SPE-195858-MS

A New Optical Sensor Configuration Enables First Time Use of the

Mid-Infrared Optical Wavelength Region for Chemical Analysis During
Formation Tester Logging Operations

Ralph Piazza, Alexandre Vieira, and Luiz Alexandre Sacorague, Petrobras; Christopher Jones, Bin Dai, Jimmy
Price, Megan Pearl, and Helen Aguiar, Halliburton

Copyright 2019, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Annual Technical Conference and Exhibition held in Calgary, Alberta, Canada, 30 Sep - 2 October 2019.

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
This paper presents a new optical sensor configuration using a multivariate optical computation (MOC)
platform in order to enhance chemical analysis during formation tester logging operations. The platform
allows access up to the mid-infrared (λ ~ 3.5 microns) optical region, which has previously not been
accessible for in-situ real-time chemical measurements in a petroleum well environment. The new technique
has been used in the field for the analysis of carbon dioxide and synthetic drilling fluid components such
as olefins.
MOC is a technique that uses an integrated computational sensor to perform an analog dot product
regression calculation on spectroscopic data, optically, rather than by electronic digital means. Historically,
a dot product regression applied to spectroscopic data requires both a spectrometer and a digital computer in
order to perform a chemical analysis. MOC sensors require neither and because the key sensor component,
the multivariate optical element (MOE), is a stable temperature robust solid-state element, the technique
is well suited for downhole petroleum environments. A new dual-core configuration using two MOEs
designed to work in parallel enhances the sensitivity of the measurement enabling a mid-infrared analysis.
Spectroscopic measurements were performed on 32 base oils that were reconstituted to reservoir
compositions over a wide temperature and pressure range up to 350°F and 20,000 psi for a total of
12 combinations for each base oil. Live gas compositions (i.e. reservoir conditions) were achieved by
adding multiple methane, ethane, propane, and carbon dioxide charges to each base fluid. The reconstituted
petroleum fluids were further mixed with olefin-based synthetic drilling fluid (SDF). This rigorous
experimental design data therefore allowed for solid state MOEs to be designed to operate under the same
reservoir conditions. Laboratory validation showed measurement accuracy of +/-0.43 wt% for a range of
0 to 16 wt% CO2 and +/-0.4% from 0 to 10 wt% SDF. A wireline formation tester optical section was
modified with the MOC dual-core configuration to enable the mid infrared detection of both carbon dioxide
and olefins. This formation tester was then deployed in five wells collecting seven samples from various
locations. The downhole SDF and carbon dioxide measurements were subsequently found to be in good
2 SPE-195858-MS

agreement with laboratory analysis with field results for valid pumpouts showing an accuracy of 0.5 wt%
CO2 and 1.0 wt% olefins cross a range of 1.2 to 22 wt% CO2 and 1.4 to 9.7 wt% SDF.
Carbon dioxide is an important component of petroleum whose presence and concentration affects
completion options, surface facilities, and flow assurance, which thereby affects operational costs of
petroleum production. Olefins are a primary component of synthetic drilling fluid (SDF), although other
mid-infrared active components such as esters, ketones, alcohols, and amines also can be present. High
concentrations of SDF in openhole formation tester samples lower the quality of laboratory phase behavior
analysis and thereby force greater monetary risk in development of assets, especially when conducting
reservoir production simulations. Therefore, it is important to monitor both components during formation
tester sampling operations.

Introduction
Accurate compositional measurements of reservoir petroleum fluid are necessary for various exploration
and production activities, such as ensuring a well is safely drilled, confirming new discoveries, evaluating
the production potential and value of such discoveries, optimizing the capital investment for production,
and designing a field management system across multiple wells (Elshahawi et al. 2007; Dong et al. 2008).
Formation testers provide both down hole fluid analysis and also acquire samples from which laboratories
measure the composition of the reservoir fluids. However, formation tester sampling is a single opportunity
event without the realistic chance to acquire additional samples from subsequent runs, supposing the
laboratory determines those samples to be unfit. Therefore, decisions in real time must be made as: Where
to sample?, When to sample?, and How to sample? in order to optimize the sampling process. Downhole
chemical analysis helps to answer those questions. With regards to the question "When to sample?", it is
most important to ensure that those samples are of low contamination with respect to drilling fluid filtrate,
a byproduct of the drilling process. Otherwise, if samples are contaminated with the near wellbore filtrate
invasion, a laboratory analysis will not provide accurate information such as carbon dioxide concentration,
necessary to address the exploration and production activities. Downhole chemical analysis, as performed in
the visible and near infrared optical regions (NIR), can be used indirectly to estimate a contamination level
in the drilling fluid via asymptotic trend fitting, as first proposed in 1991 by Hammond. However, because
in the near infrared the drilling fluid filtrate spectroscopically is not distinguished from petroleum, no direct
measurement of drilling fluid filtrate has been possible. Instead, the contamination estimation relies on the
assumption that a pumpout trend to conform to idealized models, and the models often assume a simplified
reservoir structure and fabric with perfectly sealing mud cake and no active invasion. Unfortunately, both
the conformance assumptions and the model assumptions often prove false. However, the mid-infrared
(MIR) optical fingerprint region has great chemical specificity and as such, many synthetic components of
drilling fluid that are not present in natural petroleum are both identifiable and quantifiable from within
the petroleum matrix (Person and Zerbi 1982; Rabkin 1987). Therefore, MIR active synthetic drilling fluid
filtrate (SDF) contamination levels may be directly measured. Also, due to the chemical specificity of the
MIR response, the downhole chemical analysis can be improved yielding critical sampling information with
regards to reservoir compositional grading and reservoir compartmentalization ("Where to sample?"). Fluids
routinely compositionally grade within a reservoir such that the concentration of a given component varies as
a function of depth. Also, different components of the reservoir fluid routinely have different compositional
gradients, driven by equilibrium and/or disequilibrium processes, even within the same reservoir (Mullins
et al. 2015; Wang et al. 2015). Therefore, when sampling for a specific component such as carbon dioxide,
it is best to monitor that component directly during downhole sampling operations.
Characterization of the carbon dioxide profile within a reservoir is critical as the presence negatively
affects the value of that asset (Perez 2013; Popoola et al. 2013). Carbon dioxide lowers the BTU value of
natural gas, and it needs to be scrubbed from a fluid stream before shipping within a pipeline in order to
SPE-195858-MS 3

