SPE-212231-MS

Multiscale Reservoir Simulation of High-Resolution Models

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

Jostein Natvig, Daniel Dias, Frode Bratvedt, Shingo Watanabe, Zhuoyi Li, and Antonina Kozlova, SLB; Pavel
Tomin, Jiamin Jiang, and Xundan Shi, Chevron Technical Center

Copyright 2023, Society of Petroleum Engineers DOI 10.2118/212231-MS

This paper was prepared for presentation at the SPE Reservoir Simulation Conference held in Galveston, Texas, USA, 28–30 March 2023.

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
To quantify the uncertainty in reservoir performance, it is common to build ensembles of models that sample
the space of possible reservoirs that are consistent with the available data. To evaluate the spread of possible
outcomes, simulations experiments are run for each model in the ensemble to calculate for instance recovery
factor. The geoscreening workflow is a common way to do this systematically and in a reasonable time. It
can work as follows: First, run simulations with simplified physics to calculate recovery factor for every
model in the ensemble. Then, use recovery factor (and other quantities) to rank and select representative
models for high, medium, and low performance scenarios that can be used for full field simulations.
In this paper we present an application of the multiscale sequential fully implicit (MS SFI) framework
to simulate extremely complex high-resolution models with simplified physics. This enables us to perform
fast evaluations of geological uncertainty, such as in the geoscreening workflow. The multiscale SFI method
computes each timestep in two steps: First, it solves a nonlinear equation for pressure (and flow). Then, it
solves a nonlinear equation for saturations and mole fractions. The pressure equation is solved iteratively
using a multiscale approach.
The MS SFI method has recently been made generally available in a commercial reservoir simulator and
can easily be benchmarked with a state-of-the-art fully implicit (FI) method. The MS SFI method was used
to successfully simulate a realistic high-resolution geological model in a practical time frame, achieving
approximately 10 times speedup in CPU time compared to the FI method. This demonstrates the ability
of the MS SFI method to effectively deal with extremely complex models, enabling fast quantification
of geological uncertainty with a shorter turnaround time. In many instances the MS SFI method enables
simulation of large models at the original geological resolutions without the need for upscaling.
Finally, we demonstrate how the MS SFI method benefits a geology screening workflow and discuss
future use of the MS SFI framework to create fit-for-purpose simulation engines for other workflows.

Introduction
The FI method is considered the gold standard in terms of robustness, stability, and consistency for reservoir
simulation, and is offered in virtually every commercially available simulator today. The FI method solves
2 SPE-212231-MS

all physical processes at the same level of implicitness, and for every timestep the solution variables are
updated synchronously using an iterative solver.
Recently, an alternative solution called the SFI method (Jenny et al. 2006) has been made available in
a commercial simulator (DeBaun et al. 2005; Fjerstad et al. 2005; SLB 2022). The SFI method divides
the reservoir dynamics of each timestep into two subproblems, one for reservoir pressure and another
for the transport of reservoir fluids. This method improves performance and stability and provides a
flexible framework that can be configured with many different solver options for the pressure and transport
subproblems. For the pressure subproblem, this currently includes a multiscale restricted smoothed basis

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

(MsRSB) method (Lie et al. 2017; Møyner and Lie 2016a, 2016b; Kozlova et al. 2016) to accelerate
the solution of the pressure subproblem. The MsRSB method is especially well suited for the complex
unstructured grids used to model various operational applications targeted by commercial reservoirs
(Khataniar et al. 2022). Specialized solvers for the transport subproblem are hybrid upwind discretization
(Watanabe et al. 2016), special optimized linear solvers and a simple streamline option for geoscreening
(Shook and Mitchell 2009; Macé and Bogush 2017; SLB 2017). Finally, there are different ways to
iteratively improve the consistency between pressure and transport solution variables (Jiang and Tchelepi
2019).
The recovery of hydrocarbon resources is a complex undertaking where hard data about the reservoir
structure and attributes are scarce and soft data in the form of geological scenarios, structural interpretation
of seismic data, outcrops, and analogues are needed to build simulation models. The model building process
introduces uncertainty. A common practice today is to generate an ensemble of models that cover the space
of uncertainty in data. To quantify uncertainty in dynamic responses such as recovery factor, reservoir
simulation is needed. Unfortunately, it is infeasible to run field simulations for all models in an ensemble
to precisely capture the probability distribution for dynamic responses. Instead, only a few representative
models can be used for full field simulation. Choosing representative models is a crucial step in this process.
Selecting representative models and, more generally, ranking models with respect to heterogeneity is
an active research topic. Proposed measures of heterogeneity that can be used to rank models and predict
flow responses range from purely statistical measures like the Dykstra-Parsons coefficient to flow-based
heterogeneity measures, such as, for example, the Lorenz coefficient (Dykstra and Parsons 1950; Schmalz
and Rahme, 1950; Lake and Jensen 1991). More recently, there has been much research into flow-based
measures of heterogeneity that use simulation to quantify, such as flow capacity and sweep efficiency (Shook
and Mitchell 2009) and flow diagnostics (Shahvali et al. 2012; Møyner et al. 2014; Watson et al. 2020).
The process of screening, ranking, and selecting models is a complex task that requires a well-defined
workflow. Macé and Bogush (2017) introduced a workflow to systematically rank models in an ensemble
and select representative models for low, medium, and high recovery scenarios that accurately sample the
distribution of outcomes (SLB 2017). This geoscreening workflow uses a simplified simulation based on
streamlines to quickly generate the flow response to pressure gradients. These flow responses may be used
to estimate how structure and heterogeneity in an ensemble of models affect the distribution of outcomes
of simulation responses (e.g., recovery factor).
In this paper we present a commercial simulation engine for the geoscreening workflow presented in
Macé and Bogush (2017). The new engine uses the SFI solver framework with a multiscale method to
accelerate the pressure step and new optimized transport solvers tailored for the speed, accuracy, and
robustness required for the geoscreening workflow. To be precise, we have implemented a distributed
parallel streamline solver option that solves the transport subproblem very efficiently using time of flight
(TOF). We also have the option to use an optimized finite-volume transport for reduced physics. For the
model ranking and screening workflow, it is important that an ensemble of models can be run as quickly
as possible and without convergence issues, execution errors, etc. The new simulation engine is superior to
the FI method for this application. The new simulator is most attractive when applied for high-resolution
SPE-212231-MS 3

models; e.g., geological models with high resolution in the vertical direction, potentially by providing an
order-of-magnitude speedup compared to the standard FI solvers (Lie et al. 2017; Khataniar et al. 2022).
The paper is organized as follows. We first discuss the geoscreening workflow and describe the physics
assumed. We then give a brief introduction to the solvers used in this workflow, first the streamline TOF
algorithm with some examples. Then we provide a brief overview of SFI reservoir simulation technologies
that can be used for geoscreening; e.g., multiscale for pressure and subtimesteps for transport. Finally, we
demonstrate that the new SFI screening engine can accurately match reference results with significantly
less simulation time.

