R. BRAUNSDORF, and BRENT A. COUZENS, Shell International Exploration & Production, Houston, USA

Pore-pressure prediction employs a broad range of method-


ologies to estimate fluid pressures including porosity and
depth-based trends, seismic velocities, and multivariate
regressions. Model building is typically characterized by an
iterative sequence that includes model identification, cali-
bration, selection, and diagnostic checking. To predict effec-
tive stress and pore pressure, practitioners apply a diverse
set of functional forms that relate velocities and fluid pres-
sures, using, in addition to seismic velocity, predictive vari-
ables like porosity, depth, temperature, etc.
Velocity-based methods, in particular power-law func-
tions, are very popular in predrill pore-pressure prediction, Figure 1. A wide variety of functional forms are available to estimate
because they are simple and generally provide acceptable effective stress and pore pressure from porosity, seismic velocity, and
estimates. One effective way to prioritize model selection depth-based trends. The most popular methods are the power-law forms
is by measuring the accuracy of model predictions. We intro- (e.g., Eaton, Holbrook, and Bowers). σ = effective stress, V = elastic
duce in this article a new pore-pressure model calibration velocity, D = depth, φ = porosity, T = temperature, A = age.
and analysis method, where multiple functional forms are
calibrated by minimization of the difference between model-
predicted and actual measured pore pressures. With this
approach, the quality of the pressure estimation is quanti-
fied and ranked using prediction error statistics. Residual
plots provide additional model diagnostics, highlighting
systematic errors and the effective predictive range as a
function of potentially important independent variables.
This new technique allows the user to calibrate, compare,
analyze, and manage a large number of pore-pressure pre-
diction models given appropriate inputs. Examples high-
lighting the approach also are presented.

Pressure prediction and oil and gas exploration.


Overpressures in the subsurface pose major problems for
safety and cost-effective well design. Furthermore, geo-
pressures impact prospect and play appraisal and econom-
ics in a number of ways. A partial list of the issues associated Figure 2. Model calibration and ranking workflow.
with subsurface geopressures includes:

• drilling safety and cost: rig selection; well kicks and


blowouts; lost wells; wellbore stability problems; mud
expense and mud loss; stuck pipe; formation damage;
extra casing runs
• environmental risks
• prospectivity and trap analysis: hydrocarbon retention
and column height; sealing/nonsealing faults; top seal
capacity; aquifer continuity/pressure support during
production; volumetrics and economics
• impact on rock and fluid properties: reservoir quality;
sediment and fluid acoustic properties and quantitative
seismic interpretation

Consequently, the prediction of the occurrence and mag- Figure 3. Location of the deepwater Gulf of Mexico data set.
nitude of overpressures and the associated trap integrity (dif-
ference between the minimum horizontal stress and the tic basins is disequilibrium compaction. Disequilibrium
pore pressure) are of considerable importance in hydrocar- compaction geopressures result when part of the weight of
bon exploration. the overburden is transferred to the pore-fluid pressure dur-
ing loading. Rapid burial rates of mud-rich (low-perme-
Overpressure mechanisms. Overpressures can be caused ability) sediments promote the development of compaction
by compaction disequilibrium (undercompaction); tectonic disequilibrium geopressures. Overpressures due to dise-
compression (lateral stress); aquathermal expansion; hydro- quilibrium compaction are often recognized by higher than
carbon generation and gas cracking; mineral transformations expected porosities at a given depth. Unlike disequilibrium
(e.g., illitization); and osmosis, hydraulic head, hydrocar- compaction, most of the other mechanisms are either com-
bon buoyancy, etc. paratively minor or occur late in the burial history (e.g.,
Of these, the most important mechanism in Tertiary clas- hydrocarbon generation and mineral transformations).
1516 THE LEADING EDGE DECEMBER 2006
Figure 6. Input data display and screening of the basic calibration data
Figure 4. Input data set. set. FPGBML = fluid pressure gradient below mudline.

Figure 7. Error statistics. MVO = maximum value overprediction,


MVU = minimum value underprediction, SHP = subhydrostatic pres-
sure predictions, and SLP = sublithostatic pressure predictions.
Figure 5. Input data display and screening of the basic calibration data
set. FPGBML = fluid pressure gradient below mudline.