prevent lowering the value of the commingled production (Goldstein et al. 1999; Ulbig and Hoburg 2002;
Udina et al. 2008; Loubar et al. 2005). Also, petroleum production operational costs can be higher as a
result of carbon-dioxide-related scale, which chokes the production and requires remediation (Tjomsland
et al. 2001; Abouie et al. 2017; Tjomsland et al. 2008). Additionally, the presence of carbon dioxide can
corrode production systems, surface facilities, and pipelines not designed to handle corrosive concentrations
of carbon dioxide. The capital asset investment to construct corrosion-resistant wells, corrosion-resistant
production equipment, and surface scrubbing facilities is significantly higher compared to production
without carbon dioxide (Perez 2013). Corrosion alone costs the oil and gas industry USD 1.4 billion, with
approximately 60% of failures directly related to carbon dioxide (Perez 2013). Furthermore, lost production
related to corrosion failures costs the oil and gas industry further tens of billions of dollars in lost revenue
(Popoola et al. 2013). Concentrations greater than 2 bar partial pressure are considered highly corrosive and
from 0.2 to 2 bar, moderately corrosive (NACE 1958). In the presence of organic acids, 0.1 to 1 mM (0.2
bar to 2 bar) can also be highly corrosive (Bonis and Crolet 2005).
The mid-infrared (MIR) optical region is especially useful for spectroscopic chemical analysis (Herzberg
1945; Person and Zerbi 1982; Rabkin 1987; H. M. Hcise 1995; Schrader 1995; Bertie et al. 1995). The MIR
is known to the spectroscopist as the molecular fingerprint region for which the spectra uniquely identify
and quantify most chemicals. The MIR optical region extends from a frequency of 4000 cm-1 to 400 cm-1
or wavelengths 2.5 mm to 25 mm although sometimes the MIR is divided into two regions, at 1250 cm-1 (8
mm) with the shorter wavelength region belonging to the MIR and the longer wavelength region belonging
to the long wave infrared LWIR (Schrader 1995). Beyond the MIR exists the far infrared (FIR), at shorter
wavelengths exists the near infrared (NIR). Terminology varies somewhat by discipline with, for instance,
telecommunications breaking the NIR into the shortwave infrared (SWIR) and the NIR which, based on
detector types, occurs from 0.7 mm to 2.0 mm and the mid-infrared existing beyond 2.0 mm. The current
paper uses the chemistry infrared spectroscopy terminology of 2.5 mm to 25 mm since the purpose of this
study is chemical analysis (Schrader 1995). Conventional downhole access to wavelengths above 2.0 in
downhole settings has historically been difficult, due to the limitation of suitable high-temperature detectors.
Fundamentally, semiconductor photodiodes operate with decreasing efficiency at high temperatures. The
detectors also have decreased sensitivity toward higher wavelengths so that the composite effect is very
poor performance at mid-infrared wavelengths. Therefore, until recently, chemical analysis has been limited
from the visible at 0.4 mm to the NIR region of little more than 2.0 mm for formation testers applications.
In this study, we present a unique sensor configuration that allows chemical analysis above 2.0 μm and in
the MIR up to 3.5 μm for the direct chemical analysis of synthetic drilling fluid filtrate components and
carbon dioxide.
The estimation of contamination monitoring is most difficult for miscible pumpouts i.e. oil-based mud
(OBM) filtrate with hydrocarbon formation fluids. Standard sensors such as density, resistivity, capacitance,
acoustic, bubble point, compressibility, visible, and near infrared (NIR) can provide good contrast between
the filtrate and formation fluid for trend fitting, but none can observe a deterministic set of good markers
for the measurement filtrate in formation fluid. A good marker induces a known sensor response for each
endmember, such that the contamination level of filtrate in the formation fluid may be directly calculated
based on the sensor readout. The best markers have a strong response for the primary fluid (drilling fluid
filtrate for contamination monitoring) fluid and an absent response for the complimentary fluid (formation
fluid). For the standard sensors, there is no component marker which has a response in filtrate that does
not also have a response in formation fluid, nor is the response for any component a priori known. The
constituents of OBM may be categorized as either natural or synthetic. Natural components are those
components which may be commonly found within a petroleum reservoir, whereas synthetic components
are those components which would not be found commonly and naturally within a petroleum reservoir.
Therefore, components such as saturates or aromatics would be termed as natural whether from petroleum
or the mud system. Components such as olefins, esters, ketones, alcohols, and amines are not natural
4 SPE-195858-MS

because they are not naturally occurring in significant abundance within petroleum fluids. Essentially, these
functional groups are not thermodynamically stable in geologic time. Ideally, these compounds would be
excellent markers for synthetic drilling fluid (SDF) filtrate were they to be distinguished by the standard
sensors. In the mid-infrared (MIR) wavelength region of the optical spectrum from 2500 nm to 3500
nm, these compounds have unique signatures. As an example, Fig. 1 shows superimposed a transmittance
spectrum of the synthetic component 1-octene, an olefin, and its counterpart, the natural saturate compound
octane. The olefin Carbon-Carbon double bond influences the spectrum of 1-octene so that the spectra are
distinct. In fact, in the MIR fingerprint region, most compounds are distinct even if structurally very similar.

Figure 1—Adapted from the NIST web book (23), shown is octane a saturate compound overlain on octane an olefin. The
wavenumber scale is the inverse of wavelength and as such the figure is plotted on a reverse axis for easier comparison to
wavelength plots in this paper. Also note that the y scale is transmission so that peaks are negatively oriented compared
to absorbance which are positively orientated. Note the olefin CH bend, CC stretch, and CH stretch shown that are distinct
from the saturate counterpart. The CH stretch occurs in the mid IR and can serve as a marker for this synthetic compound.
Note that the position of the CH stretch is highly variable depending on the nature of the compound and can change in
intensity and position readily dependent on both the molecular structure of the olefin and the physio chemical environment.

In addition to olefins, compounds such as esters, ketones, alcohols, and amines are optically active in
the MIR from 2.5 μm to 3.5 μm as shown in Fig. 2. In this optical region, the typical components of
petroleum may have elevated petroleum baselines but do not have molecular signature activity in the form
of vibrational peaks. Therefore, this region is potentially useful as a marker region for SDF so long as
the baseline interference from petroleum can be subtracted. Although not shown, these 2.5 μm to 3.5 μm
MIR active compounds are not distinct from typical petroleum compounds in the visible or NIR from 0.40
μm to 2.0 μm. In addition to these common components of drilling fluid formulations, there are many
low concentration compounds such as but not limited to viscosifiers, emulsifiers, corrosion inhibitors, loss
control gelling agents, tar mat control agents, and hydrogen sulfide scavengers other than amines that may
have MIR activity from 2.5 μm to 3.5 μm. Generally, the component need not be identified to be a good
marker. However, because the formulation of synthetic drilling fluid can change substantially from basin
to basin, the MIR fingerprint of SDF likewise changes. Therefore, it is important to narrow this study to a
specific application area such as the pre-salt of Brazil where drilling fluids contain a common set of additives
as dictated by the local drilling requirements.
SPE-195858-MS 5

Figure 2—Shown are examples of MIR active compounds typically found in synthetic drilling fluids. The examples include
alcohols (n-butanol), ketones (2-Pentanone), amines (Tert-butylamine), and esters (Propyl Ester) as adapted from the
NIST Web Book (23). Note the activity between 4000cm-1 and 2860cm-1 (2.5 μm and 3.5 μm), the marker region for SDF.

Theory
Attenuation of an optical signal increases exponentially with concentration "C" as shown in Eq. 1 (H.
M. Hcise 1995). Also, as an initial light intensity, I0, impinges a sample, it is exponentially attenuated as
a function of path length through the sample, l, and by a material- and wavelength-specific attenuation
constant, ε. Therefore spectroscopy can be used to derive the concentration of a substance by measuring
the resultant optical intensity, I, for a characterized system given a known ε, using fractional transmittance
T. Used here, (decadic) absorbance, A, is the negative log10 of fractional transmittance and is linearly
related to the concentration of a chemical species, as shown in Eq. 2, and is used to construct a linear
relationship between a transformed signal and sample analyte concentrations (Chalmers and Griffiths 2002).
Chemical interference is an additive signal not caused by the analyte but rather a species that responds to
the measurement technique (Martens and Naes 1992; Massart et al. 1998). Often, all chemicals in a mixture
will obey the Beer's law relationship, including those of chemical interferences. Therefore, to mitigate
interference, Beer's law may be constructed in matrix form as shown in Eq. 3, where is the column
vector of absorption as a function of wavelength for species i, at a given concentration and is the column
vector multi-wavelength attenuation constant, known as the molecular fingerprint. Any spectrum of the set
of spectra, S, is the sum of absorbance from all optically active species including interference species, i, and
X, the response matrix, as a set of S as rows of X. A calibration response matrix is obtained from samples
of linearly independent concentrations of optically active species and a set of reference analyte properties,
which are usually concentrations, the vector y. Using a solver such as linear least squares, a regression
vector B may be calculated by Eq. 4 for a set of corresponding analyte reference values, the vector y, and
then projected against a new spectral set to obtain a predicted analyte concentration estimate, the vector ,
using Eq. 5 (Massart et al. 1998; Thomas 1994; Beebe and Kowalski 1991).
Eq. 1

Eq. 2
Eq. 3
Eq. 4
6 SPE-195858-MS

Eq. 5
A direct measurement of analyte concentration by optical spectroscopy, relevant to this paper, is shown
in in Fig. 3 and Fig. 4. In Fig. 3, the spectra of two live oils, two dead oils, carbon dioxide, a natural mineral
oil drilling fluid base, and an olefin-based synthetic drilling fluid are shown. The olefin SDF is synthetic
base oil composed mostly of olefins. Note the strong absorbance in the MIR of two peaks with intensities of
2.4 and 2.6 absorbance units, respectively, for the olefin SDF. No other fluid shown has overlapping peaks,
including the four petroleum fluids. The base mineral oil shown in Fig. 3, a primary component of natural
OBM, shows no MIR activity. Also note that four randomly chosen petroleum fluid mixtures, the live oils
648 and 633 as well as the dead oils 8 and 25, do not show overlapping peaks. However, it should also be
noted that the baseline is not zero for the other hydrocarbon fluids, and therefore, this baseline interference
will have to be corrected. Also note that carbon dioxide has strong MIR activity, but not overlapping with
the SDF.