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

Geoscreening Workflow
The workflow described in Macé and Bogush 2017 exists as a plug-in in a commercial modeling tool for
use in subsurface uncertainty analysis workflows to screen, rank, and select geological models. Geological
uncertainty results in the generation of multiple models that honor the available data. As a time-bounded
activity, subsurface study teams often do not have the time or computing resources to execute detailed full-
field reservoir simulations on all these models.
The simplest approach is the deterministic one, where a single geological realization is considered with
"most likely" values of the input variables. An upscaled version of this model is passed along the study cycle
where it undergoes activities such as history matching (i.e., calibration to existing historical data), field
development planning, reserves estimation, etc. This approach has extremely limited capacity to capture
uncertainty.
As uncertainty is inherent in subsurface, a more appropriate approach is to consider multiple realizations
of the geological model. Preprocessor packages have excelled in generating ensembles of models based on
parameterization of the input models, with discrete or probability distribution functions of selected variables.
More recently, the parallelization of these processes along with the availability of cloud-based integrated
modelling and simulation have caused a step change in the scale achievable through these workflows (Elfeel
et al. 2021; Ramatullayev et al. 2021; El Dabbour et al. 2022). The vast number of realizations that can be
generated makes evident the need for a performant, fit-for-purpose simulation engine. Otherwise, generating
representative recovery S-curves of resource volumes and recovery across multiple development scenarios
becomes a costly, prohibitive task.
As a step up in complexity, an approach is to select representative realizations that would be able to
capture the uncertainty. The simplest approach is by evaluating the static responses (e.g., stock tank oil
initially in place—STOIIP) and picking high, medium, and low probable models (P90, P50, and P10).
There are two fundamental issues in this approach: first, there is no guarantee that three cases will be
sufficient to represent the original ensemble; second, the process takes no account of dynamic model
response uncertainty, so when a realization passes to dynamic simulation there is no guarantee that the static
model ranking will be preserved. If a model has a P10 static response but closer to a P50 dynamic response,
this can lead to wasted study time or even worse, poor decisions.
Therefore, incorporating dynamic responses is most desirable if the cost and complexity of such an
approach is far less than carrying out a detailed production forecasting simulation. Simplified numerical
experiments using streamline simulation is an efficient and accurate technique to calculate flow response in
a reservoir model. For instance, streamlines can be used to calculate dynamic heterogeneity that measures
how efficiently connate fluid can be displaced by an injected fluid. This metric is well-suited to screen and
select geological models because it depends only on the reservoir shape and structure and the variation and
distribution of permeability and porosity.
In our geoscreening engine, a base case built in the commercial simulator can be used as a starting
point. A single instruction from the user overrides the input into the simulator and eases the transition
between a full simulation model and numerical flow experiments used in a screening study. Single-phase
4 SPE-212231-MS

displacement of "connate" fluid (i.e., fluid that was present in the reservoir at initialization) by "injected"
fluid (i.e., fluid entering the model from wells or boundary conditions) is calculated. Both fluids have
identical characteristics: they are incompressible with identical density and viscosity. Therefore, gravity
will not segregate the two fluids. Linear relative permeabilities and zero capillary pressure are used to
represent their rock physics interaction. Flexibility is given for engineers to alter these assumptions and
make the model as complex as necessary, but this comes at the cost of usability and may also gradually
impact performance.
To set up flow experiments, two types of boundary conditions can be applied: either a specified pressure

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

drop between geometrical boundaries of the model (e.g.: bottom-to-top, south-to-north, east-to-west, or in
the reverse directions), or a pressure drop between producers and injectors in a well configuration defined
either explicitly by the user or through pre-defined pattern schemes. Important output quantities are the total
produced connate fluid from the experiment and pore volume visited by the streamlines. These quantities can
be used to assess connectivity and to provide adequate dynamic responses to incorporate in the workflow.
The simulation is performed by a fit-for-purpose engine that takes advantage of many opportunities for
optimization in its approach, and dramatically reduces computational time. We explain this in the next
sections in detail.

Solution Methods
The simplification of the physics used in the geoscreening workflow enables optimizations in the numerical
methods used by the simulator. In the absence of gravitational and capillary forces, the single-phase flow
experiments decouple the pressure dynamics from the evolution of fluid transport. This makes the reservoir
pressure field stationary. Fig. 1 shows flow diagrams for the difference between the FI method, the SFI
method, and the SFI geoscreening method that is optimized for single-phase and two-phase models. The FI
method in Fig. 1a linearizes and solves for all solution variables simultaneously and is not optimized for
the simplifications in physics used in the geoscreening workflow.

Figure 1—Summary of (a) the FI method, (b) the MS SFI method, and (c) the MS SFI geoscreening solver.

The SFI method in Fig. 1b splits the material balance equations for reservoir fluids into a pressure
subproblem and a transport subproblem. In the pressure subproblem, we solve an equation for fluid pressure
that balances the total volume of fluids in every cell to the pore volumes at the end of every timestep. In the
transport subproblem, we solve an equation for the cell-by-cell material balance of each fluid in the reservoir,
where pressure, total fluid volume in each cell, and the total volume flux between cells are kept fixed for
the duration of a timestep. This method is better suited to optimize simulation for the physics assumed by
SPE-212231-MS 5

the geoscreening workflow and the accuracy needed in the output; e.g., by using a multiscale method to
accelerate the solution of the pressure subproblem, or a streamline method for the transport subproblem to
get subcell resolution of fluid fronts. The SFI geoscreening solver shown in Fig. 1c is a specialization of
the SFI method that is optimized for a single-phase problem with simplified physics where the pressure
field is stationary.
With the SFI geoscreening solver we can compute the reservoir pressure once at the beginning of the
simulation and only compute transport of fluids in the subsequent timesteps. This simplifies the equations
that need to be solved by the simulator and enables very optimized solvers for the transport subproblem.