Many of these mechanisms reduce effective stress below the


maximum historical value and represent a significant pre-
diction challenge as compaction is largely an inelastic
process.

Pressure prediction and velocity. Overpressure develop-


ment in sedimentary basins reduces the effective stress of
the system. As compaction is primarily a function of effec-
tive stress, a reduction in stress will retard compaction. As
a result, any measurement that senses porosity (e.g., seis-
mic or sonic velocities, density and resistivity logs) may pro-
vide a means of estimating overpressures.
Seismic velocities increase with effective stress. The pres-
sure dependence results from the closing of grain bound-
aries, microcracks, and flaws, which elastically stiffens the
rock mineral frame. Increasing pore pressure softens the elas-
tic mineral by opening grain contacts and microcracks, tend-
ing to lower velocities. Figure 8. Gamma-EB model showing VES versus velocity.
As compaction is largely an inelastic process, only early
geopressuring (e.g., undercompaction) can reliably be pre- properties differently and the methods commonly used for
dicted from porosity or one of its proxies (velocity or resis- the prediction of early overpressures often do not provide
tivity). Late overpressures (overpressures which reduce the reliable pressure estimates in these environments.
stress below the maximum historical values) impact rock Overpressures may be predicted prior to drilling in a

DECEMBER 2006 THE LEADING EDGE 1517


Figure 9. Gamma-EB model showing residuals versus measured fluid
Figure 12. Modified Eaton model showing VES versus velocity.
pressure.

Figure 13. Modified Eaton model showing residuals versus measured


fluid pressure.

Figure 10. Gamma-EB model showing fluid pressure error versus


DBML.

Figure 14. Modified Eaton model showing fluid pressure error versus
DBML.

Figure 11. Gamma-EB model showing measured versus predicted fluid


• Effective stress drives compaction (mudrock porosity loss).
pressure. • Geopressuring retards compaction.
• Mudrock porosity and velocity can be related.
number of ways, including seismic interval velocity inver- • Seismic interval velocity reflects mudrock velocity (pre-
sion, basin modeling, and pressure cell concepts. dominantly).
Each method has advantages and disadvantages. The • Seismic interval velocities can therefore be used to predict
logic underpinning velocity-based prediction is as follows: effective stress and fluid pressure.

1518 THE LEADING EDGE DECEMBER 2006


Figure 15. Modified Eaton model showing measured versus predicted Figure 16. Bowers model showing VES versus velocity.
fluid pressure.

DECEMBER 2006 THE LEADING EDGE 1519


Figure 17. Bowers model showing residuals versus measured fluid Figure 18. Bowers model showing fluid pressure error versus DBML.
pressure.
A wide variety of functional forms are available to esti- model. The following list summarizes the major assump-
mate effective stress and pore pressure from porosity, seis- tions and limitations associated with many approaches to
mic velocity, and depth-based trends (Figure 1). The most the prediction and detection of abnormal pressures using
popular methods are the power-law forms (e.g., Eaton, seismic velocities.
Holbrook, and Bowers). These models imply that under- The assumptions are:
compaction is the dominant cause of geopressure and typ-
ically underestimate pressures associated with unloading. • Undercompaction is the dominant source of overpressure.
Only Bowers’ method employs loading and unloading veloc- • Porosity loss is controlled primarily by vertical effective
ity versus effective stress relations to account for both under- stress and is reflected by vertical interval velocity.
compaction and late overpressures. Model building is • Sediments are at or near maximum historical stress. The
typically characterized by an iterative sequence including path to maximum historical stress has little or no effect.
model identification, calibration, selection, and diagnostic • Sands and mudrocks are in approximate pressure equi-
checking. librium.
In earth sciences, a model is a simplified version of a • Lithologic and fluid variations (e.g., varying net-to-gross,
geologic and physical system that would be too complicated hydrocarbon effects, diagenesis, and nonsystematic
to analyze in full detail. Several assumptions and related mudrock variability) have only a minor impact on the
drawbacks are embedded in the generation of a simple velocity-stress relationship.

1520 THE LEADING EDGE DECEMBER 2006


• Seismic interval velocities can be
made to reasonably match well-
derived (model calibration) velo-
cities. Seismic velocities provide
the necessary vertical resolution
adequately to resolve pressure
changes.