Figure 3—The MIR spectra of Olefin SDF against fluids including a natural mineral oil-based drilling fluid base
oil, carbon dioxide, and four petroleum fluids. The petroleum fluids and carbon dioxide do not have peak activity
overlapping with the olefinic Olefin SDF. Although OBM drilling fluids often contain many trace compounds
in the filtrate, these spectra of Olefin SDF and mineral oil base are collected on the unblended pure base oil.
SPE-195858-MS 7

Figure 4—Spectra of three recombination compositions of Dead Oil 28 collected at different temperature
and pressure combinations for each recombination. Charge zero contains the initial Dead Oil 37, Charge
1 contains 33.2% the Olefin SDF of Fig. 3 with the dead oil as the liquid portion, and approximately 150
SCF/BBL, with a gas composition of 33% carbon dioxide in a balance of methane. Charge 2 contains
approximately 300 SCF/BBL of the same mixture of 33% carbon dioxide in a balance of methane.

Fig. 4 shows the MIR activity of olefin SDF when added to Dead Oil 37. Dead Oil 37 is significantly
darker than the dead and live oils of Fig. 3. A gas charge of 33% CO2 in a balance of methane to push
the SDF base oil fluid into a PVT cell at 3000 psi pressure. The first charge was 150 SCF/BBL total gas-
to-oil ratio (GOR). A second 150 SCF/BBL gas only charge was then added. Note the carbon dioxide
peak at 4300 nm which doubles upon the second addition of gas. Also note that the 1.680 m methane
shoulder and the 2.300 m methane shoulder increase with each gas charge. All fluid mixtures were run with
a combination of three temperatures, (150°F, 200°F, and 250°F) and four pressures (3000 psi, 6000 psi,
9000 psi, and 12,000 psi) for a total set of 12 conditions for each run, leading to the splitting of the single
color curves. Note that the Olefin SDF peaks in the MIR increase substantially with the first charge but not
the second. Using the midpoint of the blue spectra as a baseline as a correction method for that baseline,
the increase in absorbance for the second charge at each peak due to Olefin SDF addition may be measured
as 0.8 and 0.9 absorbance units. Using the pure Olefin SDF response from Fig. 3, optical spectroscopy
estimates a 0.8/2.4*100=33.3% and 0.9/2.6*100=34.6% contamination which is very close to the added
33.2% contamination. This simple demonstration shows the power of MIR optical spectroscopy, but also
the necessity of designing a multivariate regression vector in order to deconvolute the baseline interference
petroleum samples.

Multivariate calibration and regression

The processes of multivariate calibration attempt to produce a correlative model between a set of multiple
observations for each member of a population set and a matched set of reference characteristics for
each member. The purpose is to provide an estimate of the same characteristic for new members of the
population with that model (Massart et al. 1998). For chemical measurement applications, the observations
are generally single-channel or multi-channel signal responses of one or more analytical instruments or
analytical methods, and the members of the population are material samples. The characteristic of the
material is generally the concentration of a particular component or group of components in a mixture
8 SPE-195858-MS

and is termed the analyte. A calibration set is described as a response matrix of a set of signal channel
responses, each occupying a column of the matrix for a set of samples, each occupying a row, and an
array of reference analyte concentrations for each sample row (Beebe and Kowalski 1991). Although many
observation techniques may be correlated with many characteristics of many different population types,
discussion herein will be limited to chemical measurement applications. With multivariate calibration, so
long as the interfering signal is expressed with linear independence in the sample set with respect to the
analyte and has orthogonal structure with respect to the analyte in the response columns, the regression
vector will be constructed as orthogonal to the interference. Implicitly, the signal due to interference is
subtracted by coefficients leaving the net signal (Lorber et al. 1997; Nadler and Coifman 2005). Fig. 5
shows an example of a multivariate regression vector. The vector weights are plotted as a function for the
dependent variables for which the regression is performed. The variables need not be of the same dimension
or continuous; however, as is common for spectroscopic multivariate regression vectors, the variables will
generally be of the same dimension and generally continuous. For this reason, the regression vector forms
a smooth function, continuous in first and second derivatives. This is due to the highly correlative nature
of the spectra from which the regression vector is derived. Note that the regression vector has both positive
and negative coefficients. For absorbance data, positive coefficients detect the analyte, whereas negative
coefficients subtract interference. In this manner, a multivariate regression vector may subtract interference
from other chemical species in solution including those that cause the nonzero baseline of the MIR region
in petroleum.

Figure 5—An example multivariate regression vector. This multivariate regression vector was designed for multiple
batches of a target SDF base oil and field filtrate with the goal of encompassing the span of variation with respect
to this formulation. The x axis is a variable index which is proportional to wavelength with an offset, and the y
axis is the regression coefficient. The greater the magnitude of the regression coefficient, the more important
the variable position for the measurement of the target analyte. Generally for an absorption spectrum, positive
wavelengths correlate with the analyte, whereas negative wavelengths subtract interference such as the oil
baseline interference. The regression vector shown here was developed using the partial least squares algorithm.

Highly correlative variables of spectral data and the large number of variables as compared to samples
often make the matrix inversion of the raw response matrix ill conditioned with respect to Eq. 4. A principal
component (PC) rotation of the response matrix can produce a much improved matrix for inversion and
SPE-195858-MS 9

linear model construction (Beebe and Kowalski 1991). A PC scores matrix is constructed as a linear
combination of the original response matrix by projecting variables of each sample row onto a new set of
orthogonal axes called PCs. The PC axes are constructed based the internal structure of variance of the
response matrix such that the first PC describes the greatest degree of variation across the dataset, with each
subsequent PC describing the greatest remaining variation but maintaining orthogonality to all previous
PCs. Because each successive PC describes less information, higher order PCs typically capture greater
contributions of noise from the original response matrix (Jackson 2005). The information content of the
original response matrix may then be represented with a reduced dimensionality of orthogonal new variable
scores which is inherently better conditioned for inversion. A linear dot product regression vector in the
original response variable space may be constructed from the PC scores coefficients and the PC rotational
"Eigen" vectors known (Beebe and Kowalski 1991).
With principal component regression (PCR), the PC dimensionality structure is arbitrary with respect
to the analyte concentration array. The regression vector is often constructed as a high order linear model
with a large number of PC scores variables. Partial least squares (PLS) is a calibration technique which
overcomes this limitation of PCR. PLS also designs a linear dot product regression vector from the reduced
dimensionality orthogonal scores matrix of a rotated response matrix. The PLS dimensions, however, are
rotated to capture the maximum variation of the reference analyte concentrations (Beebe and Kowalski
1991). Although the PCR and PLS regression vectors are different rotations of the response matrix, the
performance is often very similar even over a broad range of conditions and sample sets (Thomas 1994). In
fact, there are many algorithm options, both linear and nonlinear, for a linear dot product regression vector
construction. There is no inherent constraint that regression vector shape be similar when constructed by
different algorithms even when constructed with the same dataset. None the less, regression performance
is often sufficiently similar for at least some conditions even with different shapes. At a fundamental level
regression vectors access a complex rotation or linear combination of the analyte spectrum and interferences
spectra (Nadler and Coifman 2005).

Multivariate optical computing

Multivariate optical computing (MOC) is a compressive sensing multivariate linear regression technique
that performs an analog dot product calculation, in the optical domain, between a linear regression vector
and the inherent optical intensity spectrum of light emanating from a sample. The regression vector
shape is encoded as a transmission pattern for one or more optical elements. MOC most typically uses
an MOE, which is constructed as an interference filter. Laminated, thin film layers of two materials
with different refractive indexes (Ri) are deposited onto a substrate forming an interference pattern. The
interference pattern reflects a portion of the light as a function of wavelength. The goal of calibration
is to encode a regression vector as a transmission profile using a thin film stack design. A specifically
designed thin layer stack can sufficiently match a regression vector shape. The MOE regression vector
can either be predetermined—for instance, by PLS—or directly designed to a calibration set (Soyemi et
al. 2002). As spectral light passes through the MOE, the Hadamard vector product, a wavelength-by-
wavelength multiplication, of the sample intensity spectrum with the MOE transmission profile naturally
occurs. The Hadamard vector product is followed by a summation of all light wavelengths by a detector,
thereby completing the dot product. Therefore, the resultant detector signal is proportional to the analyte
concentration for which the regression vector was designed plus a sample-dependent offset (Soyemi et al.
2002). The sample-dependent offset can be subtracted either by use of light reflected from the element, a
reference spectral signal, or a secondary optical element (Soyemi et al. 2002; Nelson et al. 1998). MOEs
operate on the intensity spectrum emanating from a sample, as described in Eq. 1, not the linear absorbance
form of the Beer-Lambert law from Eq. 2. Therefore, the MOE does not strictly operate on signals linear
with the analyte concentration. However, it has been shown that higher-order linear models can model
10 SPE-195858-MS

