SPE-188313-MS

Reduced-Physics Modeling and Optimization of Mature Waterfloods

Fayadhoi Ibrahima, Agustin Maqui, Ana Suarez Negreira, Chao Liang, Feyisayo Olalotiti, Ouassim Khebzegga,
Sebastien Matringe, and Xiang Zhai, Quantum Reservoir Impact

Copyright 2017, Society of Petroleum Engineers

This paper was prepared for presentation at the Abu Dhabi International Petroleum Exhibition & Conference held in Abu Dhabi, UAE, 13-16 November 2017.

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 have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum 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 acknowledgment of SPE copyright.

Abstract
Mature waterfloods often present significant Reservoir Management challenges. After an initial boost in
oil production, water cuts tend to increase and flood performance starts to decline. Complex reservoirs that
have been producing for decades through hundreds or thousands of wells are notoriously challenging to
model. Creating and history-matching a simulation model usually take several months for subsurface teams,
and operational teams can rarely rely on these models to make reservoir management decisions. In this
paper, a novel methodology is presented that is being used in practice on large waterfloods or strong aquifer-
supported reservoirs, to support operational decisions in near real-time. The proposed technology relies on a
reduced-physics, data-driven reservoir model to quickly build and match a reservoir model that can be used
to optimize waterfloods. The first stage of the workflow involves collecting and validating the field data,
including rock and fluid properties, production, injection and pressure data as well as well information, such
as trajectories and historical perforations. The reservoir behavior is then modeled following an approach
similar to that of Thiele and Batycky (2006) in the context of streamline simulation. The model represents
the reservoir as a network of inter-well connections described by their strengths and efficiencies. Contrary
to traditional streamline-based method, the strength of connection is rather determined through the solution
of a numerical tracer test, which generalizes the method to unstructured or locally refined grids as well as
dual permeability systems, and allows the method to account for mild compressibility effects. An empirical
fractional flow model is then used to calculate the connection efficiencies. Once the model is complete and
calibrated, a cutting-edge optimization algorithm is used to optimize the production-injection strategy based
on this network of subsurface connections. Recommendations for adjustments in the production-injection
strategies are proposed and model uncertainties are computed through a novel algorithm to compute the
associated risks. A new finite-volume based time-of-flight computation algorithm is developed based on
the numerical tracer solution, which, combined with the empirical fractional flow model, can give a data-
driven production mapping algorithm.
The proposed methodology was successfully applied to many reservoirs across the world, including
several giant middle-east carbonates with hundreds of wells and decades of history. The approach
consistenly identified an optimized strategy that could deliver several percentage points of incremental oil
along with a reduction in water production. The methodology proposed is fast enough to build and match
2 SPE-188313-MS

a new model in a few days; and updating an existing model takes less than an hour as new data comes
available, avoiding expensive numerical simulations and helping engineers optimize daily production-
injection strategy of reservoirs.

Introduction
It has always been challenging to optimize secondary recovery. Indeed, it requires correctly understanding
and resolving the uncertainty of the extremely complex subsurface fluid network, and the capability of
quickly extracting such information from fast varying reservoir behaviors given changing waterflood
scenarios. Moreover, it involves being able to respond to those behaviors efficiently, to promptly
communicate between reservoir modeling and field operation, while still managing risks. Even though
waterflood has been utilized and studied for many years, most waterfloods are still operated under sub-
optimal conditions. Depending on different recovery scenarios and oil price, waterflood might be designed
for maximizing oil production, minimizing water production, or maximizing the estimated ultimate recovery
(EUR). The target may also change during the lifetime of the producing reservoir. This brings another layer
of difficulty in properly managing waterflood. Temporary boosting in production rate is usually preferred
in low oil price environment. However, this could lead to permanent damage to the reservoir because water
conning and bypassing are common under high productivity. In long term, this strategy usually results in
fast water breakthrough and low EUR. Minimizing water production rate is usually seen in fields that have
limited water processing capability. This strategy can help slow down water breakthrough and help improve
EUR.
Traditionally, full-physics numerical simulation has been implemented for understanding waterflooding,
forecasting reservoir behaviors and optimizing waterflood plan. However, full-physics simulation, in
addition to being usually computationally expensive, is sensitive to high uncertainty in input geology
and relative permeability model. The solution to this problem is to perform history match, in which
geological model is tuned to match the behaviors (e.g., pressure, production rate, water cut) of the reservoir.
Mathematically, the history matching process is a highly under-determined and overfitting problem because
the number of adjustable parameters greatly exceeds the amount of quantifiable reservoir behaviors.
For example, a mid-size simulation model contains 100,000-1,000,000 cells with millions of adjustable
parameters. In contrast, the number of observable reservoir behaviors is on the order hundreds or thousands.
The practice of using 1,000,000 parameters to fit 1,000 samples is mathematically invalid. In real-world
applications, it leads to different reservoir engineers giving different history-matched instances of the same
model and different reservoir behavior prediction. Moreover, simulation history match is time consuming,
experience-based and hence could be biased. It involves fine tuning numerous parameters and running
many tests. Ideallly, an efficient waterflood should be adjusted frequently to reflect reservoir changes. In
contrast, decision-making based on traditional simulation approaches could take years. One alternative to
full-physics is to use streamline simulation, which gives intuitive understanding of the producer-injector/
aquifer communication, and runs faster than finite-volume based simulations. However, streamline-based
simulation does not solve the high uncertainty issue in history matching.
The traditional history matching process can be considered as a manual data-driven approach. Automatic
data-driven approaches, such as capacitor-resistor models (CRM, Albertoni and Lake, 2003, Sayarpour,
2008), ultimately share the same concept as simulation history matching. The philosophy behind both is
that the observable reservoir behaviors are consequences of its internal properties and operation actions, and
both approaches try to uncover the cause (reservoir internal properties) from the consequences (reservoir
behaviors). While simulation tries to resolve the reservoir properties directly, CRM conceptualizes the
reservoir into a much simpler proxy model, and then calibrates the proxy model using automatic statistical
approaches. For example, in CRM models, the reservoir is partitioned into many control volumes, each
associated with one producer-injector pair and described by a simple model characterized by several
SPE-188313-MS 3

