SPE 128350

Modelling Subsurface Uncertainties with Experimental Design:

Some Arguments of Non-Conformists
Kazeem A. Lawal ∗ , SPE; Shell Nigeria Exploration & Production Company

Copyright 2009, Society of Petroleum Engineers Inc.

understanding of the effect of a controllable variable on a
rd
This paper was prepared for presentation at the 33 Annual SPE International Technical Conference and dependent variable is required, or where the range of
Exhibition in Abuja, Nigeria, August 3-5, 2009.
uncertainties is well known within a narrow interval. These
This paper was selected for presentation by an SPE Program Committee following review of information
contained in an abstract submitted by the author(s). Contents of the paper, as presented, have not been include production/injection management, ‘model-based’
reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The
material, as presented, does not necessarily reflect any position of the Society of Petroleum Engineers, its
control algorithms for ‘intelligent’ completions, business
officers, or members. Papers presented at SPE meetings are subject to publication review by Editorial
Committees of the Society of Petroleum Engineers. Electronic reproduction, distribution, or storage of any
planning and similar areas of the E&P business
part of this paper for commercial purposes without the written consent of the Society of Petroleum characterised by continuous data, and where the
Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300
words; illustrations may not be copied. The abstract must contain conspicuous acknowledgement of independent variables could be engineered for the desired
where and by whom the paper was presented. Write Librarian, SPE, P.O. Box 833836, Richardson, TX
75083-3836, U.S.A., fax 01-972-952-9435. objectives.

Abstract Introduction
As a consequence of limited capability for the acquisition, Risks, recognised herein as threats and opportunities,
analysis and interpretation of subsurface data, pervade the exploration and production (E&P) industry.
uncertainties pervade the Exploration and Production While the industry has apparently developed “pragmatic”
(E&P) business. To minimise investment risks, robust management strategies for economic, political, security
development plans, premised on adequate understanding and other macro-scale risks, technical risks are still an
of uncertainties, are critical. Experimental Design (ED), area of exposure. Given that technical risks are
complemented with Response Surface Method (RSM), subsurface and surface, mitigating subsurface risks is
which uses a statistical proxy equation to model the more challenging. This is a consequence of man’s limited
response (dependent variable) as a function of ability to influence natural processes, especially those that
independent variables (uncertainties), is a common predate man, such as formation of hydrocarbon reservoirs.
method for studying subsurface uncertainties.
Subsurface uncertainties (realisations), which are
In this paper, current applications of ED to subsurface elements of subsurface risks, largely relate to subsurface
modelling are evaluated from fundamental principles- data- acquisition, processing, interpretation and utilisation.
mathematical and physical consistencies of the proxy How representative of the reservoir is the data? How
equations, as well as robustness in modelling reliable are the data processing procedures/algorithms?
uncertainties. Within the context of modelling and Are the interpretation assumptions valid? How
mitigating subsurface uncertainties, major shortcomings of accurate/convergent are the simulator’s interpolation of
the ED and their implications for decision-making are data, and approximations of the conservation and
highlighted. These include inconsistency and non- equilibrium equations? These are some sources of
uniqueness of proxy models, violation of basic theoretical subsurface uncertainties in E&P projects. Undoubtedly,
physics, non-preservation of the correlation between understanding these uncertainties and their impacts is a
variables that are known to be inherently related, non- prerequisite for mitigating the associated risks.
controllability of input variables, under-estimation of the
impact of uncertainties, and the challenge of constructing In response to this challenge, various subsurface
(interpolating) realistic simulation models from an ED modelling techniques, including deterministic method (1)
output. and Bayesian design (2) and their variants (3), have been
developed. However, as evident from the trend of recent
Although ED is consistent with statistical principles, its publications on this subject, experimental design (ED),
description of reservoir physics is not satisfactory. In its coupled with response surface method (RSM), is the most
present form, reservoir complexities are apparently too popular technique. Although the recent surge could partly
overwhelming for reliable modelling or optimisation by the be explained by increased level of activities induced by oil
proxy models. Consequently, it is recommended that the price (Fig. 1), perhaps, the main reason is higher risks of
application of ED be limited to situations where a simple developing new discoveries, characterised by relatively

∗
Currently a PhD student at Imperial College London
2 K.A. Lawal SPE 128350

marginal returns and greater complexities. Typical other RSM methods such as kriging (21), artificial neural
upstream applications are parametric studies, network (22), thin-plate splines (23) and amplitude factor
performance prediction, history-matching and optimisation analysis (24), which though are more sophisticated than the
(4-14)
. least squares method, are not necessarily superior in
performance (25).
In principle, ED uses empirical means to establish a
statistical relationship between an objective variable and The main objective of this work is to analyse current
the factors influencing it (15, 16). In current E&P applications, modelling of subsurface uncertainties by ED, highlight the
the outcome of ED is a proxy model, which in theory, is a shortcomings and provide suggestions for improvement.
surrogate of the reservoir simulator. The ultimate objective The paper proceeds by identifying the respective issues,
is to minimise the number of simulation runs (experiments) augmenting the critiques with examples, and identifying
required to understand reservoir behaviour, thereby scope for more consistent and value-adding applications in
reducing project cycle time. the E&P business.

