Journal of Petroleum Science and Engineering 182 (2019) 106306

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering

journal homepage: www.elsevier.com/locate/petrol

Advanced analytics of self-colmatation in terrigenous oil reservoirs T

*
Denis Orlov , Dmitry Koroteev
Skolkovo Institute of Science and Technology, Skolkovo Innovation Center, Bolshoy Boulevard 30, Bld. 1, Moscow, 121205, Russia

A R T I C LE I N FO A B S T R A C T

Keywords: One of the most significant issues of petroleum engineering nowadays is a decline in productivity and injectivity
Machine learning of wells due to formation damage, which is typically associated with the migration and retention of fines. The
Formation damage prediction of formation damage is a challenging problem due to the lack of information about concentrations
Reservoir properties and localizations of potential mobile fines in the reservoir. To solve this problem different algorithms of machine
Permeability reduction
learning and data mining were tested: linear regressions, decision trees, random forest, gradient boosting and
Fine migration
artificial neural networks. We developed the predictive model of permeability reduction in Vendian deposits
(Russia) based on the analysis of rock properties and flooding conditions. This model allowed to describe the
dynamic of permeability reduction as a multi-parametric function of injected pore volumes of water. Three
defining parameters in the model were unique colmatation characteristics which have been predicted for each
core sample. All the features of the self-colmatation process were studied and arranged by their importance. To
build the model of permeability reduction we used two approaches. The first one was to discover all possible 2D
cross-plot correlations between colmatation characteristics and features (manual analysis). The second is ap-
plying machine learning algorithms where all features were taken into account simultaneously. The benefits and
disadvantages of both approaches were discussed in details.

1. Introduction of the Nepa-Botuoba Arch, which is the second largest dome-shaped

The unpredictable decline in productivity and injectivity of wells
due to formation damage or colmatation is an important problem of
petroleum engineering (Civan (2015); Boek et al. (2012); Muecke et al.
(1979)). The origin of the formation damage mechanisms varies from
fines migration and retention to adsorption processes. Moreover, the
mechanism of colmatation is dependant on an area of interest: near-
well or inter-well zone. In some cases, the colmatation could happen
not only in the near wellbore zone but cover a significant part of a
reservoir.
In this work we continued to investigate the self-colmatation effect
reported in Orlov et al. (2018). The focus of the research is on fines
migration in porous media in the case when inflow fluid is free of any
solid particles. Only the fine particles that are naturally present in
porous media of a reservoir rock (in-situ particles) and can be released
and entrained with the flowing fluid are considered. These entrained
particles can be trapped in pore throats and cause a significant per-
meability reduction (Civan (2015); Boronin et al. (2015)). This type of
formation damage (self-colmatation) is commonly related to flows in
the near-well area of reservoir (Chequer et al. (2018)).
In the work we investigated the Lower Vendian terrigenous deposits
regional high of Eastern Siberia. Water filtration in these reservoir rocks
has a strong self-colmatation feature: entrainment of the in-situ fines
during injection of initially clean water to core samples (Orlov et al.
(2018); Boronin et al. (2017)).
The purpose of this research is to show how machine learning can
help to maximize the effect of experimental data treatment. Machine
learning based on the idea that an intelligent algorithm can use com-
putational methods to “extract” practical information directly from data
without relying on a predetermined equation as a model. The sig-
nificant benefit of these data-driven methods predictive models is that
one may not have to perform experimental measurements or complex
mathematical simulations of phenomena, but can conduct prediction of
the results (Unsal et al. (2005)). Once a predictive model is trained it
could be effectively used for future forecasting.
There are many machine learning algorithms to build a predictive
model: Linear regression, Decision tree, Random forest, Gradient
boosting and Artificial Neural Network Boyd and Vandenberghe
(2004); Breiman (2001); Friedman (2001); Cortes and Vapnik (1995);
Schmidhuber (2015). The choice of algorithm strongly depends on the
problem and data structure and quality.
For example, in Choubineh et al. (2019) an artificial neural network

*
Corresponding author.
E-mail addresses: D.Orlov@skoltech.ru (D. Orlov), D.Koroteev@skoltech.ru (D. Koroteev).