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

For more details on the multiscale SFI in our commercial simulator, readers are referred to Kozlova et al.
(2016), Lie et al. (2017), and SLB (2022).
Each of the methods above has value in different use case applications. It is invaluable that the commercial
simulator used for subsurface workflows can transition between the formulations, as this creates a toolset for
geologists, geoscientists, and reservoir engineers to peruse, minimizing the transition and data flow between
the workflows and methodologies, and enabling different levels of realism in the physical modelling—from
simplified single-phase flow experiments to full compressible multiphase models. We will now go into the
details of the methodologies applied in the MS SFI geoscreening engine.

Streamline Transport Solution

Streamline methods have been applied in a wide variety of tools and workflows. To name a few: pure flow
path visualization for sense checking flow in reservoir models, computing dynamic measures of connectivity
for screening and optimization workflows as transport solvers for general-purpose reservoir simulation and
water flooding (Bratvedt et al. 1992; Batycky 1997; Datta-Gupta and King 2007), compositional reservoir
simulation (Crane et al. 2000) and in miscible improved oil recovery (Bratvedt et al. 2014) and polymer
flooding (Thiele et al. 2010).
Streamline tracing has been implemented in our simulator engine to support several workflows. For
instance, streamlines can be used in optimization workflows where streamlines are traced to calculate
connectivity for reports or used in optimization (Elfeel et al. 2021). For this purpose, we can use the
results from an FI solver to compute the velocity field to trace the streamlines. To be efficient in this
simulator environment, the implementation of streamline tracing supports multiprocess parallel execution
with distributed parallel subdomains.
If streamlines are used as a transport solver, it requires a formulation where pressure and transport
(saturations or components) are solved separately and in sequence. The multiscale SFI framework that has
recently been made available in the simulator offers exactly the flexibility where different transport solvers
may easily be implemented. Transport solvers based on streamlines fit perfectly into this framework. The
fluxes calculated based on the pressure solve can then be used to trace the streamlines.
For geological model screening workflows, we are interested in flow responses from the geological
models that can be used to measure dynamic heterogeneity; that is, how the structure and heterogeneity
of a geological model affect the efficiency of displacement of the connate fluid by an injected fluid with
identical characteristics. If the injected fluid front, that is, the boundary between injected and connate fluids
becomes distorted, generating preferential flow paths and stagnant regions where connate fluid is by-passed,
the displacement process is inefficient. Complex nonlinear fluid behaviors caused by gravity, capillary and
viscosity do not significantly alter the effect of dynamic heterogeneity on the displacement process (Lake
and Jensen 1991) and are therefore not needed to rank geological models.
Flow responses from incompressible fluids and tracer-like transport are preferred assumptions at the
early stages of the geoscreening workflow. A simple front-tracking solver (Bratvedt et al. 1992) is ideal
under these assumptions and with a steady-state flow we can even simplify further to use the TOF along
the streamlines (King and Datta-Gupta 1998) directly to find the position of the fluid front. In Fig. 2 we
show an example of TOF computed using streamlines with different boundary conditions. The density of
6 SPE-212231-MS

streamlines in this plot is low compared to what is needed when streamlines are used as a transport solver.
For a streamline transport solver, it is important to cover all parts of the reservoir that have flow. This usually
requires orders-of-magnitude higher streamline density than what is typically necessary for visualization
purposes.

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

Figure 2—Streamline solution applied to a modified version of the SPE10 model (Christie
and Blunt 2001) with a different set of boundary conditions to assess connectivity.

Saturations and rates for a given time are computed by accumulating the volumes from all streamlines
passing a cell and rates from streamlines ending in a producer. This is very cost effective, the cost for this
solver is neglectable compared to the cost of tracing streamlines.

Sequential Fully Implicit Method

The framework provided by the SFI method enables a flexible and configurable simulator design that can
be optimized for different applications. Currently, the version of the SFI method that has been released in
our simulator has been equipped with solvers and technology that makes it well suited for large field scale
models. In our internal testing framework, we have achieved performance that in most cases is superior to
the FI method, with some examples illustrated in Khataniar et al. 2022.
The current default implementation of SFI uses a standard implicit in time finite-volume formulations
for both the pressure and transport subproblems. For each timestep, we solve two systems of nonlinear
equations, first one for the pressure subproblem, then one for the transport subproblem. Each nonlinear
system is solved using Newton-Raphson iterations: in each iteration linearize the system of equations and
solves a linear system for increments of the solution variables in each grid cell. The pressure and transport
nonlinear solvers have separate convergence reports and separate criteria for convergence control. In case
simulation tuning is needed, this gives users the ability to assess what criteria may need to be tightened or
relaxed to achieve better accuracy, stability, and performance.
The default setup for the pressure solver is to use an MsRSB method (Møyner and Lie, 2016) to solve the
linear system in each nonlinear iteration. The multiscale method works as follows: First, as a preprocessing
step, the original fine-scale grid is divided into nonoverlapping coarse blocks forming a coarse grid. Each
SPE-212231-MS 7

coarse block is equipped with a basis function that is a numerically computed pressure response to a local
change in pressure; the response is nonzero only in the coarse block and a small collection of nearby grid
cells. The METIS algorithm (Karypis 2013) is used to define the coarse partitions. During the pressure
computation, we form a linear system for pressure in the coarse blocks, and the coarse grid pressure is
interpolated back to the fine grid using the basis functions. The computational savings come from solving
smaller linear systems for coarse pressure. For a more detailed description, the reader is referred to Watanabe
et al. in press.
For the transport subproblem, the default setup in our implementation of the SFI method is to use an

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

implicit in time finite-volume method with standard phase-based upwind weighting. For each linear system
that is in the nonlinear loop, we use a flexible generalized minimal residual iteration with an incomplete
lower-upper factorization with zero level of fill-in (ILU0) as preconditioner. For a more detailed description
see SLB (2022).
If needed, the pressure and transport steps can be repeated to ensure that the pressure and transport
solution variables are consistent. This outer iteration is controlled by a third set of convergence controls.
The sequential formulation provides several advantages here:
- Transport adaptivity can be used to reduce the cost (e.g., Lee et al. 2009) by localizing the
linearization of transport equations to where solutions are changing, only.
- Subtimestepping can be employed to facilitate convergence, improve solution accuracy, or to
improve performance. For instance, the pressure field can be solved only once and subtimesteps can
be taken for the transport solution (Fig. 3) to get sufficient resolution in, for example, fluid fronts
or production rates,
- Specialized discretization schemes, such as hybrid upwind weighting, or nonlinear solvers; for
example, a trust-region approach can be applied (Watanabe et al. 2016) to improve the stability of
the Newton-Raphson solvers.
- Topological-order Gauss-Seidel: the transport-only linear system is easier to solve, typically, basic
ILU0 preconditioning works well. More tailored special ordering strategies (Natvig and Lie 2008)
can be used to achieve better performance if the fluid physics is simple enough.