The chemical process industry (CPI), characterised by

controllable variables, has huge ED experiences (17, 18). In 1.0 Mathematical Consistency of Proxy Models
CPI applications, proxy models are based on actual plant For an analysis based on consistent theory, fundamental
input and response. Ironically, this contrasts the current correlations between the response variables should be
practice in E&P where, despite the non-controllability of preserved, even when the proxies are derived
subsurface variables (e.g. permeability and saturation independently. For example, hydrocarbon initially in-place
distributions, reservoir architecture, etc), proxies are based (HCIIP), recovery factor (Rf) and ultimate recovery (UR)
on simulator output and not actual reservoir performance. are known to have a fundamental relationship (Eq. 1).
But how reliable are our reservoir models (input data) and Using any of these as response variable, a proxy model
numerical schemes? How about grid orientation and could be developed. Intuitively, such proxies should satisfy
upscaling issues? Eq. 1. Using examples 1A and 1B, we investigate the
hypothesis that irrespective of the response variable
Kabir et al (19) queried the modelling of discrete variables
(HCIIP, Rf, or UR) used to derive the proxy equation, Eq. 1
such as stratigraphy in continuous probabilistic terms.
should be satisfied.
They also criticised the inclusion of controllable variables
(e.g. well count), which are decision factors, into the proxy
model of uncontrollable subsurface uncertainties. By UR
Rf = x100 ………………………..……………….(1)
incorporating well count into the design matrix, Peng et al HCIIP
(4)
wondered how the wells would be optimally placed
under different subsurface realisations without having
biased conclusions. There is also the tendency of such Example 1A: This is a model of a hypothetic depletion-
controllable variables masking the subsurface drive gas reservoir. Model uncertainties are structural
uncertainties (19). extent, porosity and permeability. Eqs. 2-4 are the
response surfaces of the dependent variables obtained by
Proxy models are not perfect. They may appear employing full factorial design (FFD) and regression
statistically sound, yet generate unrealistic (artefacts) method. Although not shown, the proxy predictions
forecasts. Amudo et al (20) discussed an example where compared very well (R2 > 0.98) with the simulator output.
the probabilistic water-cut prediction at the field level is
statistically correct but not at the well level- some
probabilities had more than 100% water cut! Although this
G = −399.3 + 0.028L + 0.028W + 1273.6φ …....(2)
anomaly was attributed to the quadratic nature of the proxy
model, regrettably, they did not report the performance of G p = −314.6 + 0.025L + 0.025W − 0.17k + 1107.2φ ………....(3)
a non-quadratic model to establish this claim.

Lawal and Eweje (3) raised concerns on the tendency of R f = 89 .9 − 7.433 x10 −5 L − 7 .506 x10 −5 W + 0 .001k − 5 .689φ ……
ED to mask the effects of subsurface uncertainties, and its ……………………………………………...…………...(4)
implications for decision-making. To mitigate these and
other challenges, they proposed the structured G = gas initially in-place (Bscf), Gp = gas UR (Bscf), Rf =
deterministic technique as a more pragmatic method for gas recovery factor, RF (%), L = reservoir length (ft),
modelling subsurface uncertainties. W = reservoir width (ft), φ = porosity, k = permeability (mD)

Ironically, despite the plethora of publications, very few Substituting Eqs. 2 & 3 into Eq. 1, the following results:
highlighted the shortcomings of the method. To the best of
our knowledge, there has not been any critique of the ED
−314 .6 + 0.025 L + 0.025W − 0.17 k + 1107 .2φ
from a strictly quantitative viewpoint. Consequently, Rf ≡ ……..(5)
employing quantitative arguments, this article highlights − 399 .3 + 0.028 L + 0.028W + 1273 .6φ
some issues with current applications of ED in E&P.
Clearly, Eq. 5 is in no way the same as Eq. 4. Therefore,
Because of its relative simplicity and popularity, this paper the functional relationship between Rf, Gp and G is not
focuses on the least squares regression technique of
conserved. In fact, considering only Eqs 2-4, it is difficult to
RSM. However, most of the conclusions are valid for the
3 Modelling Subsurface Uncertainties with Experiemental Design: Some Arguments of Non-Conformists SPE 128350

