You are on page 1of 17

Rock typing based on capillary pressure parameters

using supervised learning 1

Cesar Aguilar, 1 () ,


1. Reservoir Studies department, PDVSA
Aguilarcp9@gmail.com

Abstract. Improve reservoir description techniques is one of the most


challenging tasks for geoscientists and engineers. Detailed petrophysi-
cal characterization is crucial for describing hydrocarbon reservoirs. Pa-
rameters that may help in the understanding of what is controlling flo w
capacity within a reservoir can be extracted from capillary pressure
curves and linked with rock types. It has been shown that machine
learning has the potential to reduce interpreter bias and generate solid
results in the core-log domain, even in cases of limited data availability,
which is usua lly the case when capillary pressure data is involved. Th is
paper presents a methodology for rock typing using cluster analysis
technique based on supervised approach aiming to learn from reservoir
mercury injection capillary pressure tests, (MICP). M ULTI -
RESOLUTION-GRAPH-BASED CLUSTERING (MRGC ) is a very
powerful non-parametric algorithm which allows definin g the optimal
number of rock types/hydraulic units was used.

The methodology has been tested in a real data set from sandstone res-
ervoir integrating core and log data as an input data. Compared to pre-
viously classification, based on lithofacies information, it yielded to en-
hance reservoir description and permeability prediction. Six hydraulic
units were identified, without any “a priori” subjective knowledge. Re-
sults were afterwards compared with all the geolo gical information
available (core descriptions, image log, XRD, and thin sections) with
the purpose of assessing uncertainties. A detailed characterization was

Copyright 2022, IFEDC Organizing Committee


This paper was prepared for presentation at the 2022 International Field Exploration and Develop-
ment Conference in Urumqi, China, 16-18 August 2022.
This paper was selected for presentation by the IFEDC Committee following review of information
contained in an abstract submitted by the author(s). Contents of the paper, as presented, have not
been reviewed by the IFEDC Technical Team and are subject to correction by the author(s). The
material does not necessarily reflect any position of the IFEDC Technical Committee its members.
Papers presented at the Conference are subject to publication review by Professional Team of IFEDC
Technical Committee. Electronic reproduction, distribution, or storage of any part of this paper for
commercial purposes without the written consent of IFEDC Organizing Committee 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 IFEDC. Contact
email: paper@ifedc.org.
2

obtained, determining rock properties (porosity and permeability) and


SCAL properties (saturation height function) by each rock type.

Keywords: Rock typing, cluster analysis, machine learning, mercury


injection capillary pressure, MICP

Introduction

In recent years, petrophysicists have established that rock typing is crucial for estimat-
ing permeability in reservoir rocks. Any attempt that does not use rock typing for
modeling permeability has been documented to be characterized by a large data scat-
ter and it is associated with high uncertainty. Different approaches have been pro-
posed partitioning well-log responses into dist inct classes in order to improve reser-
voir characterization and permeability predictions. The simplest approach uses reser-
voir layering. Other schemes are using lithofacies information identified from core
description. Rabiller [1], Cuddy [2], Tiab [3], stated that “there is a strong non-linear
relationship between lo gs and lithological data, which proves difficult to tie any log
data to a sedimentologically relevant description”. Rabiller pointed out at the differ-
ences between the relative distances between classes in the log space and geological
space is the cause for this non- linearity. He proposed a two-step approach addressing
the dimensionality problem; first an automatic clusters pro cess should be performed
and second a manual merge between classes could be performed whether is needed
based on geological knowledge.

Capillary pressure can be defined as “The pressure differential between two immisci-
ble fluid phases occupying the same pores caused by interfacial tension between the
two phases that must be overcome to initiate flow”. Moreover, is indicative of the
pore throat sorting of a rock, which itself may define different rock types on the res-
ervoir.

If the bundled capillary tube m odel is assumed the pore throat size can be related with
capillary pressure using the following equation:

2σ cosθ
r= (1)
𝑃𝑐
where: r= pore throat size, σ cosθ is the interfacial tension multiplied by the cosine
of the contact angle and Pc is the capillary pressure.