The limitations are:

• The relationships are empirical,


and the coefficients must be cali-
brated locally. Data from an-
alogous geologic settings are
required (predrill prediction can
be difficult in wildcat settings).
• Many models extrapolate poorly.
• Many algorithms make non-
physical (subhydrostatic or super-
lithostatic pressure) predictions.
• Hydrocarbon column and lateral
pressure transmission (centroid)
effects are often unaccounted for
in model calibration.

Model calibration and ranking. In


this section, we address the following
questions with regard to model cali-
bration and ranking:

• Which are the best pore-pressure


prediction models?
• How can we efficiently rank the
models?

One effective way to prioritize the


model selection process is by mea-
suring the accuracy of model pre-
dictions. Using a new approach,
multiple functional forms or models are calibrated by non-
linear minimization of the difference between model-pre-
dicted and actual measured pore pressures (Figure 2).
With this approach, the quality of the pressure estima-
tion is quantified and ranked using prediction error statis-
tics, including rms error, standard deviation, absolute mean,
maximum and minimum error, and the number of non-
physical predictions. Residual plots provide additional
model diagnostics, highlighting systematic errors and the
effective predictive range as a function of potentially impor-
tant independent variables.
This new method allows the user to calibrate, compare,
analyze, and manage a large number of pore-pressure pre-
diction models given appropriate inputs. This utility has
input plotting and data QC capabilities and can test multi-
ple models with optimization and ranking. To highlight
with examples the available functionality, we have selected
a data set from the deepwater Gulf of Mexico (Figure 3).
The data set consists of 732 points including conventional Figure 19. Bowers model showing measured versus predicted fluid
pressure.
wireline logs, check-shot data, and high-quality RFT or
MDT pressure data from 80 wells in the deepwater Gulf of locations. For the 80-well data set, only wells containing con-
Mexico, with an additional 31 points coming from piezo- ventional wireline logs, check-shot data, and high-quality
probe data from three near-surface penetrations. Most data RFT or MDT pressure data were considered. Temperatures
come from the more heavily drilled areas of Green Canyon, were determined by linear interpolation between corrected
Garden Banks, and Mississippi Canyon. bottomhole log temperatures or by using appropriate
Wells were selected to represent a broad range of envi- regional trends.
ronmental conditions (temperature, fluid pressure gradi- Input data display and screening: In this step, the basic cal-
ent, water depth, age, burial depth) and geographic ibration data (e.g., interval velocity, depth, pressure, loca-

DECEMBER 2006 THE LEADING EDGE 1521


Figure 20. Dutta-Sims model showing VES versus velocity. Figure 22. Dutta-Sims model showing fluid pressure error versus
DBML.

Figure 21. Dutta-Sims model showing residuals versus measured fluid Figure 23. Dutta-Sims model showing measured versus predicted fluid
pressure. pressure.