claim that the dependent variables are fundamentally Eq. 12 as the nonlinear regression model for oil recovery
related. in an oil reservoir. Again, the time-derivative of Np is zero.
Example 1B: This is a huge offshore gas reservoir (Fig. N p = 19.84 + 7.088S + 2.363F + 4.95C + 4.325 A + 0.163SF
2), earlier discussed by Lawal and Eweje (3). Using FFD .…
+ 1.95SC + 1.95SA + 0.85 FC + 1.425FA + 0.313CA
augmented with two centre-points and creating the surface
………………………………………………………...(12)
with a regression method, the non-linear model in Eq. 6 is
the ‘best’ surrogate of the simulator’s UR. Eqs. 7 and 8 are
A = aquifer size, C = oil-water contact, F = fault
respectively the RF and FGIIP proxies.
conductivity, S = static model. Besides Np (MMstb), all
variables are normalised (-1 to 1).
G p = 259 .94 + 28 .24 F + 35 .16 SC + 20 .95 A + 125 . 67 C + 86 .85 S …
Technically, these examples erroneously infer that Np is
.…………………………………………….…………..(6) independent of the oil rate hence exactly the same proxy
model would have been obtained if the reference had been
R f = 72.82 − 0.25S + 7.12F − 0.40C + 3.85 A ……….(7) year 5 or 50! Consider another interpretation- it could be
argued that although there is no physical production (zero
oil-rate), statistically, oil is still recoverable! Obviously, this
G = 479.8 + 240.4S + 109.3C ….…………………..…..(8) is misleading. In simple terms, the models could not
isolate the effect of oil rate because it ‘sees’ oil rate as a
A = aquifer size, C = gas-water contact, S = static model, dependent variable (similar to oil recovery), defined
F = fault conductivity. All the independent variables are statistically by some other independent factors.
normalised (-1 to 1). As observed for Example 1A, Eqs. 6-
8 are not consistent with Eq. 1. 3.0 Integral of Proxy Models
Similar to derivatives, some variables are simply the
2.0 Derivative of Proxy Models integrals of other variables. A good example is the
Theoretically, some physical variables are simply the estimation of oil recovery over a time interval by
derivatives of other variables (‘source’). Hence, a good integrating, relative to time, the average oil rate function
measure of the reliability of a ‘source’ proxy model is the over the same interval.
ability of its derivative to yield the expected physical
For the reservoir discussed in Example 2B, another study
variables. For example, for a producing system (Fig. 3), was performed to generate a response function for the
average oil rate over any time interval, including the
average oil rate. Eq. 13 is the resulting model.
lifetime, is the time derivative of the recoverable volume
over that interval (Eq. 9).
Qo = −4894.6 + 428.5k + 92.6ho − 182.4s
.…….…...(13)
dN p − 0.012k 2 − 0.98ho 2 − 1.25 Kho
Qo = ……………………………………………...(9)
dt
Qo = oil rate (stb/d), K = permeability (mD), ho = oil column
Example 2A: The proxy model derived by Portella et al (26) (ft), s = skin factor.
is used to investigate the possibility of deriving the average
oil rate consistent with a UR proxy. Invoking a 33 factorial Considering that the derived Qo is time-independent, its
design with oil recovery as the response, Portella et al (26) integration over an interval t is simply given by Eq. 14.
obtained the following regression equation. Curiously; there is no similarity between Eqs. 14 and 12.

t
N p = 18.787 + 6.220 g + 1.046k + 3.979C
+ 1.727 gC + 2.751g 2 − 1.047k 2 − 3.926C 2
….…..……...(10) ∫
N p = Qo dt = Qo t ………………………………..….(14)
0

Np = oil recovery (MMstb) at end of the 12th year, Interestingly, a check on the Darcy equation (Eq. 15)
g = net/gross ratio, C = oil-water contact, reveals that Qo only has a linear, and not quadratic,
k = permeability. All the independent variables have relationship with k and ho. Although Eq. 13 is statistically
continuous values ranging from -1 to 1. correct, it is inconsistent with the physics of porous media
flow. Hence, its use for any reservoir engineering analysis
Applying Eq. 9 to Eq. 10, we have: is debatable.

Qo = 0 ………………………………………............….(11) Qo ∝ kho ….………………………………………….…(15)

Obviously, if the oil rate was indeed zero, there would not
have been an Np at the end of the 12th year. 4.0 ED-Based Correlations
Some situations may not warrant the immediate
Example 2B: In one of their field examples, Lawal and deployment of numerical simulators. Under such
(3)
Eweje used a 24 FFD with two centre-points to derive circumstances, which include screening and preliminary
4 K.A. Lawal SPE 128350

engineering designs, engineers typically rely on horizontal completion, the impact of vertical permeability is
correlations. ED is one method of developing statistically conspicuously missing.
significant correlations. With two examples, the physical
consistency of such ED-based correlations is examined.
5.0 Masking of Uncertainties
(27)
Example 4A: Employing factorial design, Chu Two related, but distinct, issues are discussed here. The
developed the following correlation β for the prediction of first concerns the distribution model of the input variables.
steamflood performance- recovery factor (%) and project The second is the range of input variables. Both cases are
life (years). illustrated with stock-tank oil initially in-place (STOIIP)
calculated with Eq. 18.
y = d o + d1(So − 50) + d 2 (μ − 7000) + d3 (h − 80) + d4 (is − 300)
( )
+ d5 ( A − 2.6) + d6 f g − 55 + d7 (φ − 0.31) + d8 (k − 3500) + ... N=
GRV * NTG * φ * Soi
Boi
………………...…………….(18)