https://doi.org/10.1016/j.petrol.2019.106306
Received 19 February 2019; Received in revised form 23 July 2019; Accepted 24 July 2019
Available online 26 July 2019
0920-4105/ © 2019 Elsevier B.V. All rights reserved.
D. Orlov and D. Koroteev
(ANN) model was applied to estimate the gas-crude oil minimum mis-
cibility pressure for a wide range of injected gases and crude oil com-
positions. The model was correlated with the reservoir temperature,
concentrations of volatile (C1 and N2 ) and intermediate (C2 , C3 , C4 , CO2
and H2 S ) fractions in the oil, C5+ molecular weight fractions in the oil
and injected gas specific gravity. A key benefit of the ANN model was
that minimum miscibility pressure could be determined with reason-
able accuracy for a wide range of oil and gas compositions. Statistical
comparison of predictions showed that the developed ANN model
yields better predictions than empirical correlation methods.
Work Unsal et al. (2005) proposed a methodology to predict pore
size distributions from air permeability measurements combined with a
numerical capillary model representing a porous medium. An optimi-
zation scheme based on the genetic algorithm that makes use of the
measured air permeability values was developed to predict the best
possible pore size distribution and pore arrangement. The evaluation
function was based on the minimization of differences between mea-
sured and calculated air permeability values.
In our experiments, the time-dependent pressure drop across the
core model was measured in 22 lab tests with different core properties
and flow conditions. The primary purpose of the work was to predict
the pressure drop dynamic. We found that pressure drop dynamic could
be simulated with the help of empirical equation. As a result, the pro-
blem statement could be reformulated. To predict the pressure drop
now, we need to predict only three colmatation parameters of the
empirical equation. We estimated these three parameters for 22 lab
tests and used this data as a training set to build the corresponding
predictive models with machine learning (ML).
2. Problem statement and applied methods
Our previous study Orlov et al. (2018) showed that natural cores
from Vendian deposits demonstrate self-colmatation during water fil-
tration even with relatively high (3%) constant salinity which was a
sufficient extension of observations from Lever and Dawe (1984).
Corresponding permeability reduction was registered in cores with a
wide range of porosity, absolute permeability and residual oil satura-
tion. We also performed some tests with a single-phase oil filtration.
The permeability decrease was observed only when water is flowing in
the porous space. Moreover, there were no colmatation in the case of oil
flooding. Water itself was shown to be an active driver of self-colma-
tation, which is likely because of its intense physicochemical reaction
with pore surface.
Experimental procedure is described in Orlov et al. (2018). The self-
colmatation at single-phase water flooding was studied during 22 lab
tests. Each core model was composed of two or three samples with si-
milar porosity, permeability, mineral composition and granulometric
distribution. Experiments conditions are presented in Table 1. The ac-
curacy of the permeability evaluation was equal to 7.1% and determined
by the accuracy of differential pressure gauge (0.5% ), the accuracy of
pumps rate (0.1%) , the accuracy of the core sample sizes estimation
Table 1
Conditions of the experiments.
No. Condition Value
1 model of oil N-decane (C10 H12 )
2 initial saturation residual oil saturation
3 conning pressure 50 MPa
4 pore pressure 12 MPa
5 temperature Ambient (23 °C)
6 pressure drop accuracy 0.5%
7 saturation accuracy 10%
8 NaCl concentration 30 g/l
9 Water viscosity 1 cP
10 Water density 1.02 g/cc
2
Journal of Petroleum Science and Engineering 182 (2019) 106306
(2.4%), and the accuracy of water viscosity evaluation (4.1% ).
In addition to water flooding experiments, some of the core samples
were tested to estimate a grain size distribution, mineral composition
with X-ray diffraction and pore throat size distribution with semi-
permeable membrane method for capillary pressure measurement.
Core samples characteristics we used in the experiments are in
Fig. 1. One can spot a significant variation in porosity (m = 6–21%),
absolute permeability (K = 10–1650 mD) and the residual oil satura-
tion (Sor =30–70%).
The core features do not follow a normal (Gaussian) form of dis-
tribution. It is especially noticeable for permeability data. This fact
reduces the potential of applications of some classical machine learning
techniques. Moreover, because of the complex structure of productive
intervals of Vendian deposits, these core samples do not have strong K-
m correlations. All these constraints of dataset increase uncertainties of
colmatation prediction and analysis.
All water-flooding experiments were performed at residual oil sa-
turations. Development of colmatation was characterized by the growth
of differential pressure after water breakthrough. After injection of
3.5–20 pore volumes of water, the total permeability was reduced by a
factor of 1.02–4.76.
Previous research Orlov et al. (2018) showed that the pressure drop
dynamics demonstrates five typical patterns (Fig. 2). The purpose of the
classification was to define a simple criterion for evaluating the me-
chanism driving retention and correlate it with some rock properties
and flooding conditions. Type I corresponds to the classical no-colma-
tation regime of filtration when pressure drop over the core model
decrease monotonically with water injection (ΔP (t )/ΔP0 < 1, ΔP (t ) -
pressure drop in time, ΔP0 - initial pressure drop). Types II and III
present two boundary cases: unlimited and limited formation damage
respectively (ΔP (t )/ΔP0 ≥ 1). Unlimited (Type II) formation damage
refers to the monotonic rise of ΔP (t ) with increasing growth rates
throughout the constant-rate filtration process. Type III corresponds to
decreasing permeability to a new steady level with a stabilized pressure
drop at some flow rate. Curves of subtype “a” initially have a convex
shape and characterize intensive colmatation at the earliest times of
water injection. Conversely, curves of subtype “b” initially have a
concave shape corresponding to the slow colmatation regime. These
curve shapes are probably determined by the dynamics of fines mobi-
lisation.
Colmatation is a non-steady behaviour of a flooding system with a
time-dependent normalized permeability K (t )/ K 0 . Thus, to estimate
correlations between colmatation, rock properties, and flooding con-
ditions one should find one or more unique colmatation characteristics
which can describe the dynamics of permeability reduction. We found it
more sensible to present K (t )/ K 0 as a function of injected water pore
volumes (Vinj ) and study the function's parameters dependencies on rock
properties and experimental conditions.
There is a widely used approximation representing the normalized
permeability K / K 0 as an inverse function of the retained fine con-
centration (σ) (Zeinijahromi et al. (2013); Gruesbeck et al. (1982)):
K (σ ) 1
= ,
K0 (1 + βσ ) (1)
where β is the formation damage coefficient. If β is large, even a small
retained concentration causes fast permeability reduction. σ in Eq. (1)
can be considered as the concentration of plugged fines. Knowing fines
concentration in the inlet and the outlet flow, it is possible to count σ in
each time moment and predict permeability reduction at a specific
value of β. Unfortunately, at self-colmatation, we do not have enough
information about in-situ fines concentrations. To solve this problem,
we used a method from Bachman et al. (2003), which contains an as-
sumption that σ can be related to the amount of the water (V) that
passed through the cross-sectional area of a porous sample denoted as
A:
D. Orlov and D. Koroteev Journal of Petroleum Science and Engineering 182 (2019) 106306

