Electrofacies Modelling and Lithological Classification of Coals and Mud-bearing Fine-grained Siliciclastic Rocks Based on Neural Networks

The identification of lithofacies from well is usually an interpretative process based on geophysical logs since core and sidewall samples are not usually available. Despite being always sampled and described, cuttings are useful only as a reference for determining the rocks because a number of problems occur during the drilling and sampling activities. Well logs are in situ continuous records of different physical properties of the drilled rocks, which can be associated with different lithofacies by experienced log analysts. This task needs a relatively great amount of time and it is likely to be imperfect because the human analysis is subjective. Thus, any alternative method of classification with high accuracy and promptness is very welcome by the log analysts. This paper is based on Neural Networks (NNs) applied in well data from the Leão Coal Mine, southern Brazil, in order to classify organic mudrocks, coals and siliciclastic sandstones, the main rocks present in the Rio Bonito and Palermo formations, by using their well logs as database. The training and validation set of the NN contain data from eight cored and logged boreholes. The input included 409 values of depth and logs of gamma-ray, spontaneous potential, resistance and resistivity for each electrofacies. The neural network model was the feedforward multilayer perceptron (MLP) and the neural networks were trained with variations of the backpropagation algorithm: Levenberg-Marquardt and Resilient backpropagation. Although an accuracy of approximately 80% had been achieved in the general classification, discrepant accuracies in the classification of the different electrofacies are discussed in order to better understand the reasons that affected negatively the NN performance.


Introduction
The rock sampling in a well which crosses intervals that bear mineral resources of any nature (e.g.: water, oil, gas, coal, metals, etc.) is essential for understanding its genesis and the subsequent activities of research and/or exploitation.Drill-cores and sidewall samples are obtained along intervals of interest, taking into account that continuous cores and the sidewall samples are punctual.They are used for determining the type of rock as well as for quantifying a number of properties through laboratory analyses.Nevertheless, very high costs restrict these kind of samples from being obtained throughout the well.Cutting samples are continuously recuperated, but they do have limitations regarding the correct positioning and reliable representation of the rock in the drilled succession.
As a way to increase the feasibility of continuous sampling in subsurface, the measuring of several properties by means of sensors is a usual procedure, like natural radioactivity, electrical conductivity, transit time of sound waves, among others.The measuring of properties is part of the logging operations, usually carried out throughout the well.The values obtained in these operations are processed and transformed into continuous records of variations along the well, named geophysical logs, well logs or electrical logs.
Well logs have become more accurate as technology developed and presently allow initial analysis of many aquifers and oil fields to be done based only on the log interpretation.In order to fill the lack of rock samples and to avoid harming the assessment on the interval of interest, combined log analysis is a common practice.It allows estimate, for instance, the kind of rock, to recognize permeable layers, to identify intervals bearing hydrocarbons and to determine the percentage of rock porosity.
For many decades, well log analysis was done manually and visually for the identification of rocks.The identification was carried out mainly by means of cutoffs, that is, intervals with values previously determined for a given kind of rock in one or more logs (e.g., shales and sandstones using gamma ray).The use of such method required an experienced log analyst who could relate the standards obtained to the different lithological types.Although good enough, that method demands a relatively great amount of time for the analysis of one single well or interval.It is also subject to different criteria of standard recognition as well as it is limited to human capacity to simultaneously relate data that may vary in nature.On the other hand, many researchers have developed different approaches in order to automate the process in the last years.A number of techniques has been used, such as statistical (Busch et al., 1987;Flexa et al., 2004;Sancevero et al., 2008), fuzzy logic (Toumani et al., 1994;Saggaf & Nebrija, 2003a;Ilkhchi et al., 2006) and neural network (Baldwin et al., 1990;Rogers et al., 2002;Bhatt & Helle, 2002;Saggaf & Nebrija, 2000, 2003b;Maiti et al., 2007;Schmitt, 2009;Maiti & Tiwari, 2010;Raeesi et al., 2012).Specific applications of neural networks for better understand organic and fine-grained siliciclastic rocks from well logs are rare (e.g., Huang & Williamson, 1996;Yang et al., 2004;Schmitt, 2009).Thus, a selected dataset was chosen for testing the neural network accuracy in differentiating these types of rocks.This paper aims to recognize electrofacies patterns and link them specifically to organic and fine-grained siliciclastic rocks of the Rio Bonito and Palermo formations at the Leão Coal Mine through neural network.The results were based on 28% of training data and 72% of validation data and allowed a discussion of the variable indexes of the network accuracy, since the well logs intervals of training and validation are cored and the original descriptions were used to compare with the results.Besides, the lithological features of the working site challenge the achievement of a successful quality of training and validation and also challenge the results of the lithofacies classification as a consequence.Although with a general accuracy of approximately 80% in the validation phase, a number of problems were identified and discussed in order to better understand the quality of the results.

