Feature Ranking and Modeling of Mineral Effects on Reservoir Rock Surface Chemistry Using Smart Algorithms.

Reservoir rock minerals and their surface charge development have been the subject of several studies with a consensus reached on their contribution to the control of reservoir rock surface interactions. However, the question of what factors control the surface charge of minerals and to what extent do these factors affect the surface charge remains unanswered. Also, with several factors identified in our earlier studies, the question of the order of effect on the mineral surface charge was unclear. To quantify the mineral surface charge, zeta potential measurements and Deryaguin-Landau-Verwey-Overbeek (DLVO) theories, as well as surface complexation models, are used. However, these methods can only predict a single mineral surface charge and cannot approximate the reservoir rock surface. This is because the reservoir rock is composed of many minerals in varying proportions. To address these drawbacks, for the first time, we present the implementation of machine learning models to predict reservoir minerals' surface charge. Four different models namely the Adaptive Boosting Regressor, Random Forest Regressor, Support Vector Regressor, and the Gradient Boosting tree were implemented for this purpose with all the model predictions over 95% accuracy. Also, feature ranking of the factors that control the mineral surface charge was carried out with the most dominant factors being the mineral type, salt type, and pH of the environment. Findings reveal an opportunity for accurate prediction of reservoir rock surface charge given the enormous amount of data available.

Introduction

Reservoir rocks are made of several minerals in varying proportions, with these minerals responsible for their surface chemistry and interactions. The effect of the mineral constituents has been a subject of research as many enhanced oil recovery processes are dependent on the change in the constituent mineral properties.[1−9] Thus, it has received enormous attention over the years. Of pertinent interest is the reservoir rock surface charge and how the constituent minerals control the net surface chemistry of the rock. More so how the reservoir environment, as well as operations, affect the reservoir rock surface chemistry. This is emphasized because one of the most challenging flow assurance problems in the reservoir (asphaltene deposition and adsorption) is induced by electrostatic interactions between the reservoir rock and precipitated asphaltene molecules. Thus, to develop a robust strategy to mitigate asphaltene adsorption, the role of the reservoir rocks’ constituent minerals and their surface charge contributions must be well understood.

Several studies have reported the effect of clays,[7,10−15] iron minerals,[1,16−20] with emphasis on how these minerals provide surfaces for asphaltene adsorption. Niriella and Carnahan[21] studied bentonite particles at pH of 4–10 and varying salt solution ionic strength using electrophoretic measurements. The authors report a negative surface charge for bentonite at all pH and all salt concentrations tested. Ma et al.[22] report fluid adsorption on synthetic and natural carbonates using adsorption experiments and surface characterization techniques. Findings reveal that the adsorption was more significant in the natural sample than in the synthetic, thus, a function of the rock mineral constituents. Al-Hashim et al.[23] reports the surface charge modification of calcite and dolomite using Arabian Gulf seawater, and contrary to observations reported by Alotaibi and Nasr-El-Din[24] and Mohammed et al.[3] an all-positive dolomite surface in DI water was observed. These may be due to differences in the rock minerals used or sample preparations. Gopani et al.[8] report the change in the surface charge of carbonate rock samples due to a change in the composition of ions (Ca2+, Mg2+, and SO2–) in brine and their ionic strength. These studies highlight the controlling role the rock mineral constituents play in molecule adsorption.

Recently, we studied the effect of pH, salt type and concentration on clays,[7] iron minerals[25] and another reservoir constituent minerals[3] surface charge. In these studies, zeta potential measurements were used to infer conditions and well operations that provide surface conditions that would promote asphaltene adsorption. Also, mechanisms responsible for surface charge modification of the studied minerals was identified to include adsorption of ionic species on the mineral surface, double layer collapse around the particle and mineral dissolution. Furthermore, trends and correlations between pH, salt type and concentrations were also established to corroborate inference of surface conditions which serve as a precursor for asphaltene adsorption. But even though the mechanism responsible for the observed changes was identified, the degree of significance of each factor and how much contribution to the reservoir rock surface charge was not quantified. This information is critical to the development of enhanced oil recovery (EOR) processes as well as the development of fluids such as inhibitors, and chelating agents.