nonlinearity (Nelson et al. 1998; Næs and Isaksson 1992; Liu et al. 2009; Gemperline et al. 1991; Centner
and Massart 1988). As such, MOEs operating on single-beam intensity transmittance can still provide a
reasonable measurement, as long as they can reproduce a higher-order regression vector.
Because MOEs are constructed similar to that of interference filters, the MOC system is robust and shares
the same strong form and construction with the capability of performing without failure under a wide range
of conditions as that of a filter spectrometer. Conventional MOC sensors based on the single core MOE
technique have an advantage of sensitivity, compared to dispersive techniques, since all wavelengths of light
strike the detector simultaneously and thereby provide greater optical energy throughput. A conventional
equal band dispersive filter spectrometer only allows a fraction of optical energy of no greater than 1/n2 to
reach the detector. Therefore a 20 channel filter spectrometer typically would allow 1/400 of the sensitivity
of a conventional MOE system. The MOE also provides another advantage over filter spectrometer systems.
The regression vector may be encoded as the transmission function with nanometer wavelength resolution
on par with that of a laboratory spectrometer. In fact, it is often not the resolution of the transmission function
that is the limiting factor in regression vector performance, but the resolution of the spectral library from
which the regression vector is designed. If then, for instance, the regression vector was designed from a
FTIR instrument, then the MOC system should perform as well as if an unknown sample were analyzed by
that FTIR system with the digital regression vector applied. This, of course, only compares the instrument,
and it should be noted that often laboratory methods have the additional advantage of sample preparation
not afforded to downhole instruments. Also, whereas an MOC sensor may perform as well as the laboratory
instrument from which the MOE was designed, there are many types and grades of laboratory instruments
both spectroscopic and non-spectroscopic. The MOC sensor may only be compared to the instrument from
which the MOE was designed, and not, for instance, a gas chromatograph which is known for superior
accuracy over most spectroscopic instruments or a mass spectrometer which is also known for superior
sensitivity over most spectroscopic instruments.
This work makes use of the dual-core MOE application illustrated in Fig. 6. A multivariate regression
vector is designed in order to determine an analyte, i.e. carbon dioxide or SDF, using a spectral library. The
spectral library must contain mixtures of the analyte in the presence of potential interferences for which the
regression will be applied. It is not possible to design against unobserved interferences, although technically
those interferences need not be identified and named. After the regression vector is designed, the regression
vector is encoded into an MOE as a transmission function. For the dual-core form, the regression vector is
divided into two transmission functions such that a weighted difference, i.e. α1 and α2, of those transmission
functions is the regression vector, i.e. Eq. 6. Because the regression vector is linear and separable, Eq. 7,
the dot product may be separately applied to each of the transmission functions of the MOE, Eq. 8, and the
weights α1 and α2 applied directly to the detector signals, Eq. 9, providing a direct estimate of the analyte,
Eq. 10.
SPE-195858-MS 11

Figure 6—Diagram of the new dual-core multivariate optical computing design and application process.
Reference sample solutions are measured using a conventional laboratory optical instrument to provide a
spectral library from which a regression vector is designed. The regression vector is encoded as the linear
combination of two optical elements’ transmission functions. As light passes through each optical element
and strikes each detector, the output voltage is proportional to the dot product of the sample spectrum and
the optical element. Because the weighted difference of each optical element transmission function is the
regression vector, Rv, the weighted difference between the detector output is the analyte concentration.

Eq. 6
Eq. 7
Eq. 8
Eq. 9
Eq. 10
The dual-core form has significant advantages over previous single core designs. First, the shape of
the regression vector may take more complex forms for a smaller number of layers. The complex shape
advantage allows more complex interference correction. The smaller number of laminated layers has an
advantage in allowing more optical energy to reach the detector. The inherent sensitivity of an MOE is
exactly related to the total optical energy reaching the detector. Now, the dual-core designs are generally
of much increased sensitivity generally by a factor of 4. Since the sensitivity of the composite MOC
sensor is the product of the MOE and detector sensitivity, the detector may be much less sensitive than a
filter spectrometer and still achieve good performance. This allows detectors in an MOC system, such as
thermopiles which are 100 times less sensitive than a semiconductor photodiode but that operate in the mid-
infrared optical region, with overall MOC sensitivity still generally greater than that for a corresponding
filter spectrometer. It should be noted that in addition to instrument sensitivity, the overall method sensitivity
considers the inherent analyte sensitivity and any sample preparation such as pre-concentration.

MOE Design
For this study, dead oil samples were used as a base fluid to be reconstituted using a small-volume pressure,
volume, temperature, x-composition (PVTX) system outfitted with a Fourier transform infrared (FTIR)
spectrometer. This system and the method of use has been described previously (Jones et al. 2017). For
this study 743 spectra were collected over a wavelength range including 2.5 to 3.5 μm. Carbon dioxide
was added up to approximately 17 wt% and SDF up to about 30%. A total of ten synthetic-based oils or
12 SPE-195858-MS

synthetic drilling fluid filtrates that are commonly used in Brazil pre-salt wells were obtained, six from the
operator, and four from the service company. The blended components were all sourced locally from within
Brazil, and although the chemical composition is not specified, were all olefin-based. The samples include
three base oils from the service company and one field filtrate, and all field filtrates from the operator.
Additionally, 16 pre-salt dead oils were received with the initial live oil composition. The oils were selected
specifically for diversity to span the range of compositional properties. SDF randomly paired with dead
oil and was mixed with the dead oils. These spectra were used for the design of the carbon dioxide and
SDF MOEs. Outlier detection and resampling was used for this final selection of spectra in both calibration
sets (Valderrama et al. 2007). The base fluids used to construct the carbon dioxide calibration set contained
medium oils of <22.5 to 30 API weight and light oils of <30 to 40 API weight, which had a gas-to-oil ratio
(GOR) less than 2223 scf/bbl. The calibration set contained 31 medium oil samples as base fluids that were
reconstituted to different live oil compositions targeted on the reservoir composition from which the base
oils were sampled. Because there is little optical activity of methane, ethane, and propane relative to that
of carbon dioxide or SDF, only methane was added as the dominate hydrocarbon gas component (Pedersen
and Peter 2007; Dandekar 2006) to attain matrix conditions for the in-situ reservoir fluid. Table 1 shows the
range of compositions for each of the calibration sets. The spectra were collected at 65.5, 93.3, and 121.1°C
and at 20.684, 41.369, 62.053, and 82.727 MPa, which spans the pressure and temperature range for the
intended MOC system use.

Table 1—The composition range of recombined components into petroleum fluid base oils.
The GOR shows the relative concentration of recombined fluids to the petroleum base.

Property Carbon Dioxide Range

Min Max

Temp. (°C) 65.5 121.1

Pressure (MPa) 20.684 82.727

CO2 (g/cc) 0 0.1060

Methane (g/cc) 0 0.1556

Ethane (g/cc) 0 0

Propane (g/cc) 0 0

Saturates (g/cc) 0.2618 0.5963

Aromatics (g/cc) 0.0833 0.2294

Resins (g/cc) 0.0065 0.2919

Asphaltenes (g/cc) 0.003 0.0257

Synthetic Drilling Fluid (g/cc) 0 0.1779

GOR (scf/bbl) 0 2223

A design consideration for the MOE is the MOC sensor detector sensitivity. If the MOE is directly
designed to the FTIR reference data, then the MOE will be optimized to the inherent signal-to-noise ratio
(SNR) of the laboratory system. This can result in accurate regression vectors that fail in the presence of field
sensor noise and drift. The MOC sensor, which uses a thermopile detector, does not have the SNR of the
laboratory FTIR liquid-nitrogen-cooled mercury cadmium telluride (MCT) detector. The digital reference
spectra Xt used in the optimization have a larger SNR than is inherent to the MOC sensor. The magnitude
to which analyte prediction is degraded by noise-related error is inversely proportional to the sensitivity of
the regression vector. The MOC sensitivity is also related to the quantity of light that reaches the detector.
This quantity of light is influenced in part by the MOE design. To design an MOE regression vector for
SPE-195858-MS 13