Neural Networks and Electrofacies Modelling
Neural Networks (NN) are groupings of processing units, called neurons or nodes, structured and interconnected, whose functionality is similar to a neural structure of intelligent organisms.The NN have a high computational power due to its parallel and distributed structure and its capacity to learn and/or make generalizations, what makes it possible to solve complex problems in a vast range of scientific knowledge (Haykin, 1999).
In view of its non-linear structure, the NN is capable to capture the most complex characteristics from the data, which is not always possible if one uses the traditional statistical techniques or other deterministic methods.The greatest advantage of neural networks over conventional methods, such as the statistical one, is to carry out an analysis without knowledge of the intrinsic theory of the matter.Other great advantages are to analyze relations which are not fully known among the variables involved in the modelling and use a well-established technique for application in well log data.
The use of numerical-computational methods for the classification of rocks based on well logs started in the early 70's and it was boosted from the late 80's.Those first works were based on numerical (Harris & McCammon, 1971;Doveton & Cable, 1979) and statistical methods, especially the principal component analysis (Wolff & Pelissier-Combescure, 1982) and the discriminant analysis (Busch et al., 1987;Delfiner, 1987).The classification of rocks based on artificial intelligence methods applied to well logs came up in the early 90's, following the works that had been carried out on neural networks (Derek et al., 1990;Baldwin et al., 1990;Rogers et al., 1992;Benaouda et al., 1999;Zhang et al., 1999;Saggaf & Nebrija, 2000;Siripitayananon et al., 2001;Bhatt, 2002;Bhatt & Helle, 2002;Yang et al., 2004;Maiti et al., 2007;Maiti & Tiwari, 2010) and fuzzy logic (Toumani et al., 1994;Xuyan & Zhanglong, 1998;Saggaf & Nebrija, 2003a;Hsieh et al., 2005).The basic principle is to identify the log patterns associated with facies or facies associations in different well logs, named electrofacies, and structure a neural network in order to recognize such patterns in all the succession or study area.Saggaf and Nebrija (2000) performed a supervised analysis to identify rocks previously identified from cores.The automatic method proposed by the authors was based on competitive neural network and had similar results to those performed by experienced log analysts.Furthermore, the method provides a way to measure the degree of confidence of the analysis.Another application on electrofacies modeling estimated missing logs in wells with a incomplete suite (Saggaf & Nebrija, 2003b).The authors recognized the interdependence among different logs using as example a well with a complete log suite and applied a neural network to estimate the missing logs.The high accuracy of the results was evaluated by blind tests in which both estimated log and the real one were compared.An interesting application of neural networks was performed by Yang et al. (2004).They estimated the clay content, grain density, total organic carbon and cement using mudstone and carbonate values obtained from well logs.

Resilient Backpropagation and Levenberg-Marquardt Algorithms
The Resilient Backpropagation (Rprop) is an algorithm that performs batch supervised training in multilayer perceptron-like networks.This algorithm works in order to eliminate the negative influence of the partial derivative value in the weight adjustment.This influence occurs because the output value of a neuron of approximately 0 (or 1) and the expected output of 1 (or 0) imply in a derivative of approximately 0. Thus, the weight for this neuron will be minimally adjusted (Braga et al., 2007).The Rprop is capable to eliminate this problem using just the signal of the derivative, not its value.The signal indicates the direction of the weight adjustment, either increasing or decreasing the previous weight.The range of the weight adjustment is given by the "actualization value" ji t  , as shown by Equation 1.
The is defined as an adaptation process dependent of the signal of the error derivative in relation with the weight to be adjusted, as indicated by Equation 2.