Figure 3—Illustration of subtimestepping in geoscreening, where a fixed pressure

field is solved once, and transport is solved many times within that period.

After the transport nonlinear iterations are complete, we check if pressure and transport solution variables
are sufficiently consistent and may either proceed to the next timestep or repeat pressure and transport steps
once again. A summary of the multiscale SFI method is presented in Fig. 1b.

Solution Method Optimizations for Geoscreening

The geoscreening SFI solver is set up to be efficient for the simplified physics setup described in the
Geoscreening Workflow section. This setup ensures that the reservoir pressure in each grid cell does not
change during the simulation. Therefore, the SFI geoscreening solver computes pressure once with a large
timestep at the beginning of the simulation. This pressure solution uses the multiscale MsRSB solver. For
transport, two different choices exist.
The first choice is a finite-volume solver that uses subtimestepping to separate the timestepping in the
transport subproblem from the large timestep in the pressure solver. Each subtimestep is like a regular
8 SPE-212231-MS

timestep; we solve a system of nonlinear equations using Newton-Raphson iterations, and in each nonlinear
iteration we solve a linear system using an optimized linear solver.
The second choice is to use a streamline method to trace streamlines and compute TOF. Then for each
timestep that is requested, TOF is used to map a solution profile along each streamline and compute the rate
of connate and injected fluids in each producer or boundary condition.

Computational Stratigraphy Model

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

To test the performance of the geoscreening engine, we run a direct simulation of high-resolution synthetic
model created using the computational stratigraphy (CompStrat) modeling framework (Willis and Sun 2019;
Willis et al. 2021a, 2021b; Li et al. in press). CompStrat is Chevron proprietary technology for reservoir
characterization and modeling that can produce very realistic geomodels with high resolution, especially in
the vertical direction. A typical model may have hundreds to a few thousand layers, translating into hundreds
of millions to a few billions of cells. Consequently, reservoir simulation of CompStrat models is particularly
challenging because of the cell count and the extreme aspect ratio (thinness) of the cells.
In this section we compare SFI geoscreening method against the FI method in the model shown in Fig.
4a. The extent of the model is about 800 m in the lateral direction and about 18 m in the vertical direction
and the grid has more than 4.4 million active cells. There are about 40 cells in the lateral directions and 2,800
cells in the vertical direction, which gives an average cell aspect ratio of more than 3,000. The permeability
field is highly heterogeneous with a ratio of vertical to lateral transmissibility of 1e7:1.

Figure 4—Computational stratigraphy model. The figure shows (a) the J-directions
transmissibility field and (b) a 2D cross section showing the J-direction transmissibility
field. The scale on the vertical axis is increased 50 times to make details visible.

In the next sections we run single-phase and two-phase experiments, first on a 2D cross-section of this
model and afterwards on the full 3D model. In the single-phase experiments, the connate fluid is displaced
by an injected fluid with identical characteristics. The fluid properties of these two fluids are listed in Table
1. For the rock physics, we use linear relative permeability curves and zero capillary pressure. Fig. 7 and
Table 2 shows the fluid and rock properties for the two-phase experiments. For simplicity, we use "oil" to
refer to the connate fluid and "water" to refer to the injected fluid in both cases.
SPE-212231-MS 9

Table 1—Fluid and rock properties for the single-phase experiments.

Description Unit Value

Surface mass density kg/m3 998.98

Viscosity cp 1
Compressibility 1/bar 0
Formation volume factor bbl/STB 1
Rock compressibility 1/bar 1.0e-8

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

Table 2—Fluid and rock properties for the two-phase experiments.

Description Unit Value

Surface mass density oil kg/m3 875.30

Surface mass density water kg/m3 998.98
Water compressibility 1/bar 3.98e-5
Water formation volume factor bbl/STB 1.0132
Water viscosity cp 0.3985
Pore volume compressibility 1/bar 1.20e-4

Cross-Sectional Model
To illustrate a potential workflow, a synthetic 2D cross-sectional model is used in this section. This model
is created as an I-cross section of the computational stratigraphy model (Fig. 4a). The high contrast in
permeability and the high horizontal to vertical grid aspect ratio makes the transmissibility values span five
orders of magnitude as shown in Fig. 4b.
Single-Phase Flow Experiment. We conduct a simple flow experiment where an in-place fluid is displaced
by an identical injected fluid. The injection is along the left edge of the domain and production at the
right edge. This flow experiment is designed to evaluate the reservoir connectivity efficiently with a fast
streamline TOF solver as a geoscreening workflow. The flow physics is simplified as described in the
previous section; the two fluids are incompressible with identical viscosities and linear relative permeability.
We specify simulation time in fractions of one pore volume injected (PVI) fluid. For this case, the
simulation time is 0.5 PVI. Two simulation runs with eight cores are conducted with (1) the FI method
and (2) the SFI geoscreening method with a streamline TOF solver. Fig. 5 shows the saturation of injected
fluid at the end of the simulation for the two solvers. The solution computed with the FI method (Fig. 5a)
has significant numerical dispersions in the injected fluid front while the streamline TOF solver Fig. 5b
maintains a sharp front between the injected and displaced fluids with minimal grid resolution and grid
orientation effects because the transport equation is solved along streamlines at higher resolution than the
grid resolution. Field responses for the fraction of injected fluid in the producer is shown in Fig. 6 (left).
The main objective of this flow experiment is to quickly evaluate the flow connectivity inside the geological
model and visualize the connection by streamlines and quantify the reservoir connectivity by the front
arrival time. In this example, the FI run finished in 192 seconds while streamline TOF solver finished in 31
seconds, which corresponds to 6.2 times speedup, as shown in Fig. 6 (middle).
10 SPE-212231-MS

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

Figure 5—Single-phase flow experiment for cross-sectional model. The plot shows the injected
fluid saturation computed with (a) the FI method, (b) the MS SFI geoscreening method with
streamline TOF transport solver, and (c) the injected fluid saturation plotted on streamlines.