an actual MOC sensor, a merit function needs to contain a penalty related to the evolution low-sensitive
designs as well as the accuracy of the regression vector. Eq. Eq. 12 calculates the error of an individual
spectrum prediction as the deviation from the reference value yi. Eq. Eq. 13 calculates a detector intensity
of a spectral signal Xi through a transmission function equal to either Ta(t) or Tb(t). Because there is also
an error associated with each detector signal, Eq. Eq. 14 calculates this error for each spectral signal as the
sum of squares of the error for Detector A, DA, and Detector B, DB. Eq. Eq. 14 amplifies this noise by the
inverse of each of the dual-core MOE's contribution to sensitivity, α and β using the MOC characterized
SNR. The final mean squared error is calculated as the sum of the squares for the noise and accuracy
contributions of every design with Eq. Eq. 15. Essentially, the mean square error (MSE) objective function
is derived to optimize a pair of MOE cores such that the error associated with accuracy and sensitivity are
both minimized. MSE scales linearly with SEC accuracy, and typically the minima for MSE will also be
the most accurate design as well.
The SNR is empirically determined as the lower 95% confidence limit, a value of 500 for the NIR and
200 for the MIR, as measured from 85 MOC sensors. The SNR is used to constrain the design solutions to
meet acceptable real-world sensor sensitivity criteria. The binary scheme of weighting devised for Eqs. Eq.
11, Eq. 12, Eq. 14, and Eq. 15 could be more generally replaced with a sum of squares response for SEC
accuracy and noise. However, because crude oils are of significant optical density variation, this scheme
was derived to provide weight to the source of errors as a winner-takes-all approach for a particular oil
type. Designing an MOE with the consideration of noise allows easy identification of the most optimal
design and potentially helps evolve designs that otherwise would not if noise was not considered during
the evolutionary design process.
Eq. 11
Eq. 12
Eq. 13
Eq. 14

Eq. 15

Carbon dioxide has predominant spectral features in both the NIR (1.3 to 2.5 μm) region as well as the
MIR region (2.5 to 3.5 μm). The FTIR transmittance PLS regression for both the separate regions provides
a similar prediction of 4.5% over the full range from 0 to 0.1060 g/cc, which is approximately 0.6 wt% for
an 0.8 g/cc oil. The MIR PLS calibration and regression vector are shown in
Fig. 7. However, the NIR region regression vector is more complex and requires latent variables (seven)
compared to the MIR regression vector, which requires four latent variables. This can be expected as a result
of the larger amount of hydrocarbon interference in the NIR (Mueller et al. 2006). Because the regression
vector in the NIR must be orthogonal to more interference, the effective signal is reduced. A net analyte
signal (NAS) analysis was performed on both regions independently, and the results are illustrated in
14 SPE-195858-MS

Figure 7—PLS model prediction plot identifying a theoretical PLS calibration

error of 0.00478 g/cc (a), and corresponding four-PC regression vector (b).

Fig. 8. The sensitivity of a generic multivariate regression vector to a calibration set can be assessed
using NAS theory. Sensitivity can be assessed across multiple instruments and/or multiple wavelength
ranges (Lorber et al. 1997; Nadler and Coifman 2005; Faber 1998; Bro and Anderson 2003). Therefore, the
sensitivity of a calibration for carbon dioxide in the NIR can be directly compared to that for the MIR. The
norm of the NAS (NNAS) for each regression vector is plotted versus the carbon dioxide concentration,
with the slope being indicative of the overall sensitivity. From the Fig. 8, the MIR region offers a 27 ×
stronger signal response compared to the NIR region despite the overall light throughput being weaker in
the MIR region resulting from the black body lamp emission profile. As previously stated, SDF has no
unique optical signal in the NIR and therefore no sensitivity in the presence of petroleum. As such, a similar
analysis is not necessary to validate the use of the MIR for SDF detection.

Figure 8—Figure, norm of NAS vs. measured carbon dioxide concentration for the NIR (a) and MIR (b) spectral
regions. The range of the MIR NAS indicates ~27 × stronger sensitivity compared to the NIR spectral region.
SPE-195858-MS 15

Using leave-one-out cross-validation, the PLS analysis identifies a root mean square (RMS) error of
cross-validation (RMSECV) of 0.0048 g/cc (or 4.5% of the range) using a four-level model. The good cross-
validation result suggests the dataset is of sufficient rank for modeling. This helps establish a performance
limit for the dual-core design process and will help evaluate the design results with reference to a global
optimum. Also using a leave-one-out cross validation, the PLS analysis identifies a 0.0153 g/cc RMS error
for SDF corresponding to 8.5% of full range or 2.2% limit of calibration. Whereas the limit of detection
is the lowest detectable amount, the limit of calibration is the lowest quantifiable amount and is usually
substantially higher than the limit of detection. This performance was at the limit of desired specification,
so a second regression vector was designed over a limited range of approximately 11% wt% SDF, with
performance of 7.7% accuracy as determined by leave-one-out cross validation. This corresponded to
+/-0.83 wt% limit of calibration. The accuracy over the limited range was improved, and further, the lower
range was believed to be more important in determination of SDF contamination. Therefore, a hybrid
approach was undertaken in the MOE design for which lower concentrations were weighted relative to
higher concentrations of drilling fluid filtrate.
To design a set of MOE channels for an MOC sensor, the spectra are transformed to the single-beam
spectra that the detector of the MOC sensor would observe. To accomplish this, the spectra are normalized
to fractional transmittance and then convolved with the radiometric contributions of optical components
along the optical path of the sensor, including the lamp, bandpass, windows, and detector. The absorbance
spectra used for calibration are shown in Fig. 9. The activity at 2.650 and 2.900 μm is caused by the carbon
dioxide combination (v1 + v3) band (Kiehl et al. 1985). From 2.5 to 3.5 μm, there is significant SDF activity.

Figure 9—Absorption spectra of reconstituted oils which includes carbon dioxide and SDF.

The designed carbon dioxide dual-core regression vector is similar in performance to that of the PLS
regression vector, with a relative accuracy of 5.4% compared to 4.5%, of full range, respectively. The
designed SDF dual-core regression vector is similar in performance to that of the PLS regression vector,
with a relative accuracy of 10.7 wt% for the full range and 7.7 wt% over the limited range compared to
the MOE which is 8.6% for the lower weighted range. Although the lower weighted range had a maximum
concentration of 25 wt% SDF, the 95% upper limit was 14.5 wt% so that the estimated limit of calibration
is 1.25 wt%.
16 SPE-195858-MS

Field Tests
For validation, the dual-core carbon dioxide and SDF MOEs are placed into an MOC system similar to the
one shown in Fig.. The MOC sensor for which the MOE is designed has been described previously (Jones et
al. 2012; Jones et al. 2016; Jones et al. 2013; Jones et al. 2015; Eriksen et al. 2013). Briefly, the sensor uses
a tungsten light source providing a nearly black body profile focused through the system by a gold back
parabolic reflector. Light traverses two sapphire windows spaced at 1mm, and then is split into two paths by a
beam splitter. The first path passes through an outer concentric circle of MOE channels as shown in Fig. and
simultaneously, the reflected beam split light is turned through an inner concentric circle of MOE channels.
The complimentary MOE pairs for each of the carbon dioxide and SDF are placed in the corresponding radial
position of concentric circles such that the beam split light passes through each of the dual cores for a given
analyte simultaneously. After passing through the inner and outer MOE, the light strikes a corresponding
thermopile detector. The thermopile detectors are wired to subtract the reference before the detector chip
amplifier boosts and digitizes the signal. A motor rotates the carousel at 6 RPM and takes a 50ms snapshot
of the detector signal when the MOE is in full uneclipsed view of the detector as determined by an angular
resolver. The signal from each snapshot is used to calculate an analyte concentration with the use of Eq.
10. Because the optical throughput, sensitivity, and exact α1 and β2 terms are slightly different MOC sensor
to MOC sensor, the MOC sensor is calibrated. The sensor used for this field test was calibration with 14
fluids, and then was placed in a wireline formation tester (WFT), which is used to acquire reservoir samples
during the process of formation pumpout.

Figure 10—Multivariate optical computing sensor. The fully assembled sensor is shown to the left with
the carousel containing complimentary inner and outer concentric circles of dual cores shown on the left.