Parameters that may help in the understanding of what is controlling permeability


within a reservoir can be extracted from capillary pressure curves. Thomeer [4], de-
veloped a method for the analysis of mercury injection capillary p ressure (MICP)
data. He observed that the data from the MICP experiment, for simple rock types,
3

could be represented by a hyperbola when plotted on Log-Log plot. The data to be


used is: (1 ) volume of mercury injected (X axis), and (2) applied pressure between the
wetting (air) and non-wetting phases (mercury) as mercury intrudes into the pore
space fraction (Y axis). Considering that for a rock sample, the bulk volume is known,
so that the volume of mercury injected can be re-expressed as a fraction of the total
sample bulk volume, Bv. To fit the hyperbola to the data, the value of two asymp-
totes, B∞ , Pd, are required. The Thomeer Hyperbola is expressed as:

Bv Pc
Log ( ) x Log ( ) = K (2)
B∞ Pd

Where; K is a hyperbola shape factor.

The asymptotes are B∞ is the percent bulk volume occupied by mercury at infinite
applied pressure and Pd, the displacement pressure required to f irst intrude mercury
into the largest pore throat. Thomeer choses to express constant K = Log [exp ( - G)]
such that Equation 2 becomes:

Bv /B∞ = exp [−G/Log ( Pc /Pd )] (3)

Where; G is the Pore Geometrical Factor and determines the hyperbola shape. In
practice, Bv and Pc data from the MICP experiment are fit by equation (3) to deter-
mine Pd , B∞ and G, for individual samples.

Using a weighted regression on data from 279 rock samples, Thomeer was able to
find a relationship between the three parameters and air permeability (Ka):

𝐵∞ 2
𝐾𝑎 = ( 3.8068 𝑥𝐺 −1.3334 ( ) ) (4)
𝑃𝑑

The following is a summary of the parameters that could be found useful for rock
typing. The disp lacement pressure Leverett is the minimum pressure required for the
non-wetting fluid to begin d isplacing the wetting fluid from the largest pores. The
displacement pressure may be estimated graphically by extending the slope of the
plateau to a capillary pressure curve to 100% wetting phase saturation. The asymp-
totes B∞ is the percent bulk volume occupied by mercury at infinite. This is not irre-
ducible water saturation because air is a compressible fluid. Fig. 1 left side.

Pittman [5] has studied the pore throat size at a particular saturation when the pore
network is connected (Fig. 1 right side) and derived empirically equations in order to
relate the pore throat size (PTR ), porosity and permeability. Furthermore, he used th is
4

PTR as a classification technique. According to the PTR he uses f ive classes: Mega
>2.5 microns, Macro <2.5 but >1.5 microns, Meso <1.5 but >0.5, micro <0.5 but >0.1
and nano<0.1 microns. However, this number of classes and range should not be con-
sidered suitable for every reservoir existent.

Another relevant parameter that can be related to permeability is the Swanson param-
eter [6] a relationship between non-wetting phase volume and capillary pressure,
wherein (Vnw/Pc)max at the apex of a capillary pressure curve is correlated with
permeability. Fig. 1. left side.

Fig. 1. Typical p lot for parameter extraction from capillary curve. Mercury saturation
Vs Capillary pressure, left side. Pore throat radius Vs Mercury saturation, right side

Theologou [7], develop a suitable software package for interpretation of special core
analysis and integration with logs, with this tool loading, correcting, and proce ssing
mercury injection capillary pressure is possible. Moreover, upon comparison of dif-
ferent approaches, he found that the most flexible and reliable curve form for hyper-
bolic modelin g of the capillary curves was the three constant hyperbolic model pro-
posed by Donaldson [8], having the form:

𝐴 + 𝐵 ∗ 𝑆𝑤
𝑃𝑐 = (5)
1 + 𝐶 ∗ 𝑆𝑊

Where A, B, C are three adimensional constants for hyperbolic fitting.


5

A d isadvantage of the hyperbola methods is the inability to easily model rocks with a
bimodal pore throat sorting. In this case two or more curves with different characteris-
tics need to be combined and used.