key parameters fitted using production/injection history. As a purely data-driven approach, CRM ignores
pre-known geological information. Therefore, significant field history is necessary. Some of the CRM
assumptions such as constant control volume over time are not realistic in most real waterfloods. Therefore,
CRM is suitable for mature, small/mid-sized waterfloods that have not been under major changes in well
controls.
To address the problems of traditional simulation and data-driven technologies, we develop a novel
methodology that can efficiently combine the advantages of both approaches. The core waterflood
surveillance algorithm contains two major steps. First, we use an improved numerical tracer algorithm to
compute the hydraulic communication between producers and injectors/aquifers based on a given geological
model. This step is deterministic similar to numerical simulation. Second, we adopt a data-driven approach
to calibrate the waterflood efficiency of each significant connection. Afterwards, a waterflood reservoir
is conceptualized into a connection-based network proxy model with the model parameters given by
deterministic and statistical approaches. Fast forecast and optimization are enabled by the simpler proxy
model. Some detailed discussions can be found in Wen et al. (2016), Zhai et al. (2016a), Zhai et al. (2016b)
and Maqui et al. (2017). This paper provides an in-depth review of the entire technology, including some
newer development such as aquifer subdivision algorithm, uncertainty analysis, improved time-of-flight
estimation and data-driven production mapping.
The paper is organized as follows. The first section introduces the inter-well connectivity core
algorithms, based on well allocation factors and connection efficiencies, including how waterflood operation
optimizations can be conducted by solving linearized problems and how we can consider multiple aquifer
strengths. Then we present an algorithm for quantifying uncertainties in the well connectivities, inherent
to limited data collection. Finally, before discussing a field application, the last technical section describes
how we leverage the tracer concept to obtain fast production mapping.

Fast reservoir diagnostic via inter-well connectivity

A compact and convenient way to describe the allocation dynamics of a waterflood reservoir is to use
a network proxy model embedding the well allocation factors (strengths) and efficiencies (oil fractional
flows) of each directed pair of injector-producer in the reservoir. This reduced-physics model has gained
momentum with the revival of streamline simulation (Thiele and Batycky, 2006, Pollock, 1988). Well
allocation factors (WAF) and efficiencies are natural metrics along streamlines and can be quickly estimated.
However, the accuracy of streamline-based methods can be deeply affected by complex (yet common)
physics such as dual porosity systems, or heterogeneous numerical geometries such as unstructured grids.
Fortunately, it is possible to estimate the network WAFs and efficiencies by directly working on Eulerian
grids, bypassing direct streamline simulations while preserving computational speed and flexibility in the
reservoir model.

Pressure and tracer formulation

The well allocation factor (WAF) from injector i to producer j, WAFij, quantifies the pressure communication
between the well pair (i,j) and is defined by

(1)

where qij is the total instantaneous liquid rate that flows from injector i to producer j, and qi is the total liquid
injection rate of injector i, i.e. we have the relation , where J (resp. I) is the set of producer
(resp. injector) indices. Note that 0 ≤ WAFij≤1 and , reflecting the conservation of mass in
the network. However, in general, WAFji ≠ WAFij, as WAFji quantifies the strength of connection between
producer j towards the specific injector i (among all injectors). Still, the following relation stands,
4 SPE-188313-MS

(2)
As in streamline simulation, the first step in estimating WAFs is to recover the (single-phase) Darcy
velocity field under fluid viscosity μ and permeability tensor K. If q is the source term, the
pressure field p can be computed by numerically solving the steady-state pressure equation (Aziz and Settari,
1979),

(3)