Figure 6—Single-phase flow experiment for cross-sectional model. The plot shows the field water
cut (left), the total CPU time (middle), and the timestep length (right) with the FI method (green solid
line) and the MS SFI geoscreening method with streamline TOF transport solver (blue dashed line).

Two-Phase Flow Experiment. After the simple flow experiments in the geoscreening workflow are
conducted and geological uncertainty is assessed by geologists and geophysicists, reservoir engineers can
pursue to include more complexity of fluid and rock modeling. Our holistic framework of MS-SFI engine
in our simulator enables us to experiment with the more complex flow dynamics.
In this example we replace the single-phase fluid model with a two-phase oil-water system to see how test
results change with more accurate fluid physics. The oil and water phases have different phase viscosities,
surface densities, formation volume factors, and relative permeability curves (Fig. 7, Table 2) The simulation
scenario is the same, the in-place oil is displaced by injected water, injection happens at the left boundary
of the domain, and production from the right boundary.

Figure 7—Fluid properties for the two-phase experiments.

In this case, we compare the FI method with an SFI geoscreening solver with a finite-volume transport
solver to account for gravity effect between oil and water fluid density differences in the flow dynamics. For
SPE-212231-MS 11

each solver we have run a simulation with 24 cores. The total simulation time is set to 791 days (0.2 PVI) to
see the water break through at the producer. The water saturation at the end of simulation is shown in Fig.
8. In this case, the flow experiment includes gravity that may reduce the effect of numerical dispersion in
the FI method Fig. 8 (left). The SFI geoscreening finite-volume transport solver Fig. 8 (right) has numerical
dispersion similar to the FI method in this case.

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

Figure 8—Two-phase oil-water flow experiment with cross-sectional model. The plot shows the water saturation at 0.2 PVI
computed with the FI method (left) and the MS SFI geoscreening method using a finite-volume transport solver (right).

Fig. 9 (left) shows the water cut for the two solvers. The agreement of the simulation results validates
the implementation of finite volume transport solver for geoscreening application.

Figure 9—Two-phase flow experiment for cross-sectional model. The plot shows the field water cut
(left), the total CPU time (middle), and the timestep length (right) with the FI method (green solid
line) and the MS SFI geoscreening method with finite-volume transport solver (blue dashed line).

Fig. 9 (middle) shows that FI run can finish in 5,984 seconds while the SFI geoscreening run with finite-
volume transport finishes in 487 seconds, which corresponds to 12 times speedup. The speedup is largely
due to better stability, which results in larger timestep lengths, Fig. 9 (right). In comparison with the simple
flow experiment solution, the water injection displacement is affected by the gravity segregation.

Full 3D Stratigraphy Model

In this section we will apply the SFI geoscreening solver to directly simulate single-phase and two-phase
displacement scenarios on the full 3D fine-scale CompStrat model. Simulation of the CompStrat model is
challenging even when the fluid model and recovery physics are simplified.
12 SPE-212231-MS

Single-Phase Flow Experiment. The first experiment uses a single-phase fluid model with properties
listed Table 1. A fluid is injected near one corner, displacing the in-place fluid. A producer is placed in the
opposing corner. The two fluids have identical viscosities and densities, and linear relative permeability
curves. The FI solver is run with 100 cores on our internal cluster, while the SFI geoscreening solver with
finite-volume transport is run on 10 cores. The total simulation time is 1 PVI or 22,622 days.
Fig. 10 shows the field-level dynamic responses for the two solvers. All responses are in excellent
agreement. The SFI geoscreening run can finish in 276 seconds while the FI method can finish in 20,893
seconds, which corresponds to two orders of magnitude speedup (75 times to be precise). The total CPU

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

time vs. simulation time can be seen in Fig. 11. We also want to emphasize that in this example, the FI
method uses 10 times more cores than the SFI geoscreening method.

Figure 10—Simulation results for the computational stratigraphy model with single-phase flow. The plots show the FI
method (green solid line); and for the MS SFI geoscreening method with finite-volume transport (blue dashed line).

Figure 11—The computational stratigraphy model with single-phase flow. The plots show total CPU time
(left), total CPU time with zoomed y-axis (middle) and timestep length (right), with the FI method (green solid
line) and for the MS SFI geoscreening method with finite-volume transport (blue dashed line). The timestep
length reported for MS SFI geoscreening method is the timestep used for the finite-volume transport.

Two-Phase Flow Experiment. In the second experiment, we run a water-flooding scenario. Like in the first
example, water is injected in a well placed near one corner ("Inj" to the left in Fig. 12) and oil is produces
from a well in the opposing corner ("Prod" to the right in Fig. 12); the initial oil saturation is 0.8, and the
injection of water last simulation period is 1,522 days or approximately 0.14 PVI. The water saturation at
the end of simulation is shown in Fig. 12. For this case, the pressure is computed on every timestep and
the transport solver uses the hybrid upwind discretization for viscous and gravity flows (Lee et al. 2015) to
SPE-212231-MS 13

stabilize the nonlinear convergence and to prevent flip-flopping of upwind directions for flux. Both the FI
solver and SFI solver are run with 64 cores on our internal cluster.

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

Figure 12—Computational stratigraphy model. The figure shows water saturation profile at the end of simulation.

For this experiment, the solutions from the multiscale SFI method are in good agreement with the FI run
in terms of the field responses (Fig. 13). However, the FI method struggles to converge and cuts the timestep
size to prohibitively small values, while the multiscale SFI method can converge longer timesteps, as can
be seen in Fig. 14 (right). The runtime, number of timesteps, and discarded timesteps are listed in Table
3. Also, note the dramatic difference in CPU time between the FI and multiscale SFI runs in this test case:
here, the multiscale SFI method is six times faster than the FI method.