Fig. 1. Distributions of core parameters.

K / K 0 − Rmin 1 (Kmin = K 0*Rmin ).

= ,
1 − Rmin 1 + α (V / A)n
where α, n, and Rmin are the three governing parameters of the model. α
represents the intensity of the damage. n defines the shape of the da-
mage curve. Rmin sets an asymptotic limit for permeability reduction
(2) We have used Eq. (2) as the basic approximation to investigate
correlations between α, n, and Rmin and rock properties together with
flooding conditions. Results of the best approximation for each ex-
periment are presented in Orlov et al. (2018) and the frequency

Fig. 2. Different types of colmatation. Coloured curves - three experimental dependencies for Type I, IIa and IIIa. On the subplot (left-upper corner) all discovered
pressure drop types are presented schematically.

3
D. Orlov and D. Koroteev
Fig. 3. Distributions of colmatation parameters.
distributions of α, n, and Rmin are in Fig. 3.
3. Colmatation prediction results and discussion
Described approach allows defining dependences of all three gov-
erning parameters (α, n, and Rmin ) on different features of the experi-
ments. We used 16 features for further analysis: porosity m, absolute
permeability K abs , pore volume Vpor , initial oil saturation Sor , pressure
drop type, 3 characteristics of pore throat size distribution, 3 char-
acteristics of grain size distribution, 4 characteristics of mineral content
distribution, and stratigraphical horizon or age of deposits (Talakh,
Khamakin or Botuobinsk). The characteristics of pore throat size dis-
tribution were fractions of pore throats with sizes dp = 6–12 μm, dp
= 12–30 μm, and dp > 30 μm. And the characteristics of grain size
distribution were fractions of grains with sizes dg = 0.05–0.01 mm, dg
= 0.1–0.25 mm, and dg = 0.25–0.5 mm. Contents of microcline, do-
lomite, anhydride, and quartz were used as 4 characteristics of mineral
concentration distribution. Kaolinite in sandstones is the major mineral
causing formation-damage, but we didn't include its content as char-
acteristic of mineral concentration distribution because for all core
samples kaolinite content was constant. So, kaolinite is a weak feature
which does not influence on predictive models training process.
For the first step we have tried to find correlations between each of
α, n, and Rmin and each of 16 features independently (using the cross-
plot charts). To evaluate the accuracy of the methods we have used the
squared Pearson correlation coefficient (R-squared) (Benesty et al.
(2009)):
2
⎛ ∑ (x − m x )(y − m y ) ⎞ that colmatation enhancement occurs when α and n are increasing, and
R2 =
⎜ ∑ (x − m )2 ∑ (y − m )2 ⎟
⎝ x y ⎠ (3)
where m x is the mean of the vector x (real values) and m y is the mean of
the vector y (predicted values).
4
Journal of Petroleum Science and Engineering 182 (2019) 106306
Unfortunately, it was not possible to find any correlations between
colmatation coefficients and horizon types because of a small dataset
and non-Gaussian distribution of features in each type. We could not
find the correlations between α, n, and Rmin and porosity and pore
volume: the R2 was very low in both cases (<0.1). All analytical data
concerning the influence of any of 13 residual features on colmatation
intensity is presented in Orlov et al. (2018) and shortly in Table 2. We
have estimated R2 for Type II and Type III separately. Moreover, within
each type, it has been obtained more tight correlations with a higher
value of R2 .
The intensity of the formation damage α and the exponent n are
increasing, and the factor Rmin is decreasing with the colmatation am-
plification. We assume a weak correlation between colmatation para-
meters and a feature of consideration if we have a mismatch in any
trend direction (if the trend of one of the colmatation parameters
contradicts to the others). Low level of correlation (R2 < 0.1) or a
mismatch in trend direction was marked with bold text in Table 2.
To estimate the influence of 13 parameters on the colmatation
process (feature importance arrangement), we should obtain a single
numerical criterion using R-squared statistics for all three parameters of
colmatation Eq. (2). We used the following rule for averaging:
R2 = 0.5γ (R2TypeII + R2TypeIII ) (4)
where R2TypeII , R2TypeIII - are means with respect to α, n, and Rmin for each
pressure drop Type. γ = 0.66 in case with mismatch in trend direction,
and γ = 1 in case without mismatch. We used γ as an additional
parameter to take into account mismatch in trend direction. We assume
Rmin is decreasing (according to Eq. (2)). If we observed these trends for
α, n, Rmin as a function of some feature increasing or decreasing, we
used γ = 1 (3/3). However, if one of these trends showed the opposite
trend, we used γ = 0.66 (or 2/3). R2 for α, n, and Rmin was estimated
D. Orlov and D. Koroteev Journal of Petroleum Science and Engineering 182 (2019) 106306