Then, instead of tracing streamlines, we follow Shahvali et al. (2012) and alternatively rely on steady-
state tracer concentrations, defined by the following equation,
(4)
where c is the tracer concentration of a fictitious tracer exclusively injected, for instance, at an injector i
(i.e., c=1 at injector i and c = 0 at each injector i’ ≠ i) and finding its way to the different producers due to
the velocity field v. Consequently, the tracer concentration at any point in the reservoir is the average tracer
concentration of all streamlines passing through this point, establishing the relationship between streamlines
and tracer concentrations. Similarly, to conduct tracer tests from producers, we can reverse the direction of
the velocity field and solve the same tracer equation.
Numerically, Shahvali et al. (2012) computed the tracer concentrations from the following modified
steady-state tracer equation,
(5)
ideal for finite-volume discretization with, for instance, two-point flux approximation, leading to the
numerical equation,
(6)
where D is the numerical divergence operator, Q = diag(q) is the diagonal matrix of flux vector q defined
on each face of the grid and A is the numerical upwind operator interpolating cell values to face values.
However, the latter formulation requires strict incompressibilty, ∇ · v = 0, which can be violated either for
physical (close to wells, faults or reservoir boundary) or numerical (rounding errors) reasons. To mitigate
this issue, we have proposed (Zhai et al., 2016) to solve instead the following steady-state tracer equation,
equivalent to (4),
(7)
The corresponding numerical linear equation to solve is
(8)
leading to numerical stability and tolerance to mild compressibility. This improved numerical tracer
algorithm has been proved accurate and robust using the pressure solution of equation (3)) with both direct
linear and iterative solvers such as adaptive algebraic multigrid (Bell et al. 2015). We have also tested
pressure solution that includes slight compressibility, i.e.,

(9)
SPE-188313-MS 5

Finally, the WAFs are attributed as follows. We denote by the numerical tracer concentration at cell
k when performing a tracer test from injector i, Gj the set of cells belonging to producer j and Gi the set of
cells belonging to injector i. By mass conservation, for any cell k, we have the following equation,

(10)

The tracer concentration from injector i assigned at producer j, cij, is then defined as the average
concentration of all the cells that belong to producer j, i.e.,

(11)

cji, the tracer concentration at injector i when performing a tracer test from producer j, is defined in a similar
fashion. cij (resp. cji) is the fraction of flux collected by producer j (resp. injector i) from tracer injection at
injector i (resp. producer j), matching the definition of WAF. In other words, we verify that

concluding the streamline-free, flexible and fast WAF computation algorithm.

Figure 1 gathers examples of (stationary) field pressure distribution and tracer concentrations on three-
dimensional grids for different reservoirs, using the methodology.

Figure 1—Example of stationary field pressure distribution (left figure) on a 3D gridded reservoir.
The black lines are well trajectories. Warm colors correspond to high pressure. Example of tracer
concentration (right figure) from a producer well (in green), essentially supported by the aquifer.

Data-driven fractional flow calibration

To complete the network proxy model, we need to be able to estimate the well connection efficiency, EFFij,
defined as follow for a pair (i,j) of injector i and producer j,

(12)

where (resp. ) is the instantaneous oil (resp. water) rate through the streamline connecting injector
i to producer j. In other words, EFFij is the instantaneous oil cut of the connection (i,j) and informs the
reservoir engineer about the performance of the waterflood from injector i with respect to helping producer
j recover oil rather than water. Because the notion of efficiency only pertains to well connection, EFFji, the
efficiency from producer j to injector i, is defined by
(13)
6 SPE-188313-MS

With no access to streamlines for efficiency estimation, we proceed in a novel fashion. The idea is to
parameterize the efficiency with a decline function with few hyper-parameters and calibrate them with
production history-matching. Among other possible decline functions, we choose here a sigmoid-like
function, g(x;a,d) = (1 + e(x-d)/a)-1), where a > 0 and d > 0 are the hyper-parameters, and write

(14)

where t is time and CWIij(t) is the cumulative water injection at time t for the connection (i,j), which is
defined by

(15)

dij corresponds to the necessary cumulative water injection in order to reach 50% efficiency, while aij dictates
the sharpness of the efficiency decline: 4aij is the additional cumulative water injection required to go from
90% (EFFij = (1 + e-2)-1) efficiency to 10% efficiency (EFFij = (1 + e2)-1), assuming that CWIij = dij - 2aij
and CWIij = dij +2 aij are achievable for this connection.
By definitions of the well allocation factor and the efficiency, the oil production rate at producer j, ,
can be expressed as follows,

(16)

Consequently, we can calibrate the efficiency model, connection by connection, by minimizing the cost
function corresponding to the mismatch between the actual historical production
and our parameterized model. In other words, we seek to find the hyper-parameters for each producer
j, that minimizes the following cost function (total squared misfit),

(17)

The advantages of this approach are twofold: the model is simple because only two (time-independent)
hyper-parameters need to be fitted per connection (thus allowing to mitigate potential overfitting), and
the history-match can be performed with minimal computational effort because the cost function (17) is
analytical, despite the optimization problem being non-linear (Nocedal and Wright, 2006).
Figure 2 illustrates the reservoir partitioning based on control volumes obtained by solving tracer
concentration equations, the corresponding network proxy model, as well as a detailed inter-well
connections showing WAFs and efficiencies, and oil/water production history-matching performance at the
well level.
SPE-188313-MS 7

Figure 2—Reservoir partitioning based on tracer control volumes (upper left), and the corresponding inter-well network model
(upper right; green/blue for producer/injector). Inter-well connections from a producer (lower left), with WAFs represented
by the pie size and efficiencies by pie chart fractioning. History-matched results (lower right) for oil (green) and water (blue)
productions. Solid lines correspond to actual production history, while dotted lines are the results of history-matching.