In the oil fields, an enormous amount of data which includes well logs, core samples, PVT and water samples are taken in the early stage of the fields. This data for example, the core samples provides information about the mineralogy of the reservoir as well as the various fraction of the minerals. Also, the formation water samples provide insights into the chemistry of the brine as well as the pH of the reservoir environment. These pH values change due to different well operations like acidizing, seawater injection, low salinity water injection etc. And with all this information about the fields and reservoirs, the strategy presented in this work explores the use of this information which often lies redundant in data rooms of most companies to infer the surface charge of the minerals around the wellbore. This information can help as a qualitative surveillance strategy to monitor the wettability alterations around the wellbore during the life of a field. Wettability alteration is said to be caused by asphaltene which is defined as a solubility class soluble in aromatics and insoluble in alkanes.[26] Wettability alterations due to asphaltene is been attributed to the adsorption of asphaltene molecules on the mineral surface via electrostatic interactions and hydrogen bonding. So, to mitigate asphaltene adsorption to the surface, conditions that promote its adsorption must be controlled. More so, the process of determining the mineral surface charge is laborious and time-consuming. This has prompted the use of theories like the surface charge complexation models[11,16,27] and the Deryaguin–Landau–Verwey–Overbeek (DLVO) theory[28−33] for surface charge predictions. However, these techniques can only predict individual mineral surface charge but not a mixture of them, thus, does not reflect conditions in the reservoir. Also, the deployment of machine learning tools for reservoir property estimations is gaining acceptance in the industry. Recently, genetic programing was deployed for the thermophysical property estimation of bitumen/light hydrocarbon system, with good correlation achieved.[34]

To address the drawbacks of the existing techniques, for the first time, we present the implementation of machine learning algorithms for mineral surface charge prediction and feature ranking. This will aid the identification of surface charge controlling factors and rank factors according to their contribution to the mineral surface charge. These will answer questions such as what factors control the rock surface charge? In what order is the dominance of these factors? and what well operations affect the reservoir rock minerals the most? The models employed in this research include the Adaptive Boosting Regressor (ADA), Random Forest Regressor (RFR), Support Vector Regressor (SVR), and the Gradient Boosting tree (GBT). These models are used to train, test, and validate predictions using statistical measures applied to measure their performances. The proposed models will serve as a stepping stone to ongoing studies on establishing the relationship between the mineral surface charge and rock dielectric constant which will further be related to mineral surface wettability. So, this study is pivotal to the implementation of machine learning tools to ease rock property predictions based on rock constituent minerals and well operations.

Results and Discussion

Preliminary Data Analysis and Feature Selection

Analysis of the data set using the Turkey methods as earlier stated resulted in no data being eliminated as an outlier and the pair plot of the features is as shown in Figure S1. To bolster the correlation between the features and the target variable, a correlation matrix is presented in Figure . Figure shows the RMSE associated with using the number of features for modeling the target (zeta potential). So, as observed the four input features will be used for the modeling however, this does not answer the question of the degree of importance of the features in relation to the zeta potential. Also, based on the results, Figure shows that there exists a positive and negative correlation between zeta potential and the input variables. In the implementation of the recursive feature selection algorithm, weak features are removed until the optimum number of features is reached. Results presented in Figure , illustrates the RMSE associated with the use of the different number of features. A reduction in the RMSE is observed with an increase in the number of features thus, in our case, the optimum number of features is four which implies that all input variables are to be used.

Figure 2

Heat map depicting the correlation between features and the target variable.

Figure 1

Optimum number of Input variable selection for target variable modeling.

One of the critical questions this study aims to answer is the identification of factors that control the rock mineral surface charge. To achieve this aim, two data sets were utilized. The first includes pH, mineral types, and zeta potential, with the zeta potential being the target variable. The pH in the sense it is used in this research represent different well operations that are implemented.

during the life of producing well namely, acidizing, stimulation, alkaline flooding, and low salinity flooding all of which induce pH change around the wellbore. Figure presents a Pareto chart representing the ranking of features’ effect on zeta potential. From the results of the base data set, it is observed that the feature with the highest effect is the mineral type followed by the pH. This goes to say that the surface charge of a reservoir rock is more dependent on the rock mineral constituents. This observation is congruent with reports by Amroun and Tiab.[35] The authors alluded to the fact that the rock mineral constituents rather than the bulk mineralogy control the rock surface charge. The importance of the reservoir rock mineral constituents cannot be over-emphasized in the determination of rock mineralogy, as 0.01% of a particular mineral can cover the pore surface and control the behavior of interactions.[25] Induced pH change is due to different well operations implemented which affect the constituent mineral surface charge. The impact of these well operations such as nanoparticle injection,[36] low salinity water flooding,[37] viscoelastic surfactant injection[38] etc. results in mineral surface dissolution, adsorption of ionic species and wettability alteration.