For the initial test, the WFT was lowered to a location within the petroleum well having a pressure of
67.07 Mpa and temperature of 89.9°C. The calibration conditions for the MOC sensor in the laboratory
differed slightly at 24.4°C (6.7% absolute temperature deviation) and 5.02 Mpa (7.5% absolute pressure
deviation) from the field conditions. In principal, the MOC sensor is designed to be temperature robust,
but it was intended to provide close conditions for calibration. Fluid is withdrawn from the formation
through a hydraulically sealed probe using a mechanical pump in an effort to clean the near-wellbore fluid
of drilling fluid filtrate contamination, a byproduct of the drilling process. Throughout the pumpout, the
contamination is reduced and the concentration of drilling fluid filtrate increases. After 228 minutes, a
sample of the formation fluid is captured. Fig. shows the measured methane, carbon dioxide, and GOR
throughout the pumpout as determined by the MOC sensor without averaging. The captured sample was sent
to a laboratory for analysis by gas chromatography and reference bulk properties, also shown in Fig.. The
final MOC sensor readings are presented as an average for the last 50 sensor measurement points acquired
SPE-195858-MS 17

over 500 seconds. Pseudo normalization (Nelson et al. 1998)is used to compensate for any drifting of the
MOC system. The in-situ measured methane concentration matched the laboratory value within 2%, and
the measured carbon dioxide value matched the laboratory value within 20%. The GOR is a measure of the
total hydrocarbon and carbon dioxide gas divided by the amount of oil. The MOC sensor also contains a
previously described (Jones et al. 2012; Jones et al. 2016; Ellis and Singer 2008) single-core hydrocarbon
GOR MOE, which, although not the subject of this study, is useful for reference in the discussion. From the
laboratory analysis, the total gas of the GOR is composed of 56.23% methane, 23.3% carbon dioxide, with
the balance being hydrocarbon components of ethane through pentane. The sampled fluid contains residual
drilling fluid filtrate equal to 3.22 wt%, based on a gas chromatograph analysis, as compared to that of a pure
filtrate sample (Hy-Billiot et al. 2002), which was collected at the well site. This compares well to the value
of 3.9% as determined by the SDF MOE. Additionally, the pH of the acquired filtrate sample is measured
as 8.9. Three additional field tests took place with a total of seven zones sampled for comparison. A fifth
field test attempt was excluded from analysis due to difficulties in maintaining a seal with the formation
tester during pumpout such that whole mud, as opposed to mud filtrate, saturated the flow line.

Figure 11—In-situ field test of the dual-core methane (C1) and carbon dioxide (CO2) MOC sensor at 88.9°C and
67.07 Mpa. The left axis is carbon dioxide and methane components in g/cc, and the right axis is GOR in scf/bbl.

Results and Discussion

For the first field test, a very interesting pattern is observed. Methane differed by 0.07% absolute
(2% relative), and carbon dioxide differed by 1.8% absolute (20% relative). Both methane and carbon
dioxide values, as determined by laboratory analysis, are within the range of each calibration. Real-time
determination of methane and carbon dioxide is highly valuable. Carbon dioxide levels were determined
to be significantly higher than that at which carbon dioxide becomes a corrosion problem. For methane,
the concentration determined is sufficient to provide sampling information. During the sampling pumpout
process, drilling fluid filtrate, which is present in the near-wellbore region, is flushed through the WFT by
mechanical pumping action and expelled into the wellbore. As the pumpout continues, the composition of
the fluid in the WFT grades from drilling fluid filtrate to formation fluid. Drilling fluid filtrate contains no
dissolved gas, whereas the formation fluid usually has some dissolved gas. In Fig., it can clearly be observed
that the level of gas to oil increases throughout the pumpout asymptotically to a constant value of 1412 scf/
bbl. Carbon dioxide accounts for 23.3% of the GOR, or 324 scf/bbl. Although the MOC-measured value of
carbon dioxide is low at 20% or 65 scf/bbl, this is only 4.6% of the total GOR and less than the measurement
accuracy of GOR. Nearly 80% of the carbon dioxide increase occurs after 91 minutes of pumpout, a curious
18 SPE-195858-MS

observation indeed. The GOR after 91 minutes increases from 1125 to 1412 scf/bbl, which is accounted for
by the increase in carbon dioxide.
To explain this strange observation, it should be noted that drilling fluid filtrate is caustic, containing
multiple oil- or water-soluble basic compounds designed to scavenge acidic compounds, such as hydrogen
sulfide or carbon dioxide. Drilling fluid can contain water-soluble and/or oil-soluble acid scavenging
compounds to help reduce corrosion and bind with acid species to protect against hydrogen sulfide. These
reactions can be reversible with a reduction in pressure to recover carbon dioxide or hydrogen sulfide
(Harrison et al. 1999; Garcia and Lordo 2007). It might be the case that before 91 minutes, the concentration
of the drilling fluid filtrate was sufficient that carbon dioxide was largely bound in a complex. For caustic
chemicals used to scavenge acidic species in drilling fluid filtrate, such as but not limited to amines, the
optical spectroscopy of bound carbon dioxide is likely different from that of unbound carbon dioxide (Vogta
et al. 2011; Idrisa et al. 2014). A spectroscopic signature for bound carbon dioxide therefore would likely
not have registered as free carbon dioxide for an MOC sensor. The WFT contained a density sensor (Jones
et al. 2007; Palmer et al. 2008; Gao et al 2010), which recorded a 100% filtrate fluid as having a density of
0.8150 g/cc and a 3.22% filtrate fluid as 0.7016 g/cc. Using a linear mixing model for drilling fluid filtrate
contamination in a formation fluid (Hsu et al. 2008; Zuo et al. 2015), the contamination at 91 minutes,
just before the carbon dioxide level increase, can be calculated as 14.28% by using the measured density
of 0.71452 g/cc. If one assumes a filtrate component binding of carbon dioxide as a limiting reagent, then
14.28% binds with 100% of the carbon dioxide, and 3.22%, the measured contamination in the sample would
bind with 22.6% of the carbon dioxide, thereby accounting for the difference between the carbon dioxide
measured by the MOC sensor versus the laboratory. If corrected for this potential drilling fluid filtrate effect,
the concentration of carbon dioxide in the reservoir fluid by measure of the MOC sensor would be estimated
as 9.35 wt% compared to the laboratory measure of 9.04 wt%, a difference of only 0.31 wt%.
Applying this correction technique by using the contamination estimation available at the time of
sampling, a carbon dioxide value of 8.3 wt% is calculated. In a total of five field trials over pumpout stations,
carbon dioxide was determined with an accuracy of +/-0.5 wt% standard error as calculated by Eq. 16 over
a concentration ranging from 6.0 wt% to 23 wt% as shown in Fig.. SDF was determined to within +/-1.0
wt% accuracy by the standard error calculation of Eq. 16.

Eq. 16

Figure 12—Field test of the SDF ICE core was completed on four wells. The standard error accuracy between
the lab and MOC measurements for field tests 1-4 was +/-1.0 wt% absolute vs. the standard error between lab
and asymptotic fit to the density measures which provided a +/-9.8 wt% absolute standard error measurement.
Comparison of sensor measurements of carbon dioxide dissolved in petroleum down hole with GC measurements
on live fluid samples captured with a WFT at the same time. The standard error for CO2 measurement
was determined from these measurements to be 0.5 wt% CO2 over a concentration range ~6-23 wt% CO2.
SPE-195858-MS 19