Forecast and Optimization of Inter-well Connections

Another advantage of the network proxy method is that the calibrated efficiency may be used to predict
future reservoir behavior. If one defines the oil production and injection rates for the next time step, the
pressure and tracer equations can be solved to recalculate the WAFs. With the WAFs, the connection
efficiency can be extrapolated following the historical trend and a reservoir forecast is obtained. In the case
of optimization, the future production and injection rates are the variables to be determined as to improve
the waterflood. This can be accomplished by weakening strong but inefficient connections (high water cut)
and strengthening efficient ones. To obtain physically reasonable results a set of user-defined constraints,
such as feasible well production and injection rates, is applied to the optimization problem. The optimizer
then runs numerous iterations with adjustments to the well rates. For each iteration, a reservoir behavior
forecast is produced. With this strategy, the nonlinear optimization problem is reduced to a sequence of
linear optimization processes which make it computationally inexpensive and extremely fast. For instance,
the water production from producer j at iteration , can be linearized as follows,

(18)

where is the producer oil rate at iteration k, while and are respectively the WAF and
cumulative water injection at iteration k - 1. can itself be linearized and expressed as,

(19)
where Δt is the optimization time step from iteration k - 1 to k. If the latter time step is small compared to the
field production history (e.g. for mature fields) and no major operational or geological changes are observed,
we can even assume that the cumulative water injection remains unchanged, i.e. .
8 SPE-188313-MS

To avoid unphysical results or harmful rates, constraints for the optimization variables must be properly
set. For example, if production rates are not constrained, the optimizer might suggest rates above the conning
limit of the reservoir. If no constraints are placed on the injection rates, the optimizer might provide a
rate that cannot be supplied by the surface facilities, or could damage the geological formation. Economic
constraints can also be implemented to restrict any change of rates or number of allowed operations to fall
within the budget. The available constraints include, but are not limited to: water production and injection
rates, oil production targets, voidage replacement ratios, economic limits, and number of allowed operations.
Grouping may also be incorporated to apply specific constraints to certain group of wells, or just on a well-
by-well basis. Further details on the optimization algorithm, including mathematical description with the
objective function and constraint implementation may be found under Wen et al. (2016) and Zhai et al.
(2016b).

Aquifer Subdivision Algorithm

In order to realistically model dynamics in the reservoir, the aquifer, as a pressure or flux boundary, should
as well be included. For example, when an aquifer is imposed as a constant pressure boundary condition,
it might behave as injector by providing support to producer, or behave as producer by accepting injected
water, or both at the same time. Similar to injector-producer connection, injector-aquifer connection and
producer-aquifer connection can also be characterized by WAF and EFF. However, unlike well, aquifer
usually occupies much larger areas, and the local behavior of the aquifer is different from place to place.
In our simplified reservoir proxy model, treating the aquifer as a single well could lead to over-simplified
results. Thus, we developed an aquifer subdivision algorithm that divides an aquifer into sub-regions, using
k-means clustering (see e.g., Forgy, 1965), to quantify the contribution of each sub-region to the field
production. Each sub-region i will therefore be treated as if it was an injector or producer and is represented
for each role as a column vector in the two matrices of WAF = (WAFij) and oil efficiencies EFF = (EFFij)
(it only plays one role at a time). An example of aquifer subdivision is shown in the left panel of Figure
3. Adding this subdivision capability gives us the flexibility to use connection-based reservoir models that
only rely on the aquifer support.

Figure 3—Cluster-based aquifer subdivision algorithm (left) and full-field application (right).

The aquifer sub-division algorithm has been applied to a giant reservoir in Latin America. This reservoir
has a high-water production due to its close proximity to a strong aquifer and to a very favorable geological
context. All the production in this field is supported by its aquifer. We subdivided the bottom aquifer into 100
sub-regions. Figure 3 illustrates the results of the algorithm (right panel). Each light-blue bubble represents
an aquifer sub-region with the size proportional to its net contribution to liquid rate production. From
SPE-188313-MS 9

this figure, we can easily identify regions that contribute the most to water production and visualize their
connections, and ultimately design a strategy to reduce the field water production.