Table 2 features. For manual analysis, we excluded horizon types because of its
Analysis of rock properties influence on colmatation parameters. categorical data type. However, ML allows taking this categorical data
Parameter Type R2 for α R2 for n R2 for Rmin into account. To do it, we used “one hot encoding” procedure. One hot
encoding is a process by which categorical variables are converted into
K abs Type II 0.5 0.7 0.52 a form that could be provided to ML algorithms to do a better job in
Type III 0.5 0.7 0.52
prediction. Instead of one feature with three values (Talakh, Khamakin
Sor Type II 0.73 0.007 0.31
Type III 0.73 0.04 0.001
or Botuobinsk) we used three features with two values (0 or 1).
dp >30 μm Type II 0.6 0.8 0.48 Additional 3 features were Talakh, Khamakin and Botuobinsk horizons.
Type III 0.71 0.49 0.15 The pressure drop was excluded from the features because it was post-
dp = 12–30 μm Type II 0.55 0.015 0.11 experimental information (the result of permeability depletion ana-
Type III 0.8 0.46 0.61
lysis). Instead of pressure drop, we built prediction models for α, n, and
dp = 6–12 μm Type II 0.73 0.29 0.25
Type III 0.91 0.35 0.27 Rmin .
dg = 0.25–0.5 mm Type II 0.62 0.67 0.24 A small dataset is an essential problem to train and test the models.
Type III 0.38 0.06 0.63 Another problem for machine learning algorithms is missing data in the
dg = 0.1–0.25 mm Type II 0.42 0.52 0.29
dataset: an ideal case is to have all features for each experiment. The
Type III 0.81 0.39 0.002
dg = 0.05–0.01 mm Type II 0.56 0.49 0.36
grain size distribution, pore throat distribution and mineral composi-
Type III 0.52 0.26 0.14 tion analysis were performed not for all the core samples. The actual
quartz Type II 0.28 0.02 0.78 number of experiments with all 17 features was 5 tests out of 19. Thus,
Type III 0.84 0.59 0.23 the first aim was to extract maximum information from our dataset.
microcline Type II 0.21 0.02 0.71
To fill the missing data in dataset matrix, we applied a sequence of
Type III 0.79 0.63 0.27
dolomite Type II 0.77 0.06 0.4 predictive procedures based on artificial neural network (ANN).
Type III 0.49 0.03 0.05 Implementation of ANN has been done in Python environment. Python's
anhydrite Type II 0.7 0.68 0.27 scikit-learn library provides ANN representation in MLPRegression
Type III * * * method (Pedregosa et al. (2011)). The structure of missing data in-
* not enough data for correlation. cluded three characteristics of pore throat sizes distributions, three
Bold stands for bad correlation (wrong trend or low R-squared). characteristics of grain sizes distributions and four characteristics of
minerals distributions. To predict all three or four characteristics si-
independently for each parameter from corresponding cross-plots. multaneously, we used Multi Output Regressor with from Python's
Based on this analytic approach it was possible to investigate all the scikit-learn library with features normalization by removing the mean
features of the self-colmatation process and arrange them by their im- and scaling to unit variance. In MLP Regression standard values of
portance (Fig. 4). The absolute permeability and the fraction of pore hyperparameters were used: 100 hidden layers, quasi-Newton methods
throats with a diameter larger than 30 μm had the most substantial for weight optimization, L2 penalty (regularization term) parameter -
influence on colmatation. The smallest correlation was for the residual 0.001, learning rate schedule for weight updates - 0.001. Algorithm for
oil saturation and the fraction of grains with size 0.01–0.005 mm. missing data prediction presented on Fig. 5. For missing data predic-
The second approach for investigating the influence of all the fea- tion, we additionally used as a feature porosity, absolute permeability,
tures on the self-colmatation has been dedicated to machine learning pore volume and residual oil saturation. The sequence of the procedures
algorithms. The primary benefit of this approach is a possibility to build is the following. On the first step, we have predicted missing pore throat
complex prediction models of α, n, and Rmin , where all features can be sizes distributions characteristics (for 6 tests) using limited grain sizes
taken into account simultaneously. distributions data (11 tests). On the second step, an extended dataset
For all predictive models, we used 17 features: 14 features described (16 tests) have been used to predict all the missing grain sizes dis-
earlier except pressure drop and horizon types and three additional tributions characteristics (3 tests). On the last step, we have predicted