Conclusion
The MOC sensor and MOE regression vector provide many advantages compared to conventional
spectroscopic equipment using typical PLS regression vectors. The MOC system is compact and robust yet
provides similar spectroscopic accuracy as a laboratory system. In fact, for this study, the MOC based on
a dual-core MOE has been validated to better accuracy than that of an FTIR using a conventional four-
factor PLS regression. However, this performance equivalence is likely dictated by the complexity of the
regression vector necessary and data preprocessing required. It is conceivable, and even probable, that
various applications could require a more complicated MOE regression vector than that which is possible
with a composite transmission function achievable by an interference pattern. Additionally, complex
spectral preprocessing methods, such as but not limited to derivative spectroscopy, as is common with digital
representations, have not been shown with MOC. With these limitations aside, in many cases, such as during
well monitoring of carbon dioxide and methane, the performance equivalence of MOC with laboratory
systems is more than sufficient, and the robustness of the MOC system enables applications not otherwise
easily achievable.
The MOC has also been field tested with results for methane that were well within the expected accuracy
range, as was laboratory validated. However, the carbon dioxide field results differed substantially from
those expected based on laboratory validation. The field results for carbon dioxide might, in fact, be correct,
and it is possible that it is the first observation of a physical process not previously suspected that can
occur during a formation sampling pumpout. Specifically, it is proposed that components associated with
the mud filtrate can bind the carbon dioxide. The components would then release the carbon dioxide with a
pressure reduction. The mechanism for such a phenomenon is currently the subject of a different laboratory
study. The proposal of such a mechanism seems self-consistent with the observed behavior of the pumpout
trend, and the observed behavior can be used to correct the MOC measured carbon dioxide to a level far
more consistent with the measured laboratory concentration of carbon dioxide. Nonetheless, the uncorrected
MOC-measured values of carbon dioxide and methane are highly useful for assessment of production issues
associated with carbon dioxide or various exploration and production activities, including WFT sampling.
The wells drilled spanned a diverse range within the pre-salt of Brazil. Therefore, because of the success
in determination of SDF filtrate contamination, the SDF variation must have been sufficiently characterized.
Likewise, the variation of pre-salt oils must have been sufficiently characterized. The accuracy determined
from field tests fall within the estimates of the theoretical MOE design with the accuracy of the four well
tests of 1.0 wt% nearly matching the MOE design accuracy of 1.25%. Therefore, we believe there is not a
significant difference between the MOC sensor performance and the PLS laboratory performance. Also, the
direct contamination assessment with the MIR sensor was significantly more successful than direct density
fitting by the asymptotic method. A noted limitation may be the lack of universality in the nature of the
SDF markers, however, the MIR sensor design is easily modifiable. MOE cores may easily be substituted
for those designed to any locality. The larger limitation of the current method is that it is only applicable to
SDF and not generic OBM containing only natural components.

References
Abouie, Ali, et al. Comprehensive Modeling of Scale Deposition by Use of a Coupled Geochemical and Compositional
Wellbore Simulator. Society of Petroleum Engineers. 4, s.l. : Society of Petroleum Engineers, 2017, SPE Journal, Vol.
22, pp. 1225 - 1241. SPE-185942-PA.
Beebe, K. R. and Kowalski, B. R. An Introduction to Multivariate Cialibration and Analysis. 17, 1991, ANALYTICAL
CHEMISTRY, Vol. 59, pp. 1007-1017.
Bertie, John E., Zhang, Shuliang L. and Keefe, C. Dale. Measurement and use of absolute infrared absorption intensities
of neat liquids. 2, s.l. : Elsevier, 1995, Vibrational Spectroscopy, Vol. 8, pp. 215-229
Bonis, Michel R. and Crolet, Jean Louis. Why So Low Free Acetic Acid Thresholds, in Sweet Corrosion at Low PCO2?
Houston : NACE, 2005. Corrosion. pp. NACE-05272.
20 SPE-195858-MS

Bro, Rasmus and Anderson, Charlotte M. Theory of net analyte signal vectors in inverse regression. 12, s.l. : Wiley, 2003,
Journal of Chemometrics, Vol. 17, pp. 646-652.
Centner, Vitezslav and Massart, D. Luc. Optimization in Locally Weighted Regression. 19, s.l. : American Chemical
Society, 1988, Analytical Chemistry, Vol. 70, pp. 4206-4211.
Chalmers, John M. and Griffiths, Peter R., [ed.]. Handbook of Vibrational Spectroscopy: Theory and instrumentation.
Chichester : John Wiley & Sons Ltd, 2002. Vol. V1. ISBN: 978-0-471-98847-2.
Commerce, U.S. Secretary of. NIST Standard Reference Data. NIST Chemistry WebBook, SRD 69. [Online] 2018. [Cited:
March 26, 2019.] https://webbook.nist.gov/.
Dandekar, Abhijit Y. Petroleum Reservoir Rock and Fluid Properties. Boca Raton : Taylor & Francis Group, CRC Press,
2006.
Dong, C., et al., New downhole-fluid-analysis tool for improved reservoir characterization. 6, s.l. : Society of Petroleum
Engineers, 2008, SPE Reservoir Evaluation and Engineering, Vol. 11, pp. 1107-1116. SPE-108566-PA.
Ellis, Darwin V. and Singer, Julian M. Well logging for Earth Scientists, 2nd Ed. Dordrecht : Springer, 2008.
Elshahawi, Hani, et al., The Power of Real-Time Monitoring and Interpretation in Wireline Formation Testing–Case
Studies. 3, s.l. : Society of Petroleum Engineers, 2007, SPE Reservoir Evaluation and Engineering, Vol. 10, pp.
241–250. SPE-94708-PA.
Eriksen, Kåre Otto, et al. Field Tests of a New Optical Sensor Based on Integrated Computational Elements for Downhole
Fluid Analysis. New Orleans, Louisiana : Society of Petroleum Engineers, 2013. SPE Annual Technical Conference
and Exhibition. pp. SPE-166415-MS. SPE-166415-MS.
Faber, Nicolaas (Klaas) M. Efficient Computation of Net Analyte Signal in Inverse Multivariate Calibration Models. 23,
s.l. : American Chemical Society, 1998, Analytical Chemistry, Vol. 70, pp. 5108-5110.
Gao, Li, et al. Sensitivity of a High-Resolution Fluid-Density Sensor in Multiphase Flow. Florence, Italy : Society of
Petroleum Engineers, 2010. SPE Annual Technical Conference and Exhibition. SPE-133405-MS.
Garcia, John M. and Lordo, Samuel A. Chemistry and Impacts of Chemistry And Impacts Of Commonly Used Amine-
Based H2S Scavengers On Crude Unit Towers And Overheads. Nashville, Tennessee : NACE International, 2007.
NACE International Conference and Expo: Corrosion. pp. NACE-07571. NACE-07571.
Gemperline, P. J., Long, J. R. and Gregoriou, V. G. Nonlinear multivariate calibration using principal components
regression and artificial neural networks. 20, 1991, Anal. Chem., Vol. 63, pp. 2313-2323.
Goldstein, N., et al. Real-Time Optical BTU Measurement of Natural Gas at Line Pressure. Denver, Colorado :
International Symposium on Fluid Flow Measurement, 1999. 4th International Symposium on Fluid Flow
Measurement.
H. M. Hcise, Dortmund. Infrared spectra of gases. [ed.] Schrader and Bernhard. Infrared and Raman Spectroscopy:
Methods and Applications. Weinheim : VCH Verlagsgescllschaft, 1995, p. 264.
Hammond, P. S. One- and two-phase flow during fluid sampling by a wireline tool. 3, s.l. : Springer, Kluwer Academic
Publishers., 1991, Transport in Porous Media, Vol. 6, pp. 299-330.
Harrison, James R., et al. Novel Lime-Free Drilling Fluid System Applied Successfully in Gulf of Thailand. Amsterdam,
NE : Society of Petroleum Engineers, 1999. SPE/IADC Drilling Conference. pp. SPE-52817-MS. SPE/IADC-52817-
MS.
Herzberg, Gerhard. Molecular Spectra and Molecular Structure: Infrared and Raman Spectra of Polyatomic Molecules.
New York : Van Nostrand, 1945, Vol. 2, pp. 266–269.
Hsu, Kai, et al. Multichannel Oil-Base Mud Contamination Monitoring Using Downhole Optical Spectrometer. Austin,
Texas : Society of Petrophysicists and Well-Log Analysts, 2008. 49th Annual Logging Symposium. SPWLA-2008-
QQQQ.
Hy-Billiot, J., et al. Getting the Best From Formation Tester Sampling. San Antonio, TX : Society of Petroleum Engineers,
2002. SPE Annual Technical Conference and Exhibition. pp. SPE-77771-MS. SPE-77771-MS.
Idrisa, I., Jensa, K. and Eimer, D. A. Speciation of MEA-CO2 Adducts at Equilibrium Using Raman Spectroscopy.
Supplement C, s.l. : Elsevier, 2014, Energy Procedia, Vol. 63, pp. 1424–1431.
Jackson, E. J., A user's guide to principal components. New York : Wiley, 2005.
Jones, C. M., et al. Laboratory Quality Optical Analysis in Harsh Environments. Kuwait City, Kuwait : Society of
Petroleum Engineers, 2012. Presented at the SPE Kuwait International Petroleum Conference and Exhibition. pp.
SPE-163289-MS. SPE-163289-MS.
Jones, C. M., et al. Measurement and Use of Formation Fluid, Saturate, and Aromatic Long Beach, : Society
of Petrophysicists and Well Log Analysts, 2015. SPWLA 56th Annual Symposium. pp. SPWLA-2015-EE.
SPWLA-2015-EE.
Jones, Chris, et al. Collecting Single-Phase Retrograde Gas Samples at Near-Dewpoint Reservoir Pressure in Carbonates
Using a Pump-Out Formation Tester With an Oval Pad. Anaheim, California : Society of Petroleum Engineers, 2007.
SPE Annual Technical Conference and Exhibition. SPE-110831-MS.
SPE-195858-MS 21