Probabilistic Model Calibration and Uncertainty Quantification

Subsurface model calibration problems are known to be ill-posed in general (Oliver and Reynolds, 2008).
It is customary to consider a probabilistic approach to the problem by which all plausible models, all
conditioned to the available data, are generated together with their respective probabilities. The idea is
to estimate the distribution of the parameters aij) and dij), by designing a workflow in which all possible
calibrated model realizations can be evaluated. Bayesian statistics have proven to be suited for probabilistic
model calibration, where observed information is utilized to update a prior belief of the system. In situations
where a prior model is completely uncertain, estimating the likelihood distribution function provides
sufficient insight into parameter uncertainty quantification. For our application, the likelihood function takes
the form of a multivariate Gaussian distribution, given as:

(20)

where d represents observed data (well oil production rate) and g(X) is the model response based on the
parameter vector X, a concatenated vector of aij and dij for all (i, j) connections. Finally, the data error
(or covariance) matrix is denoted by CD. In most cases, the probability distribution function (PDF) f(X) is
complex and difficult to express in closed form. For this reason, probability sampling algorithms, such as the
Markov Chain Monte Carlo (MCMC) algorithm were proposed to efficiently estimate these complex PDFs.
At each iteration step, we compute the acceptance probability of a proposal X based on an instrumental
distribution, q(X,Xk)) with a Markov chain at a current state Xk), using the Metropolis-Hastings criterion:

(21)

Traditional random walk Metropolis has poor performance for multi-model posterior distribution
functions, as in our applications. A multi-objective MCMC variant can help circumvent this problem
(Olalotiti-Lawal and Datta-Gupta, 2015) by using the Differential Evolution MCMC (DEMC) algorithm
(Ter Braak, 2006) for sampling the multivariate probability distribution in parallel. One important strength
of DEMC is the elimination of the difficult task of proposal generation. Proposals are generated using a
differential evolution procedure, a global optimization heuristic algorithm. At particular state t of a Markov
chain i, a proposal to state t + 1 is generated using the current state of chain i, added to a scaled difference
between the current state of two different chains R1 and R2 (sampled from a set of chains excluding i)
without replacement:
(22)
where is a scale factor depending on the problem dimensionality, d (number of parameters).
A random ‘jitter’ e~U(-b,b) is added to improve randomness of sampling, where b is a very small number,
compared to parameter magnitudes. The algorithm proceeds as follows:
1. For each producer, determine the number of Markov chains, N. A rule of thumb is to use 3 × number
of parameters. Note that the number of parameters is twice the total number of connections to that
producer nc.
2. Initialize each chain equally with the results obtained from the deterministic calibration. The
initialization may also be done with arbitrary values
3. At each iteration step, make parameter proposals using equation (22)) for all chains
10 SPE-188313-MS

4. Evaluate each proposal using equation (20)) and compute the acceptance probabilities using equation
(21)) for each proposal; across all chains.
5. Accept a proposal if α(Xk,X) > ρ~U(0,1). Reject otherwise.
6. Repeat steps 2 to 5 multiple times. Generally, a couple of thousand iterations is usually sufficient to
attain convergence.
Note that for a producer with 4 connections, 24 Markov chains are run in parallel, resulting in 24,000
function evaluations for every 1,000 global iterations. It is therefore safe to assume convergence of the
chains, based on the typical 20% acceptance rate of the DEMC algorithm.
The algorithm is applied to a waterflood reservoir with 300 wells, and results are shown for one well.
Figure 4 (left panel) shows the oil production rate comparing the observed data with the deterministic
match and a set of 200 realizations sampled from the likelihood function using the DEMC algorithm,
assuming 80% data error. The total run time was 12 minutes with 2,000 iterations and burn-in time set
to 1,000 iterations. Local bias between model production reponse and data is reduced in the probabilistic
case. A typical relationship between parameter estimates from deterministic (shown in vertical red line) and
probabilistic calibrations for a single connection is shown in the right two panels. It is important to note
that combination of the parameters from different modes of the distribution can result in marked production
performance during production forecast.

Figure 4—Left: Oil production rate from a producer, comparing observed data with deterministic
and probabilistic calibrated model responses. Right: Parameter distribution with probabilistic
calibrated model responses. The red lines correspond to the parameter deterministic values

Tracer-weighted Time of Flight

Time of flight (TOF), the time required for a neutral particle to travel to/from a location from an injector
(forward TOF) or to a producer (backward TOF), is widely used in flow diagnostics for predicting water
break-through time, evaluating sweep efficiency, and identifying target of new well placement (Thiele and
Batycky, 2003; Batycky et al., 2008; Shook and Mitchell, 2009; Shahvali et al, 2012). It is conventionally
computed by integration along individual streamlines (Datta-Gupta and King, 2007). This approach is
undesirable to handle unstructured grid or dual permeability, heterogeneous reservoir. Instead, Shahvali et
al. (2012) computes TOF for single-phase incompressible flow by solving the following equation using
finite-volume method,
(23)
SPE-188313-MS 11

