Determination of petrophysical rock types and permeability using machine learning methods in a heterogeneous reservoir, southwest of the Iran

Introduction
The study and evaluation of hydrocarbon reservoirs, as vital arteries for energy supply in today's world, are of paramount importance. Petroleum engineers and geologists are constantly striving to understand these complex subsurface systems more accurately and comprehensively. These efforts are particularly significant for carbonate reservoirs, due to their unique characteristics and specific challenges. Carbonate reservoirs, owing to complex diagenetic processes, inherent heterogeneities, and extensive fracture structures, are among the most intricate types of hydrocarbon reservoirs. These complexities make accurate evaluation of petrophysical properties and prediction of their production behavior a major challenge. In this context, identifying and determining petrophysical rock types and their characteristics, including porosity, permeability, fluid saturation, and pore size distribution, play a key role in understanding reservoir behavior and optimizing production processes. Rock types, as fundamental building blocks of the reservoir, exhibit relatively homogeneous petrophysical properties within a defined volume of rock. Identifying and differentiating these types enables more accurate reservoir modeling and prediction of its behavior under different production conditions. However, precise determination of rock types in carbonate reservoirs, due to the high diversity of textures, structures, and diagenetic processes, requires the use of advanced and integrated methods. Traditional reservoir evaluation methods rely primarily on core data and well log information. Core data provides valuable information about the physical and chemical properties of the reservoir rock, but its preparation and analysis are costly and time-consuming, and it is usually limited to a small number of wells in the oil field. Well log information provides broader coverage in the field, but its interpretation requires specialized knowledge and experience, and its accuracy may be affected by various factors. For this reason, the use of modern techniques such as machine learning and clustering has received increasing attention as a powerful tool for analyzing reservoir data and extracting valuable information from it.
In this study, with the aim of overcoming the limitations of traditional methods and improving the accuracy of carbonate reservoir evaluation, self-organizing maps (SOM) have been used to cluster well log data and identify electrofacies. Using unsupervised learning algorithms, this method is able to identify hidden patterns in the data and classify similar data into separate groups. Electrofacies, as distinct petrophysical units in the reservoir, exhibit relatively uniform log characteristics and can be considered as representatives of rock types. By matching the identified electrofacies with core data and geological information, a more accurate model of the distribution of rock types in the reservoir can be created and its petrophysical properties can be estimated with greater precision. The ultimate goal of this study is to provide an efficient and reliable method for evaluating carbonate reservoirs using machine learning techniques and improving the accuracy of predicting their production behavior.
Materials and Methods
In this study, an integrated approach comprising machine learning-based clustering methods and artificial neural networks was employed to determine petrophysical rock types and estimate permeability in the Bangestan reservoir, located in southwestern Iran. This approach, utilizing well log data and core information, enables more accurate and efficient identification of reservoir characteristics.
The dataset used in this research includes information from nine wells in the Ahvaz oil field. Among these, five wells have core data (including porosity and permeability information) and have been used as reference wells for training and validating the models. The well logs used in this study include density (RHOB), neutron (NPHI), effective porosity (PHIE), sonic transit time (DT), and gamma (GR) logs. These logs, due to their wide coverage and sensitivity to petrophysical changes, have been selected as the main inputs for clustering and permeability estimation algorithms.
To determine petrophysical rock types, the Self-Organizing Maps (SOM) clustering method was used. This method, using an unsupervised learning algorithm, is able to classify similar data into separate groups. In this study, well log data from reference wells were input into the SOM network, and after training the network, the data were divided into 25 initial clusters. Then, by analyzing the petrophysical characteristics of each cluster and matching them with core information and hydraulic flow units, similar clusters were merged and, finally, five distinct petrophysical rock types were identified.
Hydraulic flow units, as a criterion for evaluating reservoir quality, were determined using the logarithm of the flow zone indicator (Log FZI) method. This method, using core porosity and permeability data, enables the separation of flow units with different hydraulic characteristics. Matching the identified electrofacies with hydraulic flow units, as a validation method, helped ensure the accuracy and precision of the clustering.
To estimate permeability in the studied reservoir, artificial neural networks (ANN) were used. This method, using a supervised learning algorithm, is able to learn the relationship between input data (well logs) and output data (core permeability) and, based on that, estimate permeability in other parts of the reservoir. In this study, a multi-layer perceptron neural network with a hidden layer was used. The reference well data were divided into training and testing sets. The training set was used to train the network and adjust its weights, and the testing set was used to evaluate the network's performance and determine the accuracy of permeability estimation.
Permeability was estimated in two separate ways: 1) permeability estimation for the entire reservoir interval regardless of rock types, and 2) permeability estimation for each of the identified rock types separately. Comparing the results of these two methods allows for evaluating the impact of data clustering on the accuracy of permeability estimation.
To evaluate the performance of the clustering and permeability estimation models, various statistical measures were used. To evaluate clustering accuracy, the Silhouette index was used, and to evaluate permeability estimation accuracy, the correlation coefficient (R) and root mean squared error (RMSE) measures were used. These measures enable comparison of the performance of different models and determination of the best model for permeability estimation in the studied reservoir.
 