Jones, Christopher Michael, et al. A Small-Volume PVTX System for Broadband Spectroscopic Calibration of Downhole
Optical Sensors. 7, s.l. : American Institute of Physics, 2017, Review of Scientific Instruments, Vol. 88.
Jones, Christopher Michael, van Zuilekom, Tony and Iskander, Fady. How Accurate Is Enhanced Optical Fluid Analysis
Compared to Lab PVT Measurements? Reykjavik, Iceland : Society of Petrophysicists and Well Log Analysts, 2016.
SPWLA 57th Annual Logging Symposium. pp. SPWLA-2016-JJJ. SPWLA-2016-JJJ.
Jones, Christopher, et al. Field Test of the Integrated Computational Elements: A New Optical Sensor for Downhole Fluid
Analysis. New Orleans, Louisiana : Society of Petrophysicists and Well-Log Analysts, 2013. SPWLA 54th Annual
Logging Symposium. pp. SPWLA-2013-YY. SPWLA-2013-YY.
Kiehl, J. T., Bruhl, Chr. and Yamanouchi, T. A parameterization for the absorption due to the near infrared bands of CO2.
4-5, s.l. : Taylor & Francis, 1985, Tellus, Series B - Chemical and Physical Meteorology, Vol. 37B, pp. 189-196.
Liu, Fei, Jiang, Yihong and He, Yong. Variable selection in visible/near infrared spectra for linear and nonlinear
calibrations: A case study to determine soluble solids content of beer. 1, s.l. : Elsevier, 2009, Analytica Chimica Acta,
Vol. 635, pp. 45-52.
Lorber, A., Faber, K. and Kowalski, B., R. Net Analyte Signal Calculation in Multivariate. 8, 1997, Anal. Chem., Vol.
69, pp. 1620-1626.
Loubar, K., et al. Combustion Properties Determination of Natural Gas Using Thermal Conductivity and CO2 Content.
s.l. : SAE International, 2005, SAE Technical Paper. 2005-01-3774.
Martens, H. and Naes, T. Multivariate Calibration. New York : Wiley, 1992.
Massart, D. L., et al. Chemometrics: a textbook. Amsterdam : Elsevier Science Publishers, 1988.
Mueller, Nadja, et al. Quantification of carbon dioxide using downhole Wireline formation tester measurements. San
Antonio, Texas : Society of Petroleum Engineers, 2006. SPE Annual Technical Conference and Exhibition. pp.
SPE-100739-MS.
Mullins, Oliver C., et al. Evaluation of Coexisting Reservoir Fluid Gradients of GOR, Asphaltene and Biomarkers as
Determined by Charge History and Reservoir Fluid Geodynamics. Beach, California, USA : Society of Petrophysicists
and Well-Log Analysts, 2015. SPWLA 56th Annual Logging Symposium. SPWLA-2015-OO.
NACE International. Corrosion of Oil-and Gas-Well Equipment. [ed.] J. D. Sudbury, et al., et al Dallas : American
Petroleum Institute, 1958. B0010IBZBO.
Nadler, B. and Coifman, R. R. Partial least squares, Beer's law and the net analyte: statistical modeling and analysis.
1, 2005, J. Chemometrics, Vol. 19, pp. 45–54.
Næs, Tormod and Isaksson, Tomas. Locally Weighted Regression in Diffuse Near-Infrared Transmittance Spectroscopy.
1, s.l. : SAGE, 1992, Applied Spectroscopy, Vol. 46, pp. 34-43.
Nelson, M. P., et al. Multivariate Optical Computation for Predictive Spectroscopy. 1, 1998, Anal. Chem., Vol. 70, pp.
73–82.
Palmer, Richard, et al. Advances In Fluid Identification Methods Using A High Resolution Densitometer In A Saudi
Aramco Field. Austin, Texas : Society of Petrophysicists & Well Log Analysts, 2008. 49th Annual Logging
Symposium. SPWLA-2008-MM.
Pedersen, K., and Peter, L. Christensen P. Phase Behavior of Petroleum Reservoir Fluids. Boca Raton : CRC Press, 2007.
Perez, Teresa E. Corrosion in the Oil and Gas Industry: An Increasing Challenge for Materials. 8, s.l. : Springer, 2013,
The Journal of The Minerals, Metals and Materials Society, Vol. 65.
Person, Willis Bagley and Zerbi, Giuseppe. Vibrational intensities in infrared and Raman spectroscopy: Studies in Physical
and Theoretical Chemistry. New York : Elsevier Scientific Pub. Co., 1982. Vol. 20.
Popoola, Lekan Taofeek, et al. Corrosion problems during oil and gas production and its mitigation. 35, s.l. : Springer,
2013, International Journal of Industrial Chemistry, Vol. 4.
Rabkin, Yakov M. Technological Innovation in Science: The Adoption of Infrared Spectroscopy by Chemists. 1, s.l. : The
University of Chicago Press, 1987, Isis, Vol. 78, pp. 31-54.
Schrader, Bernhard. Infrared and Raman Spectroscopy: Method and Applications. New York : VCH Publishers, Inc., 1995.
Soyemi, O., et al. Design and Testing of a Multivariate Optical Element: The First Demonstration of Multivariate Optical
Computing for Predictive Spectroscopy. 6, 2001, Analytical Chemistry, Vol. 73, pp. 1069–1079.
Soyemi, O., et al. Nonlinear Optimization Algorithm for Multivariate Optical Element Design. 4, 2002, Applied
Spectroscopy, Vol. 56, pp. 477-487.
Thomas, Edward V. A Primer on Multivariate Calibraiton. 15, s.l. : American Chemical Society, 1994, Analytical
Chemistry, Vol. 66, pp. 795-804.
Tjomsland, T, Grotle, N. M. and Vikane, O. Scale Control Strategy and Economical Consequences of Scale at Veslefrikk.
Aberdeen, United Kingdom : Society of Petroleum Engineers, 2001. International Symposium on Oilfield Scale. pp.
SPE-68308-MS. SPE-68308-MS.
22 SPE-195858-MS

Tjomsland, Tore, et al. Veslefrikk Scale Control Strategy and Economic Implications: Revisited 7 years later - did we
improve? s.l. : Society of Petroleum Engineers, 2008. SPE International Oilfield Scale Conference. pp. SPE-114086-
MS. SPE-114086-MS.
Udina, S., et al. A micromachined thermoelectric sensor for natural gas analysis: Thermal model and experimental results.
2, s.l. : Elsevier, 2008, Sensors and Actuators B: Chemical, Vol. 134, pp. 551-558.
Ulbig, Peter and Hoburg, Detlev. Determination of the calorific value of natural gas by different methods. 1-2, s.l. :
Elsevier, 2002, Thermochimica Acta, Vol. 382, pp. 27-35.
Valderrama, Patrícia, Braga, Jez Willian B. and Poppi, Ronei Jesus. Variable Selection, Outlier Detection, and Figures
of Merit Estimation in a Partial Least-Squares Regression Multivariate Calibration Model. A Case Study for the
Determination of Quality Parameters in the Alcohol Industry by Near-Infrared Spectroscopy. 21, s.l. : American
Chemical Society, 2007, Journal of Agricultural Food Chemistry, Vol. 55, pp. 8331–8338.
Vogta, M., Pasel, B. and Bathena, D. Characterisation of CO2 absorption in various solvents for PCC applications by
Raman spectroscopy. Supplement C, s.l. : Elsevier, 2011, Energy Procedia, Vol. 4, pp. 1520-1525.
Wang, Kang, et al. Differing Equilibration Times of GOR, Asphaltenes and Biomarkers as Determined by Charge History
and Reservoir Fluid Geodynamics. 5, s.l. : Society of Petrophysicists and Well-Log Analysts, 2015, Petrophysics, Vol.
56, pp. 440 - 456. SPWLA-2015-v56n5a2.
Zuo, Julian Y., et al. A Breakthrough in Accurate Downhole Fluid Sample Contamination Prediction in Real Time. 3, s.l. :
Society of Petrophysicists & Well Log Analysts, 2015, Petrophysics, Vol. 56, pp. 251 - 265. SPWLA-2015-v56n3a2.