(2) Where: 0 < < 1 < According with the rule of adaption used by Rprop, when the partial derivative of the error in relation to the weight keeps the same sign (indicating that the last adjust decreased the error), the actualization value increases by the factor and speeds up the training convergence.When the partial derivative changes the sign (indicating that the last adjust was too much), the actualization value decreases by the factor and changes the direction of adjustment.
The Levenberg-Marquardt algorithm is an approximation of the Newton's method.It improves the Gauss-Newton method by using a variable learning rate and was proposed to adjust the weights of the neural network after each cycle.This algorithm is a numerical optimization technique of high computational complexity which spends a great volume of memory, which might impede its utilization in huge networks (Haykin, 1999).
When this algorithm is applied, the weights of the network are adjusted in accordance with the Equation 3. Where: is a Hessian matrix and is the gradient.The parameter µ is multiplied by the factor β when an adjust intends to increase .When the purpose is decrease the value after each step, µ is divided by β.When µ is too large, the algorithm becomes a descendent gradient, with step 1/ µ.When µ is low, the algorithm is equivalent to the algorithm Gauss-Newton (Braga et al., 2007).
The conjugate gradient algorithms are usually much faster than variable learning rate backpropagation, and are sometimes faster than resilient backpropagation, although the results vary from one problem to another.The conjugate gradient algorithms require only a little more storage than the simpler algorithms.Therefore, these algorithms are good for networks with a large number of weights (Demuth et al., 2008).

Database and Methods
In the region of the Leão Coal Mine there is the occurrence of granitoids of Proterozoic age that constitute the basement of the Paraná Basin in that region.Those granitoids are covered by a Paleozoic sedimentary sequence which comprises the Itararé Group and the Rio Bonito, Palermo, Irati, Serra Alta, Teresina and Rio do Rasto formations (Figure 1).That succession is remarkable for its coal beds in the Rio Bonito Formation, which constitute the main Brazilian coal deposits (Gomes et al., 2003).
The database is composed of well logs and lithofacies descriptions of the cores drilled in the above-mentioned coal mine.The site is located in mid-eastern Rio Grande do Sul State (Figure 2), with a distribution in the counties of Rio Pardo, Minas do Leão, Butiá, São Jerônimo, General Câmara and Vale Verde.Despite coal mining occurs in the area since the 1940's, the National Department of Mineral Production (DNPM) and the Brazilian Geological Survey (CPRM) carried out major mineral research projects in this area between 1975 and 1983.The data come from the project named Iruí-Butiá (CPRM, 1981), with drilled in meshes from 8x8km to 1x1km (Lopes, 1990).
The logs that were used for the training, validation and classification were analyzed and selected based on the borehole and description files from the Porto Alegre's Regional Superintendence Office of the CPRM.The eight boreholes chosen for this study were LB-132-RS, LB-137-RS, LB-138-RS, LB-139-RS, LB-141-RS, LB-143-RS, LB-148-RS and LB-207-RS.
The conversion of logs from the analogue to the digital format was done by means of the original logs printed on millimeter paper.Raster images were carefully acquired in scanners and imported onto a graphic software in order to adjust the scale and perform the well log vectorization.The vectorized logs were exported as a text file (.txt) so as to relate depth values to the gamma ray values (GR), spontaneous potential (SP), resistivity (RTV) and resistance (RTC).The data set was organized in a data bank and loaded onto the Matrix Laboratory application, MatLab, afterwards for the initial stages of the neural network structuring.The data set was normalized in interval [0;1] considering the maximum and minimum values.The training dataset comprised 28% of the data, with its respective complementary set of validation of 72%, having the latter been developed with basis on information about three wells, namely LB-137-RS, LB-138-RS and LB-148-RS.The small percentage of training data is related with the small number of the electrofacies 3 available for training purposes.As training of a neural network must consider the same number of samples for each electrofacies, the training data set was limited in 28%.To assess the performance of NN in the classification, an analysis of percentage of standards that were correctly classified, incorrectly classified, and rejected along the validation phase is done.The proportion of correctly classified standards is named accuracy.A classification is correct when the classifier matches an "unknown" standard with its true class (Braga et al., 2007).When the classifier places the "unknown" standard in the wrong class, the classification is said to be incorrect.The proportion of incorrectly classified standards gives origin to an error rate.In case the standard is similar to the standards of more than one class or is not similar enough to the standards of none of the classes, the unknown standard is then rejected, being designated as non-identified.