Fig. 4. The factors affected on colmatation.

5
D. Orlov and D. Koroteev
Fig. 5. Algorithm for missing data prediction. Features - input parameters for ML predictive algorithms. Train - data for models training. Predict - missing data to be
predicted. Averaged - arithmetical mean value among the dataset to be used as a feature when we have missing data for pore throat, grain size and mineral
characteristics together.
characteristics of minerals distributions (for 9 tests) using all data re-
covered on previous steps. Completely recovered dataset matrix with 19
rows (tests) and 17 columns (features) was used then to build effective
predictive models for α and n. Also, for Rmin prediction model, we used
the dataset matrix with 17 rows and 17 columns. Two experiments were
excluded from the dataset in the last case to obtain more tight corre-
lation with absolute permeability in manual analysis. This two experi-
ments could be considered as outliners with R2 = 0.19 and R2 = 0.52
correspondently. To compare manual and intellectual analysis, we used
the same reduced dataset.
We have considered 8 machine learning algorithms: 2 linear re-
gressions with regularization (Lasso and Ridge); Decision Tree; Random
Forest; Gradient Boosting; XG Boosting; SVR (implementation of
Support Vector Machine method) and MLPRegressor (implementation
of ANN) (Boyd and Vandenberghe (2004); Breiman (2001); Friedman
(2001); Cortes and Vapnik (1995); Schmidhuber (2015)). We under-
stand that for a small amount of data case, it is doubtful to use advanced
ML instruments such as Decision Trees and ANN. However, we did not
obtain satisfactory results for more simple regression models. For ex-
ample, for α model R2 = 0.70 in XG Boosting case, R2 = 0.55 in Lasso
linear regression case and R2 = 0.6 in Ridge linear regression case. For
Rmin model R2 = 0.49 in Gradient Boosting case, R2 = 0.22 in Lasso
case and R2 = 0.3 in Ridge case. To evaluate the accuracy of the
methods we have also used the squared Pearson correlation coefficient
R2 . The same metrics have been used previously at independent feature
Fig. 6. Comparison of real and predicted colmatation parameters. (a) - cross-plot of actual and predicted α. (b) - cross-plot of actual and predicted n. (c) - cross-plot of
actual and predicted Rmin .
6
Journal of Petroleum Science and Engineering 182 (2019) 106306
analysis. There is a common practice to avoid overfitting of the model;
to keep a part of the available data as a train and another part as a test
set. The cross-validation technique has been applied to estimate models'
performance and compare them with each other. In Python's scikit-
learn K-Folds cross-validator provides train/test indices to split data
into train/test sets. K-Folds cross-validator splits dataset into k con-
secutive folds. Each fold is then used once as a validation while the k - 1
remaining folds form the training set. In all our cross-validations we
used k = 7 data splits. It corresponds to 16 rows in training and 3 rows
in testing.
The best result (R2 = 0.70) for α model was obtained for XG
Boosting algorithm. To predict α, we used XGBRegressor from Python's
scikit-learn library. XGBoost (Extreme Gradient Boosting) is an opti-
mized distributed gradient boosting library. In XGBRegressor we used
the following hyperparameters: learning_rate (step size shrinkage used
to prevent overfitting) = 0.178; max_depth (determines how deeply
each tree is allowed to grow during any boosting round) = 7; sub-
sample (percentage of samples used per tree. Low value can lead to
underfitting) = 0.91; colsample_bytree (percentage of features used per
tree, high value can lead to overfitting) = 0.51; n_estimators (number
of trees you want to build) = 1960; gamma (controls whether a given
node will split based on the expected reduction in loss after the split, a
higher value leads to fewer splits) = 0.87; alpha (L1 regularization on
leaf weights, a large value leads to more regularization) = 0.3.
Performance of the model could be demonstrated by plotting predicted
D. Orlov and D. Koroteev
values versus observed values of α (Fig. 6a). One can see that data
points are located in the vicinity of the bisector.
The best result (R2 = 0.69) for n model was obtained with Lasso
algorithm (Fig. 6b). Lasso is a linear model trained with L1 norm as
regularizer from Python's scikit-learn library with following hy-
perparameters: alpha (regularization parameter, constant that multi-
plies the L1 term) = 1.7; max_iter (the maximum number of itera-
tions) = 100; tol (the tolerance for the optimization) = 0.015.
The best result (R2 = 0.49) for Rmin model was obtained with
Gradient Boosting algorithm (Fig. 6c). Implementation of Gradient
Boosting was done as GradientBoostingRegressor function in Python's
scikit-learn library. In GradientBoostingRegressor we used the fol-
lowing hyperparameters: learning_rate (learning rate shrinks the con-
tribution of each tree) = 0.41; n_estimators (the number of boosting
stages, large number usually results in better performance) = 20;
max_depth (maximum depth of the individual regression estima-
tors) = 5; max_features (the number of features to consider when
looking for the best split) = 0.3; subsample (the fraction of samples to
be used for fitting the individual base learners) = 0.5. Other hy-
perparameters had default values.
To combine these three predictive models with Eq. (2) we can
predict permeability reduction in time K (t )/ K 0 using only 17 features of
rock properties and flooding conditions for Vendian deposits. The sig-
nificant benefit of the intelligent analysis with machine learning algo-
rithms is the dependence of colmatation parameters on all significant