Theoretical basis

Machine learning is a form of AI that enables a system to learn from data rather than
through explicit programming. However, machine learning is not a simple process. It
uses a variety of algorithms that iteratively learn from data to improve data descrip-
tion and predict outcomes. As the algorithms ingest training data, it is then possib le to
produce more precise models based on that data. A machine learning model is the
output generated when you train your machine learning algorithm with data. After
training, when you provide a model with an input, you will be given an output. A
supervised learning algorithm takes a known set of input data (the training set) and
known responses to the data (output) and trains a model to generate reasonable pre-
dictions for the response to new input data. Supervised learning should be used when
there is existing data for the output to be predicted.

According to Rabiller [9], MULTI-RESOLUTION GR APH-B ASED CLUSTERING


(MGRC) is a very powerful non-parametric algorithm which allows to define the
optimal number of hydraulic units. Based on two parameters; Kernel representative
Index (KRI) and Neighbor index (NI), the zonation process is optimized. None of
these parameters are directly introduced by the user, they are automated calculated by
the algorithm based on the data distribution. MRGC It is a two -steps algorithm. First,
based on the behavior of the underlying multivariable probabilist ic density function
the algorithm is able to recognize important structures (modes) which su ggest a dif-
ferent class. Then the user has the possibility to merge different clusters based on
geological meaning. The major advantage of uses MRGC algorithm is that does not
need any “a priori” knowled ge to perform the clusters analysis. Moreover, is capable
of recognize different shapes and sizes clusters.

Research method

In order to define optimal rock types in the reservoir the following methodology was
carried out (Error! Reference source not found.). Logs and core information were
loaded in the software creating a robust database. Sixteen (16) samples from MICP,
15 from centrifuge test (water- oil d rainage tests were performed) and 168 porosity-
permeability measurements from 1 cored well were used. All the logs available were
loaded and environmentally corrected. Special lo gs such as Image logs and NMR
were loaded and processed to secure their appropriate display and interpretation. This
task is considered as the more time-consuming task in the workflow but should be
performed carefully to guarantee final results.
6

Fig. 2. Methodology used to define optimal number of rock types.

Capillary pressure data should be corrected to reproduce reservoir conditions by stress


and cationic exchange using the following equations proposed by Hill [10]:

𝑃𝑐
𝑃𝑐𝑐𝑜𝑟 = (6)
𝜑𝑟𝑒𝑠𝑒𝑟𝑣𝑜𝑖𝑟 𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛𝑠

𝜑𝑙𝑎𝑏 𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛𝑠

𝜑𝑟𝑒𝑠 𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛𝑠
𝑆𝑤𝑐𝑜𝑟 = 1 − (1 − 𝑆𝑤 ) ( ) (7)
𝜑𝑙𝑎𝑏 𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛𝑠

1

𝑆𝑜 ∗= 𝑆ℎ𝑔(1 − (0.084 ∗ 𝐶𝑜 2 + 0.22) ∗ (𝑄𝑣) (8)

Where Co is the water salinity and Qv is the cationic exchange.

MICP data should be corrected by closure correction in tight and unconventional


reservoirs, this is not the case for our study case.
7

Using extracted MICP parameters as training data, logs were submitted to a cluster
analysis by MRGC algorithm. Rock types had automated identified by the software
and were analyzed using histograms, crossplots and the integration of all the geologi-
cal information available (core descript ions, special logs, XRD analysis, th in section
and SEM photos information). Moreover, centrifuge test results were used as blind
data (data not used in the t raining process) in order to check if predictions are con-
sistent. Finally, all the information was propagated across the well and results were
checked using histograms and crossplots.

Field application

Reservoir under study is an Eocene shaly sandstone reservoir. Sand deposits are pri-
marily of channel-type corresponding to a fluvial-deltaic deposition system. Accord-
ing to core measurement, its porosity varies between 5% and 33% with a mean value
of 22 % and its permeability varies between 0.5 md and 2100 md with a mean value
of 660 md.

The main purpose of this detailed characterization is identified new infill drilling
opportunities and by–passed oil zones in the reservoir.

1.1 Data corrections

Core information was depth matched with log. Moreover, porosity and permeability
core data must be stress-corrected to simulate reservoir-confining conditions. Stress
corrections were applied to porosity and permeability measurements followin g Jones
methodology [12].

Environmental corrections were applied to all logs and gamma ray logs were normal-
ized. Bad hole corrections were applied to the density log.