Figure 3

Data set 1 (Base case) Pareto chart showing the feature ranking. This shows the degree of importance of pH and Mineral type on the target variable modeling.

The second data set includes features such as pH, Mineral type, salt type, salt concentration, and target variable being zeta potential. The results of the feature ranking of these data set which has 1690 data points are shown in Figure . As can be observed from the figure, the rock mineral type and salt type have the same relative importance which further highlights the importance of these two features relative to other parameters. Next after the effect of both mineral type and salt type is the combined effect of salt type and mineral and the combined effect of pH and mineral. Interestingly, pH on its own has not much effect on the zeta potential of the rock however, its interaction with the minerals is impactful on the rock surface charge. This cannot be farther from the truth as this is the basis upon which acidizing and well stimulation and some surface modifying operations are based. Reservoir brine which has several salt types in varying concentrations provides ions that adsorb on the mineral surface thus, changing its surface charge. This further solidifies our position that attention should be paid to the reservoir rock constituent minerals rather than the bulk mineralogy.

Figure 4

Data set 2 Pareto charts showing the feature ranking. This depicts the degree of influence of salt type, Mineral and a combination of these factors on the target variable prediction.

Also, the rock mineral zeta potential is affected by the combined effect of the pH and mineral type. This is vividly observed in operations such as acidizing and stimulation operations which results in mineral component dissolution leaving positively or negatively charged surfaces. So, findings from the feature ranking involving both data set reveal the importance and order of effect of factors that control the rock mineral surface charge. This also would serve as a pivot for the development of fluids or operations targeted at modifying the reservoir rock surface charge to improve recovery.

Model Development and Evaluation

To examine the relationship between the optimum features identified, four machine learning models were implemented. This includes the RFR, SVR, GDT, and ADA, with their optimum sets of hyperparameters obtained via grid search cross-validation. A summary of the optimum hyperparameters in each model is presented in Table . The predictive performance of all the models using statistical parameters in Table was evaluated and the results are depicted in Table . Based on the correlation coefficient (R2), the mean absolute error (MAE) and the root mean square error, all the models performed well with above 95% R2 values. From the results, RFR and ADA are the best models to be used however, any of the implemented models can be used as all were able to predict the target variable accurately. Furthermore, to corroborate the information shown in Table , plots of the model performances for the training, testing and evaluation are shown in Figures and . Also, the target variable predictions are shown in Figures S2 and S3. What this implies is that with enough data of the conditions of pH, mineral type and salt concentrations, machine learning algorithms can be used to infer the surface chemistry of reservoir rocks accurately. Thus, the accessibility of reservoir rock surface charge using machine learning algorithms usher in a new era of property estimation given available data. This work herein serves as a stepping stone to further work in deploying machine learning tools to predict reservoir surface conditions. Further to this study is the establishment of correlation between the rock mineral zeta potential and dielectric constant which will, in turn, be related to the surface wetting characteristics of the reservoir rock. With these achievements, the control and mitigation of many flow assurance problems can be achieved.

Table 1

Hyperparameters of the Implemented Machine Learning Models

model	parameter	ranges	optimum values
RFR	the no. of trees	10–5000	1000
	the max. depth	10–100	100
	the min. number of samples at a leaf node	1	1
	the min. number of samples required to split an internal node	1–3	2
ADA	base estimator		decision tree
	max. depth of the boosted tree	10–100	40
	the min. number of samples required to split an internal node	1–3	2
	the max. number of estimators at which boosting is terminated	10–200	100
	loss function		exponential
SVR	penalty parameter, C	0.1–5.0	2
	kernel coefficient, gamma	0.1–0.9	0.1
	stopping criteria tolerance	0.001	0.001
	kernel type	RBF and polynomial	RBF
GDT	learning rate	0.01–0.05	0.03
	max. depth of a tree	1–10	2
	the min. sum of the weight of all observations required in a child	0.1–5.0	1
	the fraction of samples observations to be selected for each tree	0.1–2.0	0.95
	L2 regularization weight	0.01–0.1	0.01
	the min. loss reduction required to split an internal node	0.1–0.5	0.1
	the fraction of feature columns to randomly sample for each tree	0.1–0.9	0.75
	no. of boosting rounds	100–10,000	10,000
	early stopping rounds	20	20