Results and Discussion
In this study, we successfully identified five distinct petrophysical rock types in the Bangestan reservoir using machine learning methods. These rock types were accurately determined using the SOM clustering algorithm and matching with core data and hydraulic flow units. The clustering results showed that each of these rock types has unique petrophysical characteristics that affect fluid flow behavior in the reservoir. Rock types 1 and 2 had the best reservoir quality, rock type 3 had the medium reservoir quality and rock types 4 and 5 had the lowest reservoir quality.
Matching the identified electrofacies with hydraulic flow units (FZI) showed a high correlation between the two. This correlation indicates that the SOM clustering method was well able to separate flow units with different hydraulic characteristics.
The results of permeability estimation using artificial neural networks (ANN) showed that this method is able to estimate permeability with acceptable accuracy. Comparison of permeability estimation results with core data showed that the correlation coefficient (R) between the estimated values and the actual values is around 0.9804. Also, the root mean square error (RMSE) is around 0.0778.
Comparing the results of permeability estimation for the entire reservoir interval with the results of permeability estimation for each of the rock types separately showed that data clustering has a positive effect on the accuracy of permeability estimation. In other words, permeability estimation for each of the rock types separately has higher accuracy than permeability estimation for the entire reservoir interval. This result 
shows that considering the petrophysical characteristics of each of the rock types can help improve the accuracy of permeability estimation models.
The results of this study show that the use of machine learning methods can help improve the accuracy and efficiency of carbonate reservoir evaluation. The SOM clustering method, as a powerful tool for identifying petrophysical rock types, enables more accurate reservoir modeling and prediction of its production behavior. Also, artificial neural networks (ANN), as an efficient method for estimating permeability, enable quantitative evaluation of reservoir characteristics and optimization of production processes.
However, it should be noted that the results of this study are limited to the Bangestan reservoir in the Ahvaz oil field and may not be generalizable to other carbonate reservoirs. To generalize the results of this study to other reservoirs, more studies and examination of data related to those reservoirs are needed.
 
Conclusion
In this study, an integrated machine learning-based approach was presented for identifying petrophysical rock types and estimating permeability in the Bangestan carbonate reservoir. Using the SOM clustering algorithm, five distinct rock types were identified, each with unique petrophysical characteristics and flow behaviors. ANN models, trained separately for each rock type, were able to estimate permeability with acceptable accuracy. The results showed that data clustering and considering the petrophysical characteristics of each rock type significantly improved the accuracy of permeability estimation. This approach can be used as an efficient tool for evaluating carbonate reservoirs and optimizing production processes.
This study significantly enhances our understanding of the Bangestan reservoir characteristics and can serve as a foundation for developing more advanced models in hydrocarbon reservoir evaluation. The findings of this research may also contribute to optimizing production processes and managing oil resources, paving the way for future studies in this field.