(24)
where ϕ, τ and Vϕ are porosity, TOF and pore volume respectively. Equation (23)) is strictly valid in
continuous space because streamlines do not intersect; however, using the direct discretized finite-volume
equation (24)) is problematic because multiple streamlines can share the same grid cell. As the result, the
calculated TOF using (24) is an average TOF of all steamlines that pass through a grid cell, which was
already pointed out by Shahvali et al. (2012). This averaging effect can lead to very inaccurate results near
wells because wells could receive many streamlines from different injectors. It is impossible to isolate the
TOF from a particular injector using only equation (24)).
Krogstad et al. (2016)proposed a method to calculate the accurate TOF free from any averaging
of streamlines. The method requires solving transient tracer equation and can be very computationally
intensive. We propose a tracer-weighted TOF algorithm improved from Shahvali et al. (2012). The idea is
to use tracer concentration to scale both the flux and pore volume to the true fraction that belongs to each
specific injector-producer pair, and thus removes the averaging effect from different injection-production
pairs. This method does not prevent the averaging effect from the same well pair. However, this effect is
minor compared to averaging different well pairs. We revise equation (23)) into the following conservation
form,
(25)
Equation (23)) and (25) are equivalent in continuous space as the tracer factor c cancels out. However, the
discretization of equation (25)) gives tracer-weighted equation for solving TOF using finite volume method,
(26)
Forward and backward TOFs are obtained by using tracer from injectors and producers respectively.
Tracer c plays different roles at the two sides of the equation and is no longer cancellable: at LHS, c scales
the flux q and Q, while on the RHS it scales pore volume Vϕ.

Figure 5—Forward TOF using equation (24) (left, injectors (1-4) in purple and producer (5) in red). The reservoir is surrounded
by a constant-pressure side aquifer at all edges. The gray box is zoomed in as shown in the mid (solved by equation
(24)) and right panel (solved by (equation (26)). Only cells with tracer value c1c5>0.01% is shown in the right two panels.

We illustrate the improved TOF algorithm using a 2D synthetic problem with heterogeneous well
placement. The reservoir is 1 meter thick with a porosity of 0.25, meshed with unstructured PEBI grid
shown in Figure 5. The injection rate of each injector is equal to 1/4 TPV (total pore volume) over 100 years
and the production rate of the producer is 1.5 TPV over 100 years. Left panel of Figure 5 shows the forward
TOF calculated using equation (24). Only one single TOF value averaged from all injectors and aquifer is
obtained for each cell. In fact, the ‘pollution’ from the TOFs of other injectors is so strong that the TOF
12 SPE-188313-MS

from injector 1 to the producer (mid panel) is as large as 10 folds of the true value calculated by equation
(26) (right panel). The averaging effect is effectively removed by using equation (26).

Production Mapping
One of the most critical and challenging tasks regarding waterflood reservoirs is to evaluate where the
remaining oil is. Traditionally, finite-volume based or streamline-based numerical simulations are utilized
to compute the displacement of oil by water. Small time-increments are necessary in the simulations in
order to prevent numerical instability. The uncertainty in geology, relative permeability model and process
of history matching also lower the credibility of simulated water saturation evolution. In this section, we
propose a hydrid approach to map the saturation distribution based on the strength and efficiency of the
network proxy model and the TOF solution.
The oil efficiency EFFij of an injector-producer connection (i, j) declines with the increase of cumulative
water injection CWIij. Although this is a property evaluated at the location of wells, it is the consequence
of oil being replaced by water graduately in the entire drainage-swept volume associated with the injector
and the producer. If we assume the empirical oil efficiency model of a well-established connection does
not change significantly through time, we can map the oil efficiency of a particular connection as fractional
flow back to the entire drainage volumes in the reservoir where the fluid that contributes to the oil efficiency
is to be produced after additional volume of water is injected. The order of pore volumes being displaced
by water from injector and being produced at the producer is determined by forward and backward TOF.
Pore volume with shorter forward TOF is displaced sooner by water from injector, while pore volume with
shorter backward TOF is produced sooner at the producer. Relative permeability relations are then used to
convert oil fractional flow to saturation.
After tracer concentration is solved numerically, the reservoir is partitioned into many control volumes.
In each control volume, there is one injector (or aquifer) i at one end and one producer j at the other end. We
then solve for time-of-flight using the tracer-weighted TOF algorithm for each grid cell k in the control
volume. The assumption is that cells with the same TOF value have similar water saturation, because the
liquid in those cells have traveled the same time from injectors. It is equivalent to assuming that a single
‘super-streamline’ connects the injector and producer, and the problem is reduced to 1D. We then sort all the
cells that belong to this control volume by their TOF value, and then start to map EFFij (CWIij,aij, dij) to all
the cells. For any cell k in between the injector and the producer, we assume that its equivalent cumulative
water injection is