….……………………………………….……………....(16)
Example 5: Table 1 presents the input data for the
y = Rf (%) or project life (years), μ = oil viscosity (cP),
simulation and the STOIIP’s obtained from deterministic
So = initial oil saturation (%), h = reservoir thickness (ft),
(worst/best-case) calculations. It should be noted that the
φ = porosity (%), A = pattern size (acre), is = steam rate low and high data are the known ‘minimum’ and
(b/d),
‘maximum’ values, corresponding to P0 ξ and P100
fg = steam quality (%), k = permeability (mD).
respectively. From available data, these ranges and
Although the original paper has much longer equation (45 STOIIP’s are the limits of the system.
coefficients) than presented here, space consideration has For the probabilistic simulation, uniform, triangular, and
restricted us to only the main (linear) effects. However, the lognormal input distribution functions are evaluated. For
truncation does not affect the logic of this discussion each case, all the input variables are assumed to have the
because all the coefficients (d’s) are non-zero and do > 0. same distribution. However, the curves are bounded by
Physical consistency of the model is studied as follows. the P0 and P100 data. With triangular distribution, the
a. k = 0. For an impermeable rock, Eq. 16 predicts non- base values are assumed to be the likeliest.
zero RF and project life. One wonders the basis of Fig. 4 depicts the STOIIP distributions and the
such performance in an absolutely impermeable rock. deterministic bounds. While the uniform and triangular
b. h = 0 or So = 0. Despite a non-reservoir or dry results conform to the deterministic limits, the lognormal
discovery, non-zero RF and project life are statistically outcome is non-representative. Relative to the
feasible! deterministic P0 STOIIP, the deviations of uniform,
triangular and lognormal assumptions are 16, 54 and -62%
Example 4B: Combining ED and multivariate respectively. Corresponding deviations for P100 are -12, -
regression, Kabir et al
(28)
proposed Eq. 17 for the 21 and 146% respectively. That is, while uniform and
screening and evaluation of thin-oil reservoirs. The triangular models underestimate the range of STOIIP
correlation was reportedly tested and validated with uncertainties, lognormal exaggerates it considerably. Put
independent sets of experimental and published field data. in another way, whereas uniform and triangular models
undervalue the risks and reward, lognormal amplifies
them.
R f = −24.626 + 1.722ho + 9.687 x10−4 kh + 3.171Sorw
It is not uncommon to see some engineers (20) constrain
+ 2.062 x10−3 Lw + 0.276hGOC + 4.983 x10− 4 q − 0.026ho 2 the range of input variables, and using, for example, P10
+ 1.482 x10− 4 qho − 0.019qSorw and P90 as the bounds of the input factors for probabilistic
…………………………………………………………....(17) simulations. Ironically, from the results obtained with these
constrained data, P0 and P100 are estimated for the
Rf = RF (%), ho = oil-column thickness (ft), q = oil rate dependent variable! This is double counting. The
(stb/d) implication of this practice is investigated with the data
kh = horizontal permeability (mD), Sorw = residual oil shown in Table 2, which are P10 and P90 of the input data
saturation to water (fraction), Lw = horizontal-well length used in Table 1.
(ft), hGOC = well standoff from the gas-oil contact (ft)
Comparing the deterministic results of Tables 1 and 2, the
‘constrained range’ simulation causes 82% increase in the
In Eq. 17, where there is no oil-rim (ho = 0), oil recovery is
‘actual’ P0 and 33% reduction in the ‘actual’ P100 STOIIP.
still a possibility! Similarly, for an impermeable (kh = 0)
As observed in the earlier example, probabilistic
system, there could still be oil reserves! Surprisingly, for a
simulations cause additional deviations from the
deterministic ‘constrained’ results (Fig. 5).

β
Structurally, the correlations are essentially the same for both RF and
ξ
project life. They only differ in the coefficients (d’s). The correlations PX = there is an X % probability that the outcome would be less than a
are not limited to specific data ranges particular value
5 Modelling Subsurface Uncertainties with Experiemental Design: Some Arguments of Non-Conformists SPE 128350

model of specific probability does not resolve the foregoing

6.0 Extracting ‘Unique’ Simulation Models issues. In fact, this practice is misleading when applied to
discrete variables such as static models, fluid models, etc.
A key output from the ED is a probability distribution of the
Using the work on Chuchupa field by Rivera et al. (13),
response variable, which entails repetitive random
(29) other issues associated with the creation of ‘compliant’
sampling (Monte Carlo simulation) .
simulation models from the response distribution curve
Unfortunately,because of the randomness of the
have been highlighted by Lawal and Eweje (3).
simulation, the results (distribution) are rarely repeatable.

Furthermore, the distribution curve of the response 7.0 Dynamics of Response and Input Variables
variable implicitly assumes continuous input variables.
This is contrary to the fact that most subsurface variables In their present forms, ED and RSM implicitly ignore time
are inherently discrete, especially when aggregated into effects, as the response variable always undergoes a
say, a static model. For example, how does one transit constant change per unit perturbation of the independent
continuously in a model with uncertain stratigraphy? Or variable (e.g. Eq. 19). However, this contradicts the
how does one interpolate between different realisations of essence of ‘dynamics’ as time-dependent variations.
fault pattern?
∂R
Considering the randomness of the simulations and = Ai …………………………………………….…..(19)
∂x i
discreteness of the independent variables, deconvolving
an output data into its contributing parameters (models) is
an inverse problem. Regrettably, such inverse problems Where Ai is the coefficient of the independent variable xi in
are usually ill-posed, and have no unique solution, the R proxy model.
especially when there is non-linearity. Consequently,
∂Q o
creating a unique simulation model by combining the input Using Eq. 13 as an example, = −182.4 for all time.
variables to satisfy a particular output (e.g. P10, P50, or ∂s
P90) is questionable. Additionally, the practice of For time-dependent changes of skin factor (s), as shown in
interpolating input variables to create the model that Fig. 10 for a real reservoir, current ED theory may not be
conforms to a desired probability is even more misleading. able to reflect the effect of such dynamics on the oil rate or
other dependent variables.
Another issue is the basis on which the simulation models
should be created. Although one option is the distribution Another concern is the estimation of the time (steady-state
curve based on recovery, another equally important time) at which the calculated response, after an input
alternative is that based on recovery factor. In contrary to perturbation, would be achieved, or the path taken by the
what one might expect, these options do not yield the response variable before reaching the steady-state. For
same conclusion (run). This is investigated with the example, in Eq. 7, increasing or decreasing the aquifer
following example. size by 0.5 unit while keeping other inputs at their base
values (0), would yield RF of 70.9 and 74.7% respectively.
Example 6: For the problem discussed in Example 1B, However from the proxy, there is no idea on when the
the ED matrix and results are presented in Table 3. From reservoir would realise these RF’s, or the production
the quality of UR and RF crossplots respectively shown in profile associated with each response.
Figs. 6 and 7, the derived surfaces (Eqs. 6 and 7) are
satisfactory. Assuming uniform distribution of input factors,
8.0 Reservoir Proxy or Simulator Proxy?
UR and RF distribution curves are illustrated in Figs. 8 and
9 respectively. In principle, current application of ED to subsurface
studies merely generates a surrogate for the simulator and
The next step is the creation of P10, P50 and P90 models not the reservoir itself! This implies that uncertainties
by mapping respective probabilities to the simulation runs. associated with subsurface modelling (data) and
Table 4 shows the mapped models for UR and RF limitations of numerical simulators are simply propagated
responses. As the two responses yield different sets of to the proxy. Yet, this model is applied as the reservoir
runs, the concern is: “which set of models does one select surrogate. While this may appear reasonable where a
without compromising consistency? “ For instance, run ‘4’ ‘reliable’ history-matched model exists, it is debatable for
which approximates a P90 model in UR, is only a P4 green fields. This approach contradicts the practice in CPI
model in terms of RF! and other industries which, in spite of availability of
rigorous mathematical models, still employ actual plant
Apparenetly, the resolution of the foregoing issue is not performance data to construct proxy models of plant
trivial. While the UR-based mapping seems consistent with dynamics. This notwithstanding, such models are updated
economic analysis, RF-based models tend to be regularly as remarkable changes are observed in plant
manifestations of technical factors (e.g. recovery behaviour.
mechanisms). In resolving this issue, a pertinent question
to be anwered by the E&P industry is: “Is subsurface Although the objective of history-matching, an inverse
modelling driven by economic factors or vice versa?” problem, is to establish the most consistent and plausible
subsurface realisation which fits observed performance,
As a remark, the current practice of ‘tweaking’ the highest- the exercise is rarely unique as several, apparently
impacting variables (heavy hitters) to achieve a simulation
6 K.A. Lawal SPE 128350