tion, etc.) are read in and a variety of default plots are gen- ally we find that relationships using additional variables
erated. These plots give the interpreter an appreciation of yield better results. In practice, however, this apparent
outliers, biases in the data set, the effective range of resul- improvement is frequently counterbalanced by the uncer-
tant models, potentially important predictive parameters, tainties related with these additional calibration variables.
etc. and point out needs for additional data, further QC, etc. In our data set, the Bowers model is ranked higher than
Example plots and histograms are shown in Figures 4–6. Dutta-Sims model. Dutta-Sims includes temperature as a
Model calibration: Once the data set is successfully input, variable and these results suggest that temperature has no
a large number of models are tested, resulting in tables of (or a limited) role in pressure generation. This result may
statistics and QC plots of the optimized results (Tables 1–2, be due in part to the comparatively low, and narrow range
and Figure 7). Additional functional forms/approaches can in temperatures (mean value 68° C) of this data set. A minor-
easily be incorporated. Residual plots provide additional ity of the calibration wells in this data set approach tem-
model diagnostics, highlighting systematic errors and the peratures associated with clay diagenesis in the Gulf of
effective predictive range as a function of potentially impor- Mexico (80–100° C) or kerogen maturation.
tant independent variables. Figures 8–23 compare four dif-
ferent models: Gamma-EB, Eaton (calibrated exponent and Conclusions. New functionality has been developed to aid
base), Bowers, and Dutta-Sims. the geoscientist in calibrating, comparing, analyzing, and
Model ranking: The results of the model calibration are managing a large number of velocity-based pore pressure
used to rank the order of the individual models by provid- prediction models given appropriate inputs. We introduce
ing statistics on absolute and relative errors, nonphysical pre- a new method, where multiple functional forms are cali-
dictions, etc. (Table 3). The favored model or models can then brated by nonlinear minimization of the difference between
be used to predict pressure, stress, overburden, drilling mar- model-predicted and actual measured pore pressures. With
gin, etc. in 1D, 2D, and/or 3D with input and model uncer- this technique, the quality of the pressure estimation is quan-
tainty propagation. tified and ranked using prediction error statistics, includ-
The pore-pressure prediction models summarized in ing rms error, standard deviation, absolute mean, maximum
Table 1 use diverse numbers of fitting variables, and usu- and minimum error, and the number of nonphysical pre-
1522 THE LEADING EDGE DECEMBER 2006
dictions. Residual plots provide additional model diagnos- the road ahead” by Dutta (GEOPHYSICS, 2002). “Investigation of
tics, highlighting systematic errors and the effective pre- crustal structure and active tectonic processes in the coast ranges,
dictive range as a function of potentially important Central California” by Eberhart-Phillips (PhD dissertation,
independent variables. Stanford University, 1989). “The equation for geopressured pre-
The pressure prediction models summarized in Table 1 diction from well logs” by Eaton (SPE paper 5544, 1975). “Fracture
use different numbers of calibration parameters, and gen- gradient prediction for the new generation” by Eaton and Eaton
erally we find that relationships using additional variables (World Oil, 1997). “Porosity and pressure: role of compaction dis-
yield better results. In practice, however, this apparent equilibrium in the development of geopressures in a Gulf Coast
improvement is often counterbalanced by the uncertainties Pleistocene Basin” by Hart et al. (Geology, 1995). “The relation-
associated with these additional variables. To generate the ship between porosity, mineralogy, and effective stress in granu-
best predrill predictions, we advocate convolving input and lar sedimentary rock” by Holbrook (presented at SPWLA 36th
model uncertainties at potential well locations. In this way, Annual Logging Symposium, 1995). “Real-time pore pressure
you can test not only the impact of the model selection, but and fracture-pressure determination in all sedimentary litholo-
also the impact of data quality (e.g., seismic velocity uncer- gies” by Holbrook et al. (SPE Formation Evaluation, 1995).
tainty). “Estimation of formation pressures from log-derived shale prop-
Resultant predictions (with uncertainties) can be used as erties” by Hottman et al. (JPT, 1965). “Analysis of overpressure
input for real-time model updating (Malinverno et al., 2004). on the Gulf of Mexico Shelf” by Katahara (2003 OTC Proceedings).
Using information derived from the offset wells to define the “Integrating diverse measurements to predict pore pressure with
prior probability distribution, one can use while-drilling mea- uncertainties while drilling” by Malinverno et al. (SPE paper
surements of velocity and pressure to update the velocity 90001, 2004). “How to predict formation pressure and fracture
trend (improved velocity and depth information), and refine gradient” by Matthews and Kelly (Oil and Gas Journal, 1967). “An
the model calibration. Taken together with resistivity-based engineering interpretation of seismic data” by Pennebaker (SPE
models, one can further mitigate the risks and costs associ- paper 2165, 1968). “A global algorithm for pore-pressure predic-
ated with drilling geopressured prospects. tion” by Scott and Thomsen (SPE paper 25674, 1993). “Seismic
pressure-prediction method solves problem common in deep-
Suggested reading. “Pore-pressure estimation from velocity data: water Gulf of Mexico” by Wilhelm et al. (Oil and Gas Journal, 1998).
accounting for overpressure mechanisms besides undercom- TLE
paction” by Bowers (SPE Drilling & Completion, Paper 27488,
1995). “Detecting high overpressure” by Bowers (TLE, 2002). Acknowledgments: We thank Shell International Exploration and
“Fluid flow in low permeable porous media” by Dutta (in Production Incorporated and Shell Exploration and Production Company
Migration of Hydrocarbons in Sedimentary Basins, 2nd IFP (SEPCo) for permission to publish this work.
Exploration Research Conference, Editions Technip, 1987).
“Geopressure prediction using seismic data: current status and Corresponding author: mario.gutierrez@shell.com

DECEMBER 2006 THE LEADING EDGE 1523

You might also like