Figure 13—Simulation results for the computational stratigraphy model with water-oil two-phase
flow. The plots show the FI method (green solid line) and the MS SFI method (blue dashed line).
14 SPE-212231-MS

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

Figure 14—The computational stratigraphy model with water-oil two-phase flow. The plots show the total
CPU time (left), the total CPU time with zoomed y-axis (middle), and the timestep length (right) for the FI
method (solid green line) and the MS SFI method with a finite-volume transport solver (blue dashed line).

Table 3—Performance comparison for the computational stratigraphy model with water-oil two-phase flow.

Description FI MS SFI

Timesteps 76,477 4,347

Timestep cuts 807 187
Total CPU time 118,703 19,829

Illustration of Geoscreening Workflow in Channelized Reservoir

After demonstrating the application of the engine to a deterministic, large case for flow diagnostics in detail,
we investigate the workflow of geological screening within a synthetic full-field scale waterflood model.
Geoscreening workflows can take many variations. The one proposed here is simplistic as it aims to
demonstrate a proof of concept for the engine. The steps to be followed are:
1. Definition of a ‘base case’ geological model
2. Definition of uncertain parameters and their range and distribution
3. Set up of geoscreening SFI numerical experiment (e.g., definition of wells as boundary conditions)
4. Run realizations of geoscreening SFI under a one variable at a time (OVAT) sensitivity, evaluating
the effect at the extreme of the parameter ranges
5. Run many numerical experiments varying all the parameters using Monte-Carlo sampling
6. Perform case selection by analyzing the results
In our base case geological model, the corner point grid has 2.33 million active cells distributed at an
area of 6000 × 4800 × 100 m, with an average cell dimension of 25 × 25 × 1 m. A stochastic object
modelling approach is used, with fluvial channels and adaptive channels. Certain parameters are replaced
by a probabilistic distribution function.
The fluvial channels take a fraction of the model that varies according to a uniform distribution function
ranging from 20 to 40 percent (variable name: FC_FRAC). The channel orientation, amplitude, wavelength,
as well as the width and thickness of its bodies, vary within the channels themselves using a triangular
distribution. The same approach is taken to the adaptive channels, except instead of a fraction the number
of channels is varied (variable name: N_CH), from 15 to 25. A seed parameter adds to the randomness of
the problem. Random realizations are displayed in Fig. 15.
SPE-212231-MS 15

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

Figure 15—Nine random realizations of the facies model generated from the uncertainty workflow.

The facies parameters guide the porosity and permeability values. The reservoir is divided into two halves,
namely a top and bottom zone:
- Top zone
∘ Background facies: 1% porosity, 1 md permeability
∘ Levee: 4% porosity, 40 md permeability
∘ Channel: 10% porosity, 400 md permeability
- Bottom zone:
∘ Background facies: 5% porosity, 10 md permeability
∘ Levee: 8% porosity, 600 md permeability
∘ Channel: 14% porosity, 1500 md permeability
Two multipliers are applied, one to the top zone properties (variable name: TOP), the other to the bottom
zone properties (variable name: BOTTOM). Their distribution is uniform, with values ranging from 0.75 to
1.25. Last, the ratio of vertical to areal permeability (variable name: KVKH) is also uncertain. A uniform
distribution ranging from 0.0001 to 1 is used.
Thirty vertical wells are distributed in the reservoir, being 16 producers and 14 water injectors. These
serve as boundary conditions to the numerical experiments. The reservoir is initially fully containing connate
fluid, and the injectors displace it with injected fluid. We run all the property calculations described above,
and then inject the same pore volume as initially in the reservoir. At the end, we analyze two responses:
16 SPE-212231-MS

- The stock tank oil initially in place (STOIIP) from a simple volumetric multiplication of pore volume,
hydrocarbon saturation, and formation volume factor
- The connate fluid production cumulative (CFPC) from the dynamic numerical connectivity
experiment using the geoscreening SFI engine.
First, we perform an OVAT sensitivity to measure the responses at the extremes of their distribution. Ten
geoscreening SFI cases are run by setting one of the variables at their minimum or maximum value in its
defined range and maintaining the base-case value for all other variables. The results are analyzed through

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

tornado plots, where we can visualize the most impacting variables and how they affect the individual
responses (Fig. 16).

Figure 16—Tornado plot showing results from OVAT sensitivity into responses.

Next, 512 realizations are generated through Monte-Carlo sampling. Each realization takes about 40
seconds to run in 32-way parallel on an AMD EPYC 7532 CPU, with only about 4 seconds spent into
the pressure solution and 6 seconds for the streamline solver, with remaining time in tasks such as grid
processing, data loading, and reporting. Considering a single node was available and simulations of 32-way
had to be run sequentially, the entire 512 model ensemble would be processed in around 5.7 hours. If we
consider a total of 1,024 cores available, we would be able to run 32 simulations concurrently, and the total
elapsed time for the ensemble would drop down to less than 11 minutes.
To put these numbers in perspective, a single run of a realistic scenario, full physics setup of this model
running in FI can be completed in around 36 minutes.
To select representative cases, a simple method is to cross-plot the two responses (one static, one
dynamic), and interactively select from areas representing P10-P10, P50-P50, and P90-P90. Fig. 17 shows
this process for the 512 cases generated from the method.
SPE-212231-MS 17

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

Figure 17—Cross-plot of dynamic and static responses for selection of representative geological
cases. The colored lines outline the P10, P50, and P90 areas within a ±5 range of variation.

To further enforce that the rank is preserved, a plot can be created of the most meaning variables
from sensitivity and how these selected cases correlate. For instance, from the tornado plot we can see
that FC_FRAC (fraction of channels in the reservoir model) is of reasonable importance for both static
and dynamic responses. A small fraction of channels should correlate to a pessimistic (low volume, low
connectivity case), while a large fraction should correlate to an optimistic (high-volume, high-connectivity
case). The plots in Fig. 18 show only cases within the representative quadrants and can be used to confirm
that the selections (solid circles) are adequate or if they need to be revisited.