features simultaneously. It makes the machine learning models more
robust than the 2D cross-plot correlations only. We can also show that
for each α, n, and Rmin coefficient R2 for the machine learning models
will be better (or the same) than for manual analysis (Table 3). In the
manual analysis, we estimated R2 independently for each colmatation
parameter by comparing predictions from one-feature approximations
with real values. As an argument for one-feature approximations, we
used the strongest parameter from Fig. 4 - K abs .
Machine learning models also allow performing feature importance
analysis directly without any assumptions and approximations.
Python's XGBoost method allows arranging features due to their influ-
ence on prediction model (Friedman (2001)). The XGBoost library
provides a built-in function to plot features ordered by their importance
and provides a score indicating how important each feature was in the
construction of the boosted decision trees within the model. Results of
feature importance analysis for α, n, and Rmin are presented in Fig. 7.
Each colmatation parameter has its own most important feature: the
content of microcline for α, the absolute permeability for n, and the
fraction of pore throats with diameters 12–20 μm for Rmin . We now
consider the first five important features for α, n and Rmin . We can see
that 5 of 15 features could be found twice: the microcline content and
the fraction of pore throats sizes >30 μm (in Figure Fig. 7a and b); the
absolute permeability and the fraction of pore throats sizes 6–12 μm (in
Fig. 7b and c); and the porosity (in Fig. 7a and c). Accordingly, these
five features have the dominant influence on the colmatation process.
It is also interesting to compare importance from intelligent and
manual analysis. One can see that four of five features of dominant
influence on the colmatation could be found within the five most im-
portant features from Fig. 4. This fact proves that both approaches do
not contradict and demonstrate close results.
The example of prediction with both approaches is shown in Fig. 8.
Table 3
Models comparison.
Colmatation R2 for strongest R2
coefficient feature (manual analysis) (intelligent analysis)
α 0.62 0.70
n 0.69 0.69
Rmin 0.42 0.49
7
Journal of Petroleum Science and Engineering 182 (2019) 106306
Blue circles represent observational data. Red lines are results of semi-
log approximations. Black crosses are data predicted with ML. One can
see that scattered crosses lay closer to the blue circles then the distance
between them and red lines. The scattered nature of predicted data
could be explained by using in ML models more than one feature.
Comparison between experimental, one-feature approximated and
predicted permeability reduction for three experiments presented in
Fig. 9. One can see that the uncertainty in the permeability reduction
prediction with our model does not exceed 10%. The divergence be-
tween experimental and predicted values of pressure drop or perme-
ability reduction increasing with time or pore volumes injected in the
core. Core parameters used as features in ML algorithm for each ex-
periment are shown in Table 4. Experimental and predicted colmatation
parameters also presented in Table 5.
In some cases (Type IIa and IIIb) the divergence between experi-
mental and predicted values became very significant after a certain
moment of time: usually after injection of more than 15 pore volumes.
It happens because empirical Eq. (2) could not correctly simulate more
than one extremum on ΔP (t ) curve.
4. Conclusions
This work is a continuation of previous experimental research of
self-colmatation in Lower Vendian terrigenous deposits of Eastern
Siberia. The focus of the research is to predict the dynamics of per-
meability reduction in particular core sample using only the results of
routine and special core analysis:
– porosity (and pore volume),
– permeability,
– residual oil saturation,
– grain size distribution,
– mineral composition,
– pore throat distribution,
– and stratigraphic horizon.
The dynamics of permeability reduction was accounted in multi-
parametric function of injected pore volumes of water (Eq. (2)). Thus
the problem was reduced to the prediction of only three colmatation
characteristics α, n, and Rmin .
We have used two approaches to build the prediction models for α,
n, and Rmin :
– to discover all possible 2D cross-plot correlations between colma-
tation characteristics and 17 features (manual analysis),
– to build complex multi-feature models based on machine learning
algorithms.
The applicability of various Machine Learning algorithms to build a
relevant model for the prediction of some colmatation properties were
tested.
The main problem to build relevant predictive model was a small
dataset to train and test the models: only 22 laboratory tests of single-
phase water flooding in reservoir conditions at residual oil saturation.
An additional problem for applying machine learning algorithms was
missing data in the dataset (some features were absent for some tests).
To fill the missing data in dataset matrix, we applied a sequence of
predictive procedures based on ANN.
We have considered 8 machine learning algorithms for formation
damage prediction: 2 linear regressions with regularization (Lasso and
Ridge); Decision Tree; Random Forest; Gradient Boosting; XG Boosting;
SVR (implementation of Support Vector Machine method) and
MLPRegressor (implementation of ANN). The best performance de-
monstrated:
– XG Boosting algorithm for α prediction (R2 = 0.70),
D. Orlov and D. Koroteev Journal of Petroleum Science and Engineering 182 (2019) 106306