Using a commercial software, MICP data was corrected by stress and clay cationic
exchange using equations 6, 7, and 8. Samples with high porosity are more sensitive
to stress corrections. Contrarily, samples with low porosity are more sensitive to clay
cationic exchange Error! Reference source not found..
8

Fig. 3. MICP data corrections

1.2 Extracting parameters

Using a commercial software, hyperbolic model proposed by Donaldson (Equation 5)


was used to fit corrected MICP data. Results are displayed in a cross-plot between (1
– mercury saturation) or water saturation (X axis) versus capillary pressure (Y axis)
and permeability is used as a color scale (Blue is low permeability – red is high per-
meability). Each sample availa ble was fitted and plotted. Results obtained were con-
sidered good because model curve match its corresponding MICP real curve (Fig. 4).
Only sample 1205 fitting exhibits a small deviation with respect to the real data.
9

Fig. 4. MICP fitting using the model proposed by Donaldson.

Once the fitting is reached, several parameters could be extracted using extrapola-
tions, as was explained in the introduction section. Results are presented for each
sample and depth in the Error! Reference source not found.Error! Reference
source not found.. Parameters could be used as input data for rock typing and perme-
ability predict ion. In some cases when fitting is not reached samples should be d is-
carded. A careful quality control should be performed in order to detect lab measure-
ments problems. Some samples could be altered/fractured during this test or during
sample preparation or a whole section could be damaged during core extraction pro-
cess.
10

Table 1. Parameters extracted of the MICP data


HYPER HYPER HYPER
DEPTH SAMPLE A B C LAMBA PD PTR APEX Bv∞
8400.60 1125 -107.13 119.88 -6.84 1.63 3.38 12.59 0.08 0.15
8406.80 1205 -225.19 220.64 -3.26 2.34 11.28 1.46 0.02 0.31
8420.10 1208 -85.38 96.82 -7.32 1.53 2.83 17.85 0.11 0.14
8425.74 1219A -59.84 74.23 -5.18 1.56 2.43 21.79 0.12 0.19
8433.34 1301 -86.16 96.20 -5.43 1.64 3.38 11.41 0.09 0.18
8440.15 1303 -88.63 104.83 -6.75 1.59 2.85 17.62 0.11 0.15
8450.15 1313 -61.83 76.40 -4.37 1.63 2.64 16.13 0.11 0.23
8451.75 1315 -56.96 70.52 -5.09 1.55 2.39 22.11 0.13 0.20
8492.15 1425 -96.18 120.85 -5.53 1.73 2.90 15.71 0.09 0.18
8587.50 1714 -80.54 92.52 -5.05 1.65 3.22 11.93 0.08 0.20
8589.80 1716 -106.61 127.87 -7.86 1.60 2.86 18.66 0.10 0.13
8591.10 1801 -81.14 90.80 -6.00 1.58 3.10 13.98 0.10 0.17
8605.30 1815 -114.63 126.80 -7.42 1.62 3.42 12.67 0.08 0.13
8672.07 2019 -113.65 133.54 -7.47 1.64 3.12 15.31 0.08 0.13
8747.17 2305 -186.85 227.07 -5.36 2.14 3.92 6.75 0.06 0.19
8759.87 2318 -169.12 209.64 -6.68 1.93 3.50 10.85 0.08 0.15

1.3 Supervised learning

Based on the MRGC Algorithm using pore throat radius as training data and shale
volume and effective porosity calculated by logs as predictors, six different rock types
were identified. Predictors logs were chosen in oder to extrapolate results to other
wells in the reservoir. Different extracted capillary parameters were tested and pore
throat radius at SHg=35% exhib its the best resu lts. Each rock type has different aver-
age value of the input data and a corresponding color, as shown in the Error! Refer-
ence source not found..