unrelated, models could satisfy (within allowable error The practice of using P10 and P90 realisations as the ‘low’
bounds) this criterion. Yet, these supposed history- and ‘high’ values of the input variables for performing
matched models yield different forecasts, hence stochastic simulations has negative implications on the
uncertainties in decision-making. ‘true’ distribution of the output. By limiting the input data
range to P10 - P90, there is no theoretical justification for
In addition, numerical models including history-matching, estimating the ranges P0 - P9 and P91 - P100 in the
are very sensitive to prevailing reservoir characteristics output distribution. Consequently, we do not agree with the
and assumptions. Consequently, with a switch from one practice of ‘deliberately’ constraining the uncertainty
recovery mechanism to another, e.g. from depletion to ranges of the independent variables, as recommended by
gas-cap drive, a previously history-matched model may no Amudo et al (20). Rather, such reduction in uncertainty
longer be valid for the new mechanism (31). Further, results range, even for mature systems, should be triggered by
of history-matching tend to depend on the algorithm used improved knowledge such as availability of new data, re-
(32, 33)
. These issues, which are some of the drawbacks of interpretation of existing data or perhaps, analogue.
automatic history-matching, make it imperative to be Moreover, we opine that combinations of the extreme
cautious at relying on conclusions of ED obtained from the values would effectively define the edges of the response
so-called ‘history-matched’ models. surface (1).
Besides dynamic issues, static models have some implicit
Some Pragmatic Applications of ED in E&P
uncertainties which might be difficult to approximate as
continuous functions. For example, in a very recent Despite the drawbacks of ED in modelling subsurface
comparative study of four reservoir modelling techniques uncertainties, scope exists for deployment in the following
(object based modelling, sequential indicator simulation, areas of E&P which, contrary to subsurface variables, are
multiple point statistics, and spectral component geologic characterised by independent variables that are
modelling) using high resolution ferron sandstone outcrop continuous and controllable for the achievement of desired
data, it was discovered that none of the algorithms could responses.
accurately reproduce the proportions of the reference
model facies (34). In fact, despite being conditioned to the Management of production and injection: As
same dataset, the resulting models differ significantly in examples, optimising separator performances relative to
volumetrics and flow characteristics. chemical dosage, operating pressure, valve setting, etc,
requires a good understanding of process dynamics.
Applying fundamental principles, which couple
Discussion
conservation equations and phase equilibrium, to develop
Inconsistency with reservoir physics is a major drawback and solve complex mathematical models may not be
of implementing ED for modelling subsurface practical. Besides, in the case of chemical dosage, the
uncertainties. However equally important, is paucity of chemical vendor is not likely to ‘reveal’ the true chemical
data. Although it appears straightforward, successful composition of the product, making it impossible to model
application of ED requires large amount of quality data, reaction kinetics. However, ED provides a cheap
especially for systems, such as reservoir processes, alternative, which though limited to a narrow operating
characterised by significant non-linearities. In contrast to envelop, is suitable for most engineering purposes.
the discrete nature of most subsurface variables, in strict Model-based control: automated systems such as
sense, ED and RSM presume continuous variables. ‘intelligent’ completions require models relating the
The issue of ‘heavy hitters’ is debatable. Our experience controllable variables (e.g. position of inflow control valves,
indicates that this depends strongly on the assumed ICV) with the output variables (e.g. fluid
development option. For instance, while the ratio of production/injection rate and pressure). By empirical
vertical-to-horizontal permeability (kv/kh) may be critical for identification, ED may provide cheaper alternative
a horizontal well development, it is not likely to be modelling (17) than more rigorous first-principle models,
important for vertical well development. Imagine applying a and where available, ED-based model may also
proxy model (without kv/kh) developed with vertical complement such first-principle models. However, where
completions in a heterogeneous system to predict/optimise digital control, entailing conversion of continuous analog
the performance of horizontal completions in the same signals to discrete digital signals, is implemented, it is
system! Obviously, there is need for specific proxy models instructive to ensure appropriate sampling frequency to
which may differ considerably between development plans, minimise loss of high-frequency information (35).
and possibly, between reservoir sectors or blocks.
Business planning: Using profitability indicators such as
The dynamism of ‘heavy hitters’ is also of concern. For net present value, payout time or internal rate of return as
example, the key uncertainties under a gascap drive are dependent variables, ED may be used to develop useful
likely to differ from those of water drive or combination response surfaces incorporating effects of uncertainties
drive. Therefore, in a scenario where the development such as production, price, schedule, cost and discount-
plan had been ‘optimised’ with respect to gascap drive, rate on project performance. Such proxies would also
how exposed are the investments if water drive becomes facilitate construction of probability distribution curves of
the outcome? Green fields and poorly understood mature the profitability indicators, which enhance decision-making
fields are particularly vulnerable to this risk. compared to deterministic tools such as spider diagram
and tornado chart.
7 Modelling Subsurface Uncertainties with Experiemental Design: Some Arguments of Non-Conformists SPE 128350