(27)

where is the fraction of pore volume in grid cell k associated with connection (i, j); is the
pore volume of cell k that can conduct mobile liquid; CWIij,0) is the cumulative water injection associated
with connection (i, j); αij is a scalling factor to be determined; and the summation * is performed for all
the cells l that have (i.e., downstream compared to cell k). When k is at producer, i.e., end of the
‘super-streamline’, there is no cell at its downstream, and . In absolute incompressible
condition, αij = 1. We then apply the efficiency function EFFij to get the oil fractional flow at all cells
k belonging to this control volume. The oil saturation and water saturation can then be obtained by
using a relative permeability model. The final oil saturation is a weighted average over all ij combinations,
i.e., .
SPE-188313-MS 13

In the actual production mapping process, we tune parameters aij, dij and αij so that the total volume of
water injected less produced matches the total water volume in the ij control volume additional to the initial
water saturation, under no significant change of EFFij, i.e.,

(28)

subject to , where CWPij is the cumulative water production of the

injector-producer pair, and and are the parameters initially calibrated in the history match process.
Figure 6 shows an example of water saturation mapping using the approach described above. The
reservoir studied has very large horizontal permeability, and moderate-to-small vertical permeability. The
water saturation results show that most water invaded into the reservoir from some high permeability
channels from the side, and some water originated from the bottom aquifer (conning).

Figure 6—Water saturation map of the top zone of a single dome, purely bottom aquifer-driven
reservoir over 20 year production history. The water saturation map is created using the data-
driven production mapping algorithm discussed in this manuscript. The red dot shows the
location of bottom holes of producers with the size proportional to cumulative production volume.

Discussion and Field Applications

We have presented a comprehensive description of a technology for waterflood management that combines
reduced-physics and data-driven techniques. The technology only requires a basic geological model and
production/injection history of a reservoir under waterflood, and can then automatically perform tracer
and time-of-flight computations, history matching, waterflood forecast and optimization, and production
mapping. The technology does not need long-term human-biased simulation and history matching, and
avoids over-simplification of the reservoir encountered in other data-driven waterflood modeling. The
pressure and tracer computation is based on finite-volume method, thus fully compatible with unstructured
grid, finite compressibility and dual permeability system. The time frame of building such model from
raw data is typically 1-3 days. A comprehensive waterflood analysis, including building the network proxy
model with uncertainty quantification, creating data-driven production mapping, performing waterflood
prediction and strategy optimization, can usually be obtained within one week. Updating an existing model
with new production/injection data only takes 1-3 hours. This reduces the analysis timeframe from years to
days, and enables quick adjustments for waterflood enhanced performance.
The technology has been applied to more than 20 waterflood fields worldwide, with total oil production
exceeding 700,000 STB/D, including the Middle East, Asia, North and Latin America, with waterflood
history ranging from 5 to more than 80 years, and well counts from 50 to 2,000. One of the real field example
is a 20 km by 5 km giant carbonate reservoir under direct line drive with 200+ horizontal producers and
150+ horizontal injectors. The reservoir has been under waterflood for about six years, and is still under
active development. All injectors were initially producers, and were converted into injectors after 6-12
month production. At the time this study was performed, the water cut of the field was already about 30%.
Purely data-driven approaches like CRM failed in modeling the waterflood because of lack of sufficient
14 SPE-188313-MS

history and fast varying waterflood strategies. CRM assumption of time-independent drainage volume was
not valid in this field. Instead, we first built an automated 2.5D PEBI (perpendicular bisector) unstructured
grid to discretize the reservoir into 500,000 cells, and then applied the proposed technology to compute the
tracer solution for each producer and injector. The tracer concentration solution gave the well allocation
factor between each injector-producer pair. An automatic calibration to empirical fractional flow models
was done within one minute, and gave history matching descrepancy results of less than 10% in oil/water
production rate on well level, and less than 2% on field level. We then performed a waterflood optimization
with the target of minimizing total water production while maintaining total oil production target. The entire
optimization problem was converted into many small iteration steps, each of them being a simple linear
optimization on the network proxy model. To test the validality of the forecast and optimization results,
two identical numerical simulations were performed 6 months before the study. The first simulation had
unchanged water injection and liquid production rates (simulation A, do-nothing scenario). The second
simulation had the same total injection rate but adjusted injection rates for each individual injector and
updated liquid production rates using the monthly network-based optimization results based (simulation
B, guided scenario). We ran both simulations for 6 months without any new injector or producer. Figure 7
shows the normalized oil and water production rates of both scenarios. It is shown that simulation B indeed
has significantly lower water production rate than simulation A after 6 months, without sacrificing total
production rate. After 6 months, the do-nothing scenario witnessed an increasing watercut from 16% to
25%, while simulation B first had 8% decrease in water cut, followed by a slow increase to 15%. The sudden
decrease in water cut during the first month of optimization is because the optimization algorithm found
an improved waterflood strategy; after the first month, the field was already run at optimal/near-optimal
scenario, the water cut slowly increased due to inevitable water breakthrough.