Rock type 6 (yello w) has the best reservo ir properties with h igher porosity and low
clay volume compared with the other reservoir rocks. On the other hand, rock type 1
(green) is a non- reservoir rock, with lo w porosity and high clay volume. Even if the
training data set is constrained to 16 samples good results were reached.
11

Table 2. Average value of the input data for each rock type

EFFECTIVE CLAY
ROCK TYPES PTR
POROSITY VOLUME

1 0.13 0.3 2.99


2 0.18 0.23 11.10
3 0.23 0.17 12.3
4 0.24 0.16 13.01
5 0.26 0.13 13.37
6 0.33 0.05 19.12

Agu ilar [13] proposed that predicted rock types should be compared with all the geo-
logical information available (Core descriptions, image log, XRD, and thin section) to
observe if the results are linked together and the classification is valid. Results ob-
tained in the cored well will be showed (Error! Reference source not found. & Er-
ror! Reference source not found.). Track 1 gamma ray and caliper lo gs and track 2
show resistiv ity curves, respectively. Track 3 displays porosity logs. Track number 4
shows image logs. Track 6 and 7 display core photos and thin section photos, respec-
tively. Track 8 shows automated classification (Yello w co lor is the best quality and
green color is the wo rst quality). Track 9 displays T2 distribution from NMR log.
Track 10 shows shale volume calculation derived from gamma ray (red line), XRD
data from core in points and clay volume from spectroscopy log (filled curve), Track
11 compares total core porosity (in points) and total porosity calculated by logs. Track
number 12 shows comparison between core permeability (in points) and permeability
calculated using logs capillary parameters in red line (calculations will be further
explained in the next section) also Timur permeability is displayed in green line.

In both figures a match between geological information and automated classification


can be found. Rock with good hydrocarbon impregnation and big grain size in co re
photos, well so rted in thin section, low clay volume in XRD data, light yellow resis-
tive colors in image logs, long T2 d istribution from NMR, high porosity and high
permeability was identified as the best rock quality type. Otherwise, rocks with poor
hydrocarbon impregnation, small grain size in core photos, poor sortin g in thin sec-
tion, high clay volume in XR D data, dark conductive colors in image logs, short T2
distribution from NMR, low porosity and low permeability d isplay worst rock quality
types. A domain expert is required in order to recognize this match because geologi-
cal features could not be expressed in simple terms. Domain experts should recogn ize
systematic errors in rock type classification process and identify if the cause is due to
training data quality or quantity, predictor logs quality, dimensionality problem or all
of them and test different approaches to solve some core -log calibration problems.
12

Fig. 5. Results obtained compared with geological information. Interval 8560ft-8625ft.

Fig. 6. Results obtained compared with geological information. Interval 8425ft-8460ft.


13

1.4 Blind data testing

To test the predictive ability of the model blind data was used. If the automatic pro-
cess is robust, rock types predicted at the depth of the centrifuge capillary test will be
reasonable in terms of irreducible water saturation values and capillary curve shape
even when these points were not used during the training process. Results obtained
(Error! Reference source not found.) honors that samples identified as RT6 (yel-
low) exhibits the best properties, it means lower irreducible water saturation and the
slope of the transition zone is almost horizontal in a water saturation Vs capillary
pressure plot. When the rock type indicator decreases the irreducible water saturation
increases, also the slope of the transition zone is sharper. There wasn’t blind data
available for RT1 (green), non-reservoir rock. This kind of result defines a saturation
height function and could be used in order to model initial water saturation and there-
fore contribute to estimate “Original oil in place” (OOIP) as is showed by Marquez,
Aguilar, & Perozo. [14].

Fig. 7. Blind data test results, different colors represent rock quality types. RT6 in yellow is the
best quality rock but RT2 in gray is the worst one. There is not RT1 samples available.

1.5 Permeability calculation

Permeability was calculated using equation 4. Each capillary pressure variable was
used as training data using porosity and shale volume as predictors and propagated as
14

a continuous log, then permeability is calculated solving the equation. In a one -to-one
plot between permeability pred icted vs core permeability, resu lts obtained with super-
vised learning (red points) were superior compared with resu lts obtained using Timur
equation optimized for NMR lo g (b lue points) that was previously used in the reser-
voir Error! Reference source not found.. The slope of the linear regression for su-
pervised learning model is close to one, with lower d ispersion, higher correlation
coefficient, y ield ing to less error. This approach for permeability prediction would be
preferred rather than the previous one based on Timur equation used in the reservoir.

Fig. 8. One to one plot between core permeability and predicted permeability for results ob-
tained with supervised learning (red points) and Timur correlation results (blue points).