Conclusion 3. Lawal, A.K. and O. Eweje (2009). “On Field

Development Planning: A Comparative Study of
Main findings of this work include: Experimental Design and Structured Deterministic
Techniques”. Paper presented at Offshore West
• Data acquisition, processing, interpretation and
Africa Conference & Exhibition, Abuja, 27 - 29 Jan
utilisation are critical for robust modelling of
subsurface uncertainties 4. Peng, C.Y., R. Gupta, K. Vijayan, G.C. Smith, M.A.
Rayfield, and D.R. DePledge (2004). “Experimental
• Although statistically consistent, most ED-based proxy
Design Methodology for Quantifying UR Distribution
models do not conform to reservoir physics
Curve- Lessons Learnt and Still to be Learnt”. Paper
• Depending on the input distributions, ED may 88585 presented at SPE Asia Pacific Oil and Gas
underestimate or exaggerate the impacts of Conference & Exhibition, Perth, 18 - 20 Oct
subsurface uncertainties 5. Badru, O., and C.S. Kabir (2003). “Well Placement
• Reservoir mechanisms are too complex for reliable Optimisation in Field Development”. Paper 84191
modelling with ED-based proxy equations presented at SPE Annual Technical Conference &
• There is yet no convincing method of constructing a Exhibition, Denver, 5 - 8 Oct
unique simulation model from the distribution curve of 6. Queipo, N.V., J.V. Goicochea, and S. Pintos (2002).
the response variable. “Surrogate Modelling-based Optimisation of SAGD
Processes”. J. of Petroleum Science & Eng., 35, 83-
93
Recommendation
7. Aanonsen, S.I., A.L. Eide, L. Holden, and J.O. Aasen
Based on this study, the following are recommended: (1995). “Optimising Reservoir Performance under
Uncertainty with Application to Well Location”. Paper
• Analysis and mitigation of subsurface uncertainties
30710 presented at SPE Annual Technical
should be multidisciplinary
Conference & Exhibition, Dallas, 22 - 25 Oct
• Reduction of uncertainties should be driven by 8. Guyaguler, B. and R.N. Horne (2001). “Uncertainty
acquisition and (re)interpretation of quality data assessment of Well-Placement Optimisation”. Paper
• Known ‘minimum’ (P0) and ‘maximum’ (P100) data 71625 presented at SPE Annual Technical
should always be used to bound the input factors Conference & Exhibition, New Orleans, 30 Sep - 3
• Relatively greater efforts should be expended on Oct
reservoir geosciences and physics than the 9. Onwunalu, J., M. Litvak, L.J. Durlofsky, and K. Aziz
mathematics and statistics (2008). “Application of Statistical Proxies to Speed Up
Development Optimisation Procedures”. Paper
• Possibility of applying ED for production management,
117323 presented at International Petroleum
business planning and other applications should be
Exhibition & Conference, Abu Dhabi, 3 - 6 Nov
investigated.
10. Artus, V., L.J. Durlofsky, J. Onwunalu, and K. Aziz
Acknowledgements (2006). “Optimisation of Nonconventional Wells under
Uncertainty”. Comput. Geosci., 10, 389 – 404
The author is grateful to Jonathan Umurhohwo, for useful
discussions on Experimental Design. 11. Vincent, G., B. Corre, and P. Thore (1999). “Managing
Structural Uncertainty in a Mature Field for Optimal
Well Placement”. SPERE, 2, 377-384
Nomenclature
12. Litvak, M.L., B. Gane, and L. McMurray (2007). “Field
Boi initial oil formation volume factor, L3L-3 Development Optimisation in a Giant Oil Field in
GRV gross rock volume, L3 Azerbaijan and a Mature Oil Field in the North Sea”.
NTG net-to-gross ratio, LL-1 Paper 18526 presented at Offshore Technology
Soi initial oil saturation, L3L-3 Conference, Houston, 30 Apr - 3 May
STOIIP stock-tank oil initially in-place, L3 13. Rivera, N., N. Meza, J. Kim, P. Clark, R. Garber, A.
φ
3 -3
porosity, L L Fajardo, and V. Pena (2007). “Static and Dynamic
Uncertainty Management for Probabilistic Production
References Forecast in Chuchupa Field, Colombia”. SPERE, 10,
4, 433 - 439
1. O’Dell, M. and Lamers, E. (2003). “Subsurface
14. Koci, P.F. and J.G. Mohiddin (2007). “Realistic History
Uncertainty Management in the Harweel Cluster,
Matching of Cyclic Steam Stimulation Performance of
South Oman”. SPE 84189 presented at Annual
Several Groups of Multilateral Wells in the Peace
Technical Conference & Exhibition, Denver, Colorado,
River Field, Canada”. Paper 107201 presented at
5 - 8 Oct
SPE Europec/EAGE Annual Conference & Exhibition,
2. Passone, S. and McRae, G.J. (2007). “Probabilistic London, 11 - 14 Jun
Field Development in Presence of Uncertainty”. IPTC
11294 prepared for the International Petroleum 15. Montgomery, D.C. (1976). Design and Analysis of
Technology Conference, Dubai, 4 - 6 Dec Experiments, John Wiley & Sons Inc, New York
8 K.A. Lawal SPE 128350