Table 3

Statistical Parameters

statistical metric	expression	equation number
R2		(1)
RMSE		(2)
MAE		(3)

Table 2

Machine Learning Models Evaluation

measure	data	RFR	ADA	SVR	GDT
R2	training	1.0000	1.0000	0.9950	1.0000
	testing	1.0000	1.0000	0.9930	1.0000
	validation	1.0000	1.0000	0.9910	1.0000
RMSE	training	0.0000	0.0000	0.0873	0.0169
	testing	0.0000	0.0000	0.1001	0.0161
	validation	0.0000	0.0000	0.1131	0.0156
MAE	training	0.0000	0.0000	0.0677	0.0113
	testing	0.0000	0.0000	0.0775	0.0106
	validation	0.0000	0.0000	0.0792	0.0103

Figure 5

Model performance for training, testing, and evaluation data set (a) SVR and (b) ADA.

Figure 6

Model performance for training, testing, and evaluation data set for (a) GDT and (b) RFR.

The deployment of machine learning algorithms for mineral surface charge prediction ushers in an opportunity to easily monitor surfaces around the wellbore as different well operations are implemented in the field. As demonstrated in this study, the data set covers a wide range of mineral types found in the oil reservoirs thus, data from any of these minerals can be reliably predicted with models as in this study which covers about 12 mineral types. More so, even though good predictions were obtained, a prior understanding of mineral chemistry is critical as models often may not capture the physics of the problems. Thus, the limitation of machine learning tools like these.

Conclusions

The capability of machine learning models in predicting reservoir properties was implemented to address questions of what factors control the reservoir mineral surface charge. Haven successfully implemented different models; the following conclusions are reached based on findings from this research work.

The reservoir mineral surface charge is highly dependent on the rock constituent minerals and brine salt composition.

The well operations (pH) have a limited effect on the zeta potential however a combined effect of the mineral type and pH is significant.

Machine learning models can be successfully deployed to predict reservoir surface charge given enough data.

The RFR, SVR, ADA, and GBT are all machine learning models that can be deployed for reservoir property estimation with high accuracy.

Materials and Method

This section provides information about the methods deployed in this research. Details of the implemented machine learning algorithms can be found else.[39] Thus, this section will focus on the implementation.

Methodology

Data Description

In this study, pH, salt type and concentration, mineral types, and zeta potential data from our earlier published works[3,7,25] were used. Two sets of data sets were obtained from these studies, one which is termed the base case (with variables pH and mineral and zeta potential) and the second (with variables pH, salt type and concentration, mineral and zeta potential). The base case represents the contributions of only the pH and mineral types with the pH representing different pH inducing well operations. Two data sets were used to delineate the contributions of the type and concentration of the salt from the base as it mimics the effect of reservoir native brine and injection of smart or engineered water into the reservoir and their effect on mineral surface charges. The data sets consist of 169 and 1690 data points respectively and it is important to mention that they were divided randomly in the ratio 60:30:10 into training, testing and validation respectively.

Computational Approach

The approach adopted in this study involves cleaning the data set, optimal input selection, and the implementation of different machine learning algorithms. Good machine learning algorithms with high precision and accuracy are dependent on the quality and amount of data as well as the selection of the most correlated features. To ensure good model accuracy data set outliers were eliminated using Tukey’s method. Turkey’s method utilizes the upper (Q3) and lower (Q1) quartile to evaluate each parameter and eliminates data that fall outside the threshold of Q1 + 1.5(Q3 – Q1) and Q1 – 1.5(Q3 – Q1). Also critical to the performance of a machine learning model is the features used in the modeling, as for good model performance, a high correlation between the input and target variables is expected. Thus, feature selection was carried out, however, against the use of conventional technique (correlation coefficient and relevance factor), recursive feature selection was employed. The recursive feature selection approach accounts for the drawback of the conventional technique which is that they only account for the interaction between the target and input variables without consideration to mutual interactions with other features. Details about the recursive feature selection are presented elsewhere.[40,41]

After the feature selection, machine learning models were implemented with their performance evaluated using statistical parameters shown in Table . These include the root mean square error (RMSE), coefficient of determination (R2), and mean absolute error (MAE). In this study, the target is the zeta potential with features including pH, Mineral, Salt Type and Salt Concentration. This study employed different computational libraries in the Python programming framework such as the Pandas, Numpy, Scikit-learn and Matplotlib for the model development and analysis.