A simple technique that is commonly used is to present the examples to the multi-layer perceptron in a random order in between two seasons.Ideally, this random order guarantees that the examples that were successively presented to the network in one season rarely belong to the same class (Haykin, 1999).
The accuracy of the classification was checked by the Confusion Matrix (Tables 1, 2 and 3).It is a resource that offers an effective measure of the classification model by showing the number of correct classifications opposed to the classifications predicted for each class, about a group of training samples of each class.The number of successes in each class is located in the main diagonal of the matrix.All the other elements represent errors of omission and commission in the classification (Stehman, 1997).
160 trainings were carried out until a network topology was reached and a one-percent error acceptance was set (see Appendix 1).Network input parameters were organized in the following way: one network was trained with five variables (depth, logs GR, SP, RTC, and RTV) and another one with four: depth, logs GR, SP and RTC.In all the experiments, the output had three neurons: one to represent electrofacies 1, another to represent electrofacies 2, and another one to represent electrofacies 3. The neurons work based on a non-linear function, named activation function, which determine the output data using the input data.The main activation functions are Logistic Sigmoid, Linear and Hyperbolic Tangent.The activation functions used were the linear and the logistic sigmoid function (LOGSIG), whereas the tested learning algorithms consisted of variations in the backpropagation algorithm, such as Levenberg-Marquardt, Resilient Backpropagation and Scaled Conjugate Gradient.
In the process of validation, we have tried to analyze the NN's efficiency by means of three logged and cored boreholes (LB-137-RS, LB-138-RS and LB-148-RS) with the known lithological descriptions that did not take part in the training stage.The validation set was composed of 497 samples of sandstone, 2411 of siltstone, and 263 of coal, giving a total of 3171 values.In order to assess the ability of the network classification, the complete sections of drilling that were used in the validation stage were plotted.This was done by comparing the original description with the outputs of the neural network.

Results and Discussions
The boreholes LB-138-RS and LB-148-RS had the best performance of learning with the topology [5-15-3-3] and as variables the logs gamma rays (GR), spontaneous potential (SP), resistivity (RTV) and resistance (RTC).
The activation function used in all layers was the logistic sigmoidal, the algorithm was the Resilient backpropagation and the learning finished after 4000 cycles.
The success in the classification rate for the borehole LB-138-RS reached 76% and was the minor used in the validation stage.Electrofacies 2 reached 88,3% of success in its classification while electrofacies 1 and 3 had 51,8% and 16,5%, respectively (Table 1).Although incorrectly classified as electrofacies 1 by the neural network, the interval between 219,9 a 220,5m has a number of medium-grained sandstone intercalations within a bed of siltstone.On the other hand, the interpretation of a thinner sandstone bed than that described in the core, between 222,6 and 224,3m, has to do with both silt and mud contents present as matrix of the sandstone, i.e., filling the porous space among the sand grains.The silt class tends to be enriched in potassium feldspar and the clay content of the mud is also rich in such element.In both cases, textural and mineralogical properties altered the well log patterns and affected the lithological classification.Potassium, thorium and uranium are the main radioactive element present in minerals and affect considerably the gamma rays record in well logs.The presence of silt and clay into the sandstone modify the permeability and consequently the SP, RTV and RTC logs.
This borehole also contains seven coal beds, named as S2, S3, I, I2, I3s, I3i and I4, in the interval between 256,28 and 278,55m.Thin beds are unlikely to be classified correctly due to the shoulder effect, i.e., the logging tools are not capable to record the true values of each property because the rocks situated below and above a thin interval influence the logs, creating false values between the minimum and the maximum records of contrasting rocks bounded by sharp contacts.Thus, isolated thin beds of coals are not classified correctly as electrofacies 3.
Another question is related with the wide range of possibilities involving the term coal in the study area.Coal is also used as a generic term to designate organic carbon-enriched muddy rocks.In the Leão Mine coal is characterized by a great amount of inorganic material, such as clay, silt, quartz, feldspar and lithic grains etc.In fact, the organic rocks present in the area range from coal to carbonaceous siltstones.The variable content of the inorganic material contribute to modify the log values and, consequently, the classification of the neural network.