Figure 7—Comparison of total oil/water production rate of simulation A (do-nothing

scenario, dotted curves) and simulation B (guided by optimization based on network
proxy model, solid curves). Both simulations have the same injection rate (cyan curve).
SPE-188313-MS 15

Being able to lower water production rate while maintaining/increasing oil production rate does not only
make the field operation more profitable, it also can increase the ultimate recovery rate in the long run,
with more efficient water injection and slowed-down water breakthrough. In the above case study, such
a target was shown achievable by simply adjusting the injection and liquid production rates. This type of
field production optimization at minimal capital investment is extremely valuable for field management,
especially under the current low oil price environment.

References
Albertoni, A., Lake, L. (2003). Inferring Interwell Connectivity Only From Well-Rate Fluctuations in Waterfloods.
SPEREE, 6(1), 6–16.
Batycky, R. P., Thiele, M. R., Baker, R. O., and Chugh, S. H. (2008). Revisiting Reservoir Flood-Surveillance Methods
Using Streamlines. SPE Res Eval & Eng, 11(2), 387–394. SPE-95402-PA. http://dx.doi.org/10.2118/95402-PA.
Bell, W. N. and Olson, L. N. and Schroder, J. B. (2015). Algebraic Multigrid Solvers in Python v3.0. https://github.com/
pyamg/pyamg
Datta-Gupta, A. and King, M. J. (2007). Streamline Simulation: Theory and Practice, No. 11. Richardson, Texas: Textbook
Series, SPE.
Maqui, A. F., Zhai, X., Suarez Negreira, A., Matringe, S. F. and Lozada, M. A. (2017). A Comprehensive Workflow
for Near Real Time Waterflood Management and Production Optimization Using Reduced-Physics and Data-Driven
Technologies, SPE-185614-MS, SPE LACPE Conference, Buenos Aires, Argentina.
Nocedal, J. and Wright, S. (2006). Numerical Optimization, Springer Science and Business Media.
Forgy, E. W. (1965). Cluster analysis of multivariate data: efficiency versus interpretability of classifications. Biometrics,
21, 768–769.
Krogstad, S., Lie, K. A., Nilsen, H. M., Berg, C. F. and Kippe, V. (2016). Flow Diagnostics for Optimal Polymer Injection
Strategies. ECMOR XV – 15th European Conference on the Mathematics of Oil Recovery, EAGE, Amsterdam, the
Netherlands.
Oliver, D. S., Reynolds, A. C. and Liu, N. (2008). Inverse Theory for Petroleum Reservoir Characterization and History
Matching. Cambridge University Press, Cambridge, UK.
Olalotiti-Lawal, F., & Datta-Gupta, A. (2015). A Multi-Objective Markov Chain Monte Carlo Approach for History
Matching and Uncertainty Quantification. Society of Petroleum Engineers. http://dx.doi.org/10.2118/175144-MS.
Pollock, D. (1988). Semi-analytical computation of path lines for finite-diference models. Ground Water, 26, 743–750.
Sayarpour, M. 2008. Development and Application of Capacitance-Resistive Models to Water/CO2 Floods. PhD
Dissertation, University of Texas at Austin.
Shahvali, M., Mallison, B. T., Wei, K. and Gross, H. (2011). An alternative to streamlines for flow diagnostics on structured
and unstructed grids. SPE Journal, 17(03), 768–778. http://dx.doi.org/10.2118/146446-MS.
Shook, G. M. and Mitchell, K. M. (2009). A Robust Measure of Heterogene- ity for Ranking Earth Models: The F PHI
Curve and Dynamic Lorenz Coefficient. Paper Paper SPE 124625 presented at the SPE Annual Technical Conference
and Exhibition, New Orleans, 4-7 October. http://dx.doi.org/10.2118/124625-MS.
Ter Braak, Cajo J. F. (2006). A Markov Chain Monte Carlo version of the genetic algorithm Differential Evolution: easy
Bayesian computing for real parameter spaces. Statistics and Computing 16(3), 239–249.
Thiele, M. R. and Batycky, R. P. (2003). Water Injection Optimization Using a Streamline-Based Workflow.
SPE 84080, presented at the SPE Annual Technical Conference and Exhibition, Denver, 5-8 October. http://
dx.doi.org/10.2118/84080-MS.
Thiele, M. R., and Batycky, R. P. (2006). Using streamline-derived injection efficiencies for improved waterflood
management. SPE Reservoir Evaluation & Engineering, 9(02), 187–196.
Zhai, X., Wen, T. and Matringe, S. F. (2016a). Estimation of Inter-Well Connections in Waterflood under Uncertainty
for Application to Continuous Waterflood Optimization of Large Middle-Eastern Carbonate Reservoirs. SPE-181685-
MS, presented at the 2016 SPE Annual Technical Conference and Exhibition, Dubai, UAE.
Zhai, X., Wen, T. and Matringe, S. (2016b). Production optimization in waterfloods with a new approach of inter-well
connectivity modeling. SPE-182450-MS, presented at the SPE Asia Pacific Oil & Gas Conference and Exhibition,
Perth, Australia.