If a more simplist ic approach wants to be used a linear correlation between capillary


pressure parameters and permeability could be performed with high correlation coef-
ficient. (Error! Reference source not found.).
15

Fig. 9. Correlation between capillary pressure parameters and permeability (PTR left above,
apex right above and PD below). High correlation coefficients were reached.

The automatic classification was propagated across the well and results were verified
using crossplots and histograms. Shale volume derived by log Vs RHOB log crossplot
is shown. Error! Reference source not found.. A color code is used in order to veri-
fy classification consistency. Please note that better rock types (Yellow and orange)
show low shale volume and low density. Poor rock quality types (green and gray) do
not overlay with the best reservoir rock and exhibits the opposite log behavior (High
shale volume and high density).
16

Fig. 10. Shale volume derived by log Vs rhob log crossplot. A color code is used to illustrate
classification consistency.

Conclusions

•Based on supervised learning approach using capillary pressure data and integration
of all the geological information available it is possib le to identify the optimal number
of rock types in the reservoir under study. Even if the training data set is constrained
to few samples good results could be reached using an appropriate algorithm su ch as
multiresolution graph-based clustering (MRGC).
•Improved permeability pred iction is achieved by means of the application of this
methodology.
•A robust database with all the geological information available (Core descriptions,
image log, XRD, thin section and SEM information) provides the best conditions to
decide if the classification results are accurate. A domain expert is required in o rder to
recognize if the automated process honors geological features, but with supervised
learning this task could be performed time and cost effectively. Furthermore, match
with production from a simulation model is another approach to validate results.
17

References

1. Rabiller, P and Ye, S. A New Tool for Electro-Facies Analysis: Multi-Resolution-


Graph-Based Clustering. SPWLA 41th Annual Logging Symposium, June 4-7,2000.
2. Cuddy, S. The Application of the Mathematics of Fuzzy Logic to Petrophysics.
SPWLA 38th Annual Logging Symposium, June 15-18,1997.
3. Tiab D, Donaldson E. “Petrophysics 2 Ed”. Gulf Professional Publishing” (2004).
4. Thomeer, J.H.M. 1983. Air permeability as a function of three pore-network parame-
ters. Journal of Petroleum Technology, April, p. 809-814.
5. Pittman E. D. “Relationship of porosity and permeability to various parameters de-
rived from mercury injection-capillary pressure curves for sandstone”. AAPG 76 (2)
(1992).
6. Swanson B. F. A simple correlation between permeabilities and mercury capillary
pressures”. JPT, December (1981).
7. Theologou P, East R. D, Curtis A, Boult P, and Drake A. Effective Integration of
the Core Analysis and Wireline Datasets-Using a New Core Analysis Data Suite and
New Software. SPE 53887.
8. Donaldson E.C. et al, “Characteristics of capillary pressure curves”, J. Petr. Sc. &
Eng. 6 (1991).
9. Rabiller, P and Ye, S. The Iterative Use of Clustering and Modeling to Improve Per-
meability Prediction. SPWLA 42nd Annual Logging Symposium, June 17-20, 2001.
10. H.J. Hill; G.E. Klein; O.J. Shirley; E.C. Thomas; W.H. Waxman. Bound Water In
Shaly Sands - Its Relation To Qv And Other Formation Properties. The Log Analyst
20 (Issue 03) 1979.
11. Abbaszadeh, M and Fujii, H. Permeability prediction by hydraulic flow units - Theo-
ry and application. SPE Formation evaluation 1996.
12. Jones, S.C. two Point determinations of permeability and PV vs Net Confining Stress,
SPE formation evaluation, March 1988.
13. Aguilar, C. Supervised learning for rock typing in a siliclastic reservoir. 2021 Interna-
tional Field Exploration and Development Conference in Qingdao, China, 16–18
September 2021.
14. Márquez Codina, E, Aguilar Silva, C and Perozo Lloris, A. Determination of the Ini-
tial Water Saturation Model based on Capillary Pressure Curves by Rock Type, Re-
vista Técnica de la Facultad de Ingeniería. Universidad del Zulia, 44(3), pp. 141-153.
(2021).

You might also like