16. Myers, R.H. and D.C. Montgomery (2002). Response the Value of Information in a Deepwater Reservoir”.
Surface Methodology, 2nd ed., John Wiley & Sons Inc, Paper 79707 presented at SPE Reservoir Simulation
New York Symposium, Houston,
3 - 5 Feb
17. Marlin, T.E. (1995). Process Control: Designing
Processes and Control Systems for Dynamic 27. Chu, C. (1990). “Prediction of Steamflood
Performance. McGraw-Hill Inc., New York Performance in Heavy Oil Reservoirs Using
18. Hougen, J. (1964). “Experiences and Experiments Correlations Developed by Factorial Design Method”.
with Process Dynamics”, Chem. Eng. Prog. Paper 20020 presented at 60th California Regional
Monograph Ser., 60, 4 Meeting, California, 4 - 6 Apr
28. Kabir, C.S., M. Agamini, and R.A. Holguin (2004).
19. Kabir, C.S., A. Chawathe, S.D. Jenkins, A.J. Olayomi,
“Production Strategy for Thin-Oil Columns in
C. Aigbe, and D.B. Faparusi (2002). “Developing New
Saturated Reservoirs”. Paper 89755 presented at
Fields Using Probabilistic Reservoir Forecasting”.
SPE Annual Technical Conference & Exhibition,
Paper 77564 presented at SPE Annual Technical
Houston, 26 - 29 Sep
Conference & Exhibition, San Antonio, 29 Sep - 2 Oct
29. Hubbard, D. (2007). How to Measure Anything:
20. Amudo, C., T. Graf, R. Dandekar, and J.M. Randle
Finding the Value of Intangibles in Business, John
(2009). “The Pains and Gains of Experimental Design
Wiley & Sons, 2007
and Response Surface Applications in Reservoir
Simulation Studies”. Paper 118709 presented at SPE 30. Adenuga, A. (2008). Private Communication, 28th
Reservoir Simulation Symposium, The Woodlands, 2 - February 2008
4 Feb
31. SPE (2009). Advert on “Propagation of Uncertainties”
21. Deutscf, C.V. and A.G. Journel (1998). GSLIB
32. Dong, Y. and S. Dean (2008). “Reservoir Simulation
Geostatistical Library and User’s Guide, 2nd ed.,
Model Updates via Automatic History Matching With
Oxford University Press
Integration of Seismic Impedance Change and
22. Masters, T. (193). Practical Neural Network Recipes Production Data”. Paper IPTC 12550 presented at the
in C++, Academic Press International Petroleum Technology Conference,
23. Li, B. and F. Friedmann (2005a). “Novel Multiple Kuala Lumpur, 3 - 5 Dec
Resolutions Design of Experiment/ Response Surface 33. Cheng, H., X. Wen, W.J. Milliken, A., and A. Datta-
Methodology for Uncertainty Analysis of Reservoir Gupta (2004). “Field Experiences with Assisted and
Simulation Forecasts”. Paper 92853 presented at Automatic History Matching Using Streamline
SPE Reservoir Simulation Symposium, Texas, 31 Jan Models”. Paper 89857 presented at SPE Annual
- 2 Feb Technical Conference & Exhibition, Houston, 26 - 29
24. Li, B. and F. Friedmann (2005b). “A Novel Response Sep
Surface Methodology Based on Amplitude Factor 34. Deveugle, P. (2009). “A Comparative Study of
Analysis for Modelling Nonlinear Responses Caused Reservoir Modelling Techniques Using High
by Both Reservoir and Controllable Factors”. Paper Resolution Ferron Sandstone Outcrop Data”,
95283 presented at SPE Annual Technical Postgraduate Seminar, Earth Science & Engineering
Conference & Exhibition, Texas, 9 - 12 Oct Dept., Imperial College London
25. Yeten, B., A. Castellini, B. Guyaguler, and W.H. Chen 35. Ogata, K. (1987). Discrete-Time Control Systems,
(2005). “A Comparison Study on Experimental Design Prentice-Hall, New Jersey
and Response Surface Methodologies”. Paper 93347
presented at SPE Reservoir Simulation Symposium,
Texas, 31 Jan - 2 Feb
26. Portella, R.C.M., M.C. Salomao, M. Blauth, and R.L.B.
Duarte (2003). “Uncertainty Quantification to Evaluate
9 Modelling Subsurface Uncertainties with Experiemental Design: Some Arguments of Non-Conformists SPE 128350