Figure 18—Sense check of selected cases (solid circles) against a meaningful input parameter (fraction of channels
in the reservoir model, FC_FRAC) for both the static response (STOIIP) and the dynamic response (CFPC). The non-
filled circles correspond to unselected cases in a given P10 (green), P50 (blue), and the P90 (red) percentile areas.
18 SPE-212231-MS

Conclusions
We demonstrated a fit-for-purpose engine for geological screening applications, referred to as geoscreening
SFI. It combines a multiscale pressure solver, streamline transport solver, and several other techniques to
optimize performance and accuracy into the SFI framework, and makes it available for this tailored use
case. As result, the engine can handle large-scale, heterogeneous models with considerable speedup. This
is fundamental to unlocking workflows even under large ensembles.
We first show this application in a deterministic case, which is to a complex, heterogenous model

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

generated by the Chevron computational stratigraphy group. We note improvements in both accuracy
(reduced numerical dispersion) and speed (up to two orders of magnitude).
The same engine is applied to a full geological screening workflow package and can be used to make
assessments on geological uncertainty and select representative cases to bring forward into the next steps
of the study. With enough dedicated hardware, we show that the evaluation using simplified numerical
experiments can be achieved on an ensemble of hundreds of cases in a fraction of the time it would take
to run a single full-physics simulation case.
Though not demonstrated directly in this paper, the same engine could be extended for other use cases
such as upscaling validation, well placement through connectivity mapping, and general model sense
checking.
Another key aspect is that this workflow is built in the context of a general-purpose commercial reservoir
simulator, and in doing so makes the transition between simplified and full physics a lot more palatable to the
end user. Having a unified tool set to address different problems is something very appealing to geologists,
geoscientists, and engineers alike. Sequential methods are an ideal framework for fit-for-purpose engines,
and many others may be developed in future work, such as an engine dedicated to solving waterflood
simulations.

Acknowledgments
We would like to thank Boxiao Li (Chevron) for providing the models and several inspiring discussions.
The authors are grateful to Seong Lee (Chevron) and Hamdi Tchelepi (Stanford University) for many
constructive discussions on multiscale and sequential formulations. Many thanks go to Barış Güyagüler and
Xianhuan Wen (Chevron), as well as Sanjoy Khataniar and Kyrre Bratvedt (SLB) for their constant support.

References
Batycky, R.P. 1997. A Three-Dimensional Two-Phase Field Scale Streamline Simulator. PhD thesis, Stanford University,
Stanford, California, USA.
Bratvedt, F., Bratvedt, K., Buchholz, C. F. et al. 1992. A New Front-Tracking Method for Reservoir Simulation. SPE Res
Eng 7 (1): 107–116. SPE-19805-PA. https://doi.org/10.2118/19805-PA.
Bratvedt, F., Yi, T., Shaker, A. et al. 2014. Development of Streamline Simulation for Miscible Recovery Process and
its Application in an Onshore Oil Field–Abu Dhabi. Paper presented at the SPE Abu Dhabi International Petroleum
Exhibition and Conference, Abu Dhabi, UAE, 10–13 November. SPE-172132-MS. https://doi.org/10.2118/172132-
MS.
Christie, M. A. and Blunt, M. J. 2001. Tenth SPE comparative solution project: A comparison of upscaling techniques.
SPE Res Eval & Eng 4 (04): 308–317.
Crane, M., Bratvedt, F., Bratvedt, K. et al. 2000. A Fully Compositional Streamline Simulator. Presented at the
SPE Annual Technical Conference and Exhibition, Dallas, Texas, USA, 1–4 October. SPE-63156-MS. https://
doi.org/10.2118/63156-MS.
Datta-Gupta, A., King, M. J. 2007. Streamline Simulation: Theory and Practice. Texas: Society of Petroleum Engineers.
https://doi.org/10.2118/9781555631116.
DeBaun, D., Byer, T., Childs, P. et al. 2005. An Extensible Architecture for Next Generation Scalable Parallel Reservoir
Simulation. Paper presented at the SPE Reservoir Simulation Symposium, January 31–February 2. SPE-93274-MS.
https://doi.org/10.2118/93274-MS.
SPE-212231-MS 19

Dykstra, H., Parsons, R.L., 1950. The prediction of Oil Recovery by Water Flood. Secondary Recovery of Oil in the
United States, API, New York.
El Dabbour, M. H., Labib, A., Soliman, A. et al. 2022 Cloud-Based Agile Reservoir Modelling Enriched with Machine
Learning Improved the Opportunity Identification in a Mature Gas Condensate. Paper presented at the ADIPEC, Abu
Dhabi, UAE, October 31–November 3. SPE-211632-MS. https://doi.org/10.2118/211632-MS.
Elfeel, M. A., Goh, G., Biniwale, S. 2021. Advanced Completion Optimization ACO: A Comprehensive Workflow for
Flow Control Devices. Paper presented at the International Petroleum Technology Conference, Virtual, March 23–
April 1. IPTC-21189-MS. https://doi.org/10.2523/IPTC-21189-MS.
Fjerstad, P.A., Sikandar, A.S., Cao, H. et al. 2005. Next Generation Parallel Computing for Large-Scale Reservoir