Fig. 7. Results of feature importance analysis for α (a), n (b), and Rmin (c).

Fig. 8. Results of colmatation prediction: approximation and ML prediction. α (a), n (b), and Rmin (c) dependencies on absolute permeability.

8
D. Orlov and D. Koroteev Journal of Petroleum Science and Engineering 182 (2019) 106306

Fig. 9. Permeability reduction: experimental (blue dots), one-feature approximated (dash black line) and ML predicted (solid black line) data. (a) - sample A. (b) -
sample B. (c) - sample C. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Table 4 Table 5
Conditions of the experiments. Colmatation parameters for samples A, B and C.
Parameter Sample A Sample B Sample C Sample Parameter Experiment One-feature approx. ML prediction

m, % 12.5 11.1 17.6 A α 11 22.8 13.8

K abs , mD 93 860 220 n 1 1.11 1.14
Vpor , cm3 7.697 7.148 7.209 Rmin 0.35 0.59 0.49
Sor , % 47 50 42 B α 50 42.8 42.1
dp >30 μm , % 14.7 36.3 13.7 n 1.8 2 2
Rmin 0.5 0.27 0.38
dp = 12–30 μm , % 4.1 47.4 31.3
C α 16 30.7 24
dp = 6–12 μm , % 25.9 6.5 38.3
n 1.1 1.26 1.29
dg = 0.25–0.5 mm , % 12.9 20 57.4 Rmin 0.5 0.46 0.41
dg = 0.1–0.25 mm , % 54.2 9.95 15.7
dg = 0.05–0.01 mm , % 11.5 3.35 1.3
microcline, % 34.1 3.2 13.2 traditional 2D cross-plot analysis - colmatation parameters depend on
dolomite, % 2.1 1.5 0.2
all 17 features simultaneously. The R2 metric was higher for ML algo-
anhydrite, % 0 0.65 0.9
quartz, % 60.7 94.65 85.2
rithms than for manual analysis.
str. horizon Talakh Khamakin Botuobinsk All the features of the self-colmatation process have been arranged
by their importance using two independent methods (cross-plot corre-
lation estimations and joint XGBoost analysis). Both approaches do not
– Lasso algorithm for n prediction (R2 = 0.69), contradict each other and demonstrate close results. The most influ-
– and Gradient Boosting algorithm for Rmin prediction (R2 = 0.49). encing features are:

It is shown that machine learning algorithms allow building a little – permeability,

bit more precise prediction model of permeability reduction than – porosity,

9
D. Orlov and D. Koroteev
– fractions of pore throats with diameters 6–12 μm ,
– fractions of pore throats with diameters > 30μm ,
– and content of microcline minerals.
Fractions of pore throats with diameters 6–12 μm and > 30μm cor-
related with characteristic sizes of fines defining the self-colmatation in
Vendian deposits. Having these mean sizes of fines we can estimate the
mean value of the jamming ratio factor (ratio between mean fines size
and mean pore-throat size) and use it as a single characteristic ac-
counting for both pore-throat size distribution and particle size dis-
tribution and their influence on formation damage.
Finally, this work showed that machine learning methods could be
successfully applied for the prediction of flow behaviour when experi-
mental data is not available.
Appendix A. Supplementary data
Supplementary data to this article can be found online at https://
doi.org/10.1016/j.petrol.2019.106306.
References
Bachman, R., Harding, T., Settari, A.T., Walters, D., et al., 2003. Coupled simulation of
reservoir flow, geomechanics, and formation plugging with application to high-rate
produced water reinjection. In: SPE Reservoir Simulation Symposium. Society of
Petroleum Engineers.
Benesty, J., Chen, J., Huang, Y., Cohen, I., 2009. Pearson correlation coefficient. In: Noise
Reduction in Speech Processing. Springer, pp. 1–4.
Boek, E.S., Hall, C., Tardy, P.M., 2012. Deep bed filtration modelling of formation damage
due to particulate invasion from drilling fluids. Transp. Porous Media 91 (2),
479–508.
Boronin, S., Osiptsov, A., Tolmacheva, K., 2015. Multi-fluid model of suspension filtration
10
Journal of Petroleum Science and Engineering 182 (2019) 106306
in a porous medium. Fluid Dyn. 50 (6), 759–768. https://doi.org/10.1134/
S0015462815060058.
Boronin, S., Tolmacheva, K., Osiptsov, A., Orlov, D., Koroteev, D., Sitnikov, A., Yakovlev,
A., Belozerov, B., Belonogov, E., Galeev, R., et al., 2017. Modelling of injection well
capacity with account for permeability damage in the near-wellbore zone for oil
fields in western siberia (Russian). In: SPE Russian Petroleum Technology
Conference. Society of Petroleum Engineers. https://doi.org/10.2118/187806-RU.
Boyd, S., Vandenberghe, L., 2004. Convex Optimization. Cambridge university press.
Breiman, L., 2001. Random forests. Mach. Learn. 45 (1), 5–32.
Chequer, L., Vaz, A., Bedrikovetsky, P., 2018. Injectivity decline during low-salinity
waterflooding due to fines migration. J. Pet. Sci. Eng. 165, 1054–1072.
Choubineh, A., Helalizadeh, A., Wood, D.A., 2019. Estimation of minimum miscibility
pressure of varied gas compositions and reservoir crude oil over a wide range of
conditions using an artificial neural network model. Adv. Geo-Energy Res. 3 (1),
52–66.
Civan, F., 2015. Reservoir Formation Damage. Gulf Professional Publishing.
Cortes, C., Vapnik, V., 1995. Support-vector networks. Mach. Learn. 20 (3), 273–297.
Friedman, J.H., 2001. Greedy function approximation: a gradient boosting machine. Ann.
Stat. 1189–1232.
Gruesbeck, C., Collins, R., et al., 1982. Entrainment and deposition of fine particles in
porous media. Soc. Pet. Eng. J. 22 (06), 847–856.
Lever, A., Dawe, R.A., 1984. Water-sensitivity and migration of fines in the hopeman
sandstone. J. Pet. Geol. 7 (1), 97–107.
Muecke, T.W., et al., 1979. Formation fines and factors controlling their movement in
porous media. J. Pet. Technol. 31 (02), 144–150. https://doi.org/10.2118/7007-PA.
Orlov, D., Koroteev, D., Sitnikov, A., 2018. Self-colmatation in terrigenic oil reservoirs of
eastern siberia. J. Pet. Sci. Eng. 163, 576–589.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M.,
Prettenhofer, P., Weiss, R., Dubourg, V., et al., 2011. Scikit-learn: machine learning in
python. J. Mach. Learn. Res. 12 (Oct), 2825–2830.
Schmidhuber, J., 2015. Deep learning in neural networks: an overview. Neural Netw. 61,
85–117.
Unsal, E., Dane, J., Dozier, G.V., 2005. A genetic algorithm for predicting pore geometry
based on air permeability measurements. Vadose Zone J. 4 (2), 389–397.
Zeinijahromi, A., Nguyen, T.K.P., Bedrikovetsky, P., et al., 2013. Mathematical model for
fines-migration-assisted waterflooding with induced formation damage. SPE J. 18
(03), 518–533.