80

($/bb
70 ED S PE p a p e r s

60 O il Pr ic e

No, Pric e ($/bbl)

50

40

30

20

10

0
1971 - 1976 - 1981 - 1986 - 1991 - 1996 - 2001 - 2006 -
1975 1980 1985 1990 1995 2000 2005 2009
T im e In t e r v a l

Fig. 1: Historic Trend of ED Publications ∗ & Oil Price #

Gas Producer

Fig. 2: Simulation Model for Example 1B (3)

60,000 1.0E+08

9.0E+07
Table 1: Deterministic STOIIP Simulation (P0 - P100)
50,000
8.0E+07 High
Oil Rate (stb/d)(stb/d

7.0E+07 Low (P0) Base (P100)

40,000
Oil Cum (stb) (stb)

6.0E+07
GRV (MMft3) 2962 3833 4530
30,000 5.0E+07
Oil rate Porosity 0.15 0.22 0.32
4.0E+07
Oil recovery
20,000
3.0E+07 NTG 0.58 0.66 0.79

2.0E+07 Soi 0.52 0.72 0.88

10,000
1.0E+07
1/Boi (stb/rb) 0.72 0.78 0.89
0 0.0E+00
0 1000 2000 3000 4000 5000
Time (days) STOIIP (MMstb) 17.2 55.8 160.2

Fig. 3: Oil Rate and Recovery as Functions of Time

∗
Obtained with the SPE search engine, using the key words: Design of experiments, factorial design, response surface methodology, proxy model,
probabilistic forecasting. Valid as at Feb. 18, 2009
#
http://inflationdata.com/inflation/Inflation_Rate/Historical_Oil_Prices_Table.asp, accessed Feb 20, 2009
10 K.A. Lawal SPE 128350

Table 2: Deterministic STOIIP Simulation (P10 - P90)

Table 3: ED for Example 6

Low High GP Rf
(P10) Base (P90) Run S F C A (Bscf) (%)
1 0 0 0 0 279 72.8
GRV (MMft3) 3441 3833 4051
2 -1 1 -1 1 88 84.4
Porosity 0.18 0.22 0.29 3 -1 -1 -1 -1 68 62.5
4 1 -1 1 -1 414 61.2
NTG 0.62 0.66 0.71
5 1 1 1 1 605 83.1
Soi 0.66 0.72 0.79
6 1 1 -1 -1 191 76.2
1/Boi (stb/rb) 0.68 0.78 0.92 7 1 -1 -1 1 172 69.7
8 -1 1 1 -1 264 75.9
STOIIP (MMstb) 31.3 55.8 107.7
9 0 0 0 0 279 72.8
10 -1 -1 1 1 254 69.4

1.0

0.8 Table 4: Mapping Simulation Runs to Probability

(<x

GP Rf
0.6 P100 PX (Bscf) Run Ψ (%) Run
Probability (<x)

P10 144 7 66.5 10

0.4
P50 258 10 72.9 1, 9
P90 394 4 79.1 5

0.2 P0 Uniform
Triangular
LogNorm al
0.0 700
0 50 100 150 200 250 300 350 400 y = 1.0152x
STOIIP (M M s tb) 600 R2 = 0.9704
Fig. 4: End-points (P0, P100) of Deterministic & ED
UR
Actu al UR (Bscf)

500
Simulations
400

300
1.0
200

0.8
(<x

100
Probability (<x)

0
0.6
0 100 200 300 400 500 600 700
Predicted UR (Bscf)
0.4

Fig. 6: Cross-plot of Actual & Predicted UR

0.2 Uniform(P10-P90)

Unifom(P0-P100)
0.0
0 20 40 60 80 100 120 140 160
STOIIP (MMstb)

Fig. 5: (P10-P90) and (P0-P100) Distribution Curves

Ψ
The runs are the closest ones to the respective probabilities. There is
no interpolation between runs
11 Modelling Subsurface Uncertainties with Experiemental Design: Some Arguments of Non-Conformists SPE 128350

90
y = 1.0013x 1.0
(% R2 = 0.9799
85

Probability (<x) (<x)

0.8
A ctu al R F (% )

80

75 0.6
70
0.4
65

60 0.2
60 65 70 75 80 85 90
Pr e dicte d RF (%)
0.0
60 65 70 75 80 85 90
Fig. 7: Cross-plot of Actual & Predicted RF
Gas RF (%)

Fig. 9: Gas RF Distribution

1.0

0.8 30
Probability (<x) (<x)

0.6 25

0.4 20
Skin factor`

0.2 15

0.0 10
0 100 200 300 400 500 600
Gas UR(Bscf)
5
Fig. 8: Gas UR Distribution
0
7-Jun-2007

5-Oct-2007
6-Aug-2007

2-Apr-2008
8-Apr-2007
12-Jun-2006

7-Feb-2007
9-Dec-2006

4-Dec-2007

2-Feb-2008
15-Oct-2005

11-Aug-2006

10-Oct-2006
13-Apr-2006
12-Feb-2006
14-Dec-2005

Time

(30)
Fig. 10: Variation of Actual Skin Factor over Time