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

Simulation. Paper presented at the SPE International Improved Oil Recovery Conference in Asia Pacific, December
5–6. SPE-97358-MS. https://doi.org/10.2118/97358-MS.
Jenny, P., Lee, S.H., and Tchelepi, H.A. 2006. Adaptive fully implicit multi-scale finite-volume method for multi-
phase flow and transport in heterogeneous porous media. J Comp Phys 217 (2): 627–641. https://doi.org/10.1016/
j.jcp.2006.01.028.
Jiang, J. and Tchelepi, H.A. 2019. Nonlinear acceleration of sequential fully implicit (SFI) method for coupled flow
and transport in porous media. Computer Methods in Applied Mechanics and Engineering 352: 246–275. https://
doi.org/10.1016/j.cma.2019.04.030.
King, M. J. and Datta-Gupta, A. 1998. Streamline Simulation: A Current Perspective. In Situ 22 (1): 91–140.
Karypis G. 2013. A software package for partitioning unstructured graphs, partitioning meshes, and computing fill-
reducing ordering of sparse matrices version 5.1.0.
Khataniar, S.K., Dias, D. D, and Xu, R. 2022. Aspects of multiscale flow simulation with potential to enhance reservoir
engineering practice. SPE J. 27(01): 663–681. SPE-203996-PA. https://doi.org/10.2118/203996-PA.
Kozlova, A., Li, Z., Natvig, J. R. et al. 2016. A real-field multiscale black-oil reservoir simulator. SPE J. 21(06):
2049–2061. SPE-173226-PA. https://doi.org/10.2118/173226-PA.
Lake, L. and Jensen, J.L., 1991. A Review of Heterogeneity Measures Used in Reservoir Characterization. In Situ 15(4):
409–440.
Lee, S. H., Zhou, H., and Tchelepi, H. A. 2009. Adaptive multiscale finite-volume method for nonlinear multiphase
transport in heterogeneous formations. J Comp Phys 228(24), 9036–9058. https://doi.org/10.1016/j.jcp.2009.09.009.
Lee, S. H., Efendiev, Y., and Tchelepi, H. A. 2015. Hybrid upwind discretization of nonlinear two-phase flow with gravity.
Adv Water Resour 82: 27–38. https://doi.org/10.1016/j.advwatres.2015.04.007.
Li, B., Wen, X.-H., and Sun, T. In press. Bridging Computational Stratigraphy and Reservoir Simulation for Geologically
Realistic High-resolution Reservoir Modeling. Paper presented at the SPE Reservoir Simulation Conference,
Galveston, Texas, USA, 28-30 March. SPE-212244-MS.
Lie, K.-A., Møyner, O., Natvig, J.R. et al. 2017. Successful application of multiscale methods in a real simulator
environment. Comp Geosci 21: 981–998. https://doi.org/10.1007/s10596-017-9627-2.
Macé, L. and Bogush, A. 2017. Improved Resources and Reserves Estimation with the Combined Use of Scenario-Based
Geological Models and Flow Based Connectivity. Paper presented at the SPE Middle East Oil & Gas Show and
Conference, Manama, Kingdom of Bahrain, 6–9 March. SPE-183916-MS. https://doi.org/10.2118/183916-MS.
Møyner, O., Krogstad, S., and Lie, K.-A. 2014. The application of flow diagnostics for reservoir management. SPE J. 20
(2): 306–323. https://doi.org/10.2118/171557-PA.
Møyner, O. and Lie, K.-A. 2016a. A multiscale restriction-smoothed basis method for high contrast porous media
represented on unstructured grids, J Comp Phys 304: 46–71. https://doi.org/10.1016/j.jcp.2015.10.010.
Møyner, O. and Lie, K.-A. 2016b. A Multiscale Restriction-Smoothed Basis Method for Compressible Black-Oil Models.
SPE J. 21 (06): 2079–2096. SPE-173259-PA. https://doi.org/10.2118/173265-PA.
Natvig, J. R. and Lie, K.-A. 2008. Fast computation of multiphase flow in porous media by implicit discontinuous
Galerkin schemes with optimal ordering of elements. J Comp Phys 227 (24): 10108–10124. https://doi.org/10.1016/
j.jcp.2008.08.024.
Ramatullayev, S., Su, S., Rat, C. et al. 2021. The Intelligent Field Development Plan Through Integrated Cloud Computing
and Artificial Intelligence AI Solutions. Paper presented at the Abu Dhabi International Petroleum Exhibition &
Conference, Abu Dhabi, UAE, November 15–18. https://doi.org/10.2118/208106-MS.
Shahvali, M., Mallison, B., Wei, K. et al. 2012. An alternative to streamlines for flow diagnostics on structured and
unstructured grids. SPE J. 17 (3): 768–778. https://doi.org/10.2118/146446-PA.
Schmalz, J.P. and Rahme, H.S. 1950. The Variation in Water Flood Performance with Variation in Permeability Profile,
Producers Monthly 14 (9).
SLB. 2017. Petrel Geoscreening User Guide, Version 2017.1. Houston, Texas, USA: Schlumberger.
SLB. 2022. INTERSECT Technical Description, Version 2022.3. Houston, Texas, USA: Schlumberger.
20 SPE-212231-MS

Shook, G. M. and Mitchell, K. M. 2009. A Robust Measure of Heterogeneity for Ranking Earth Models: The F–PHI
Curve and Dynamic Lorenz Coefficient. Paper presented at the SPE Annual Technical Conference and Exhibition,
New Orleans, Louisiana, USA, 4–7 October. SPE-124625-MS. https://doi.org/10.2118/124625-MS.
Thiele, M.R., Batycky, R.P., Pöllitzer, S. et al. 2010. Polymer-Flood Modeling Using Streamlines. SPE Res Eval & Eng
13 (02): 313–322. https://doi.org/10.2118/115545-PA.
Watanabe, S., Natvig, J., and Tomin, P. In press. Strongly Coupled Prolongation in Multiscale Pressure Solver for
High-Contrast Heterogeneous Reservoir Simulation. Paper presented at the SPE Reservoir Simulation Conference,
Galveston, Texas, USA, 28-30 March. SPE-212229-MS.
Watanabe, S., Li, Z., Bratvedt, K. et al. 2016. A stable multi-phase nonlinear transport solver with hybrid upwind

Downloaded from http://onepetro.org/spersc/proceedings-pdf/23RSC/2-23RSC/D021S007R005/3092592/spe-212231-ms.pdf/1 by Stanford University user on 10 August 2023

discretization in multiscale reservoir simulator. Paper presented at the15th European Conference on the Mathematics
of Oil Recovery, Amsterdam, Netherlands, 29 August–1 September. https://doi.org/10.3997/2214-4609.201601852.
Watson, F., Krogstad, S., and Lie, K.-A. 2020. Flow diagnostics for ensemble methods. In 17th European
Conference on the Mathematics of Oil Recovery. European Association of Geoscientists and Engineers. https://
doi.org/10.3997/2214-4609.202035133.
Willis, B. and Sun, T. 2019. Relating depositional processes of river-dominated deltas to reservoir behavior using
computational stratigraphy. J Sediment Res 89 (12): 1250–1276. https://doi.org/10.2110/jsr.2019.59.
Willis, B. J., Kalla, S., and Sun, T. 2021a. Defining lithic patterns within river-dominated delta deposits for geostatistical
simulation. Pet Geosci 27(1): https://doi.org/10.1144/petgeo2019-105.
Willis, B. J., Sun, T., and Ainsworth, R. B. 2021b. Contrasting facies patterns between river-dominated and symmetrical
wave-dominated delta deposits. J Sediment Res 91 (3): 262–295. https://doi.org/10.2110/jsr.2020.131.