The log patterns in this interval are characterized by abrupt shifts of values, especially in the GR, RTC and RTV logs.
Electrofacies 1 was not classified with high accuracy as well, in the interval between 280,0 and 311,5m.The results of the neural network show a lot of intercalations of siltstones, and also coals, which decrease the real thickness of the sandstones.In this case, the presence of both feldspar grains and siltstone intercalations increased locally the values of the GR, RTV and RTC logs of the sandstone but SP was little or not affected.
Figure 3 shows the comparison between the core log made by the geologist and that obtained from the neural network classification.In the LB-148-RS the success rate was 78%, the error rate was 20% and 2% corresponds to non-identified patterns, all of them related with the validation stage.Electrofacies 2 reached 94,9% of success in its classification while electrofacies 1 had 9,8%.Electrofacies 3 was not classified in this borehole (Table 2).The incorrect classification between 245,4 and 245,9m resulted from the generalization for training purposes.Defined as siltstone in a minor scale, this interval contains a lot of millimetrical to centimetrical sandstone intercalations, only recognizable in a major scale of observation and description such as that of the core log.The neural network classified an intercalation of non-identified, coal and sandstone, from the base to the top, within the siltstone interval.It is quite important to calibrate the resolutions of the neural network, of the logging tools and the description performed by geologists.The resolution of the logging tools is up to 30cm therefore any thinner layer classified by the neural network must be inspected by a log analyst.This inspection can reveal, for instance, the reason for the incorrect classification of coal in the above-mentioned interval.In this case, coal resulted from a combination of specific values of the GR, RTV and RTC logs caused by the shoulder effect.
On the other hand, electrofacies 1 between 261,0 and 273, 29m was poorly classified by the neural network.The inspection of the core log has revealed that the electrofacies 1 in this interval comprises sandstone with milli-to-centimetrical intercalations of siltstones between 261,0 and 269,0m.From 269,0 to 273,29m silt is present as matrix of the sandstone in variable proportions, locally being considered as a sandy siltstone.For the depth of 318,9m, the network incorrectly classified the interval as electrofacies 1.However, in the descriptive profile, the layer of sandstone shows fine to medium grain size only in the range from 317 to 317,40m.
The incorrect classification of the coal beds, i.e. electrofacies 3, can be explained by their composition; in fact, they are very carbonaceous black siltstones.The comparison between the core log and that obtained from the neural network classification is shown in the Figure 4.The boreholes LB-137-RS was adjusted with a multi-layer model and had the best performance of learning with the topology [5-15-3-3], using as variables depth and the logs gamma rays (GR), spontaneous potential (SP) and resistance (RTC).The activation function used in all layers was the logistic sigmoidal, the algorithm was the Levenberg-Marquardt and the learning finished after 2000 cycles.
The success in the classification rate for this borehole during the validation stage reached 81% while errors and non-identified patterns reached 18% and 1%, respectively.Electrofacies 2 reached 93,6% of success rate and electrofacies 3 was correctly classified in 57,5% of the samples.Electrofacies 1 had the minor success rate with 43,3% (Table 3).The main reason for the incorrect classification of the electrofacies 1 situated between 223,60 and 225,40m is the presence of many intercalation of siltstones, which increase the GR and RTC values and decrease the SP one, and the small number of samples of these thin sandstone bed.As in the LB-148-RS, scale of description versus scale of sampling for training of the neural network was involved in problems for recognizing sandstones.
Several intercalations generated by the neural network classification in the interval between 240,73 and 253,86m have to do with the wide range of values of the GR, SP and RTC logs.This interval was defined as electrofacies 1 in the scale of observation and description and used as electrofacies 1 for training purposes.However, the neural network provided a different result.Electrofacies 2 and 3 were classified as centi-to-decimetric intercalations (Figure 5).The wide range of values of the well logs results of a combination of the lithological characteristics.The presence of calcite as a cement alters the pattern of the RTC and SP logs while siltstone intercalations and feldspar grains increase the GR log, promoting the classification as electrofacies 2 and 3.
The seven layers of electrofacies 3 described in the cores, named as S2s, S2I, S3, I, I2, I3 and I4 occur as intercalations among siltstone beds in the middle of the borehole LB-137-RS, between 253,90 and 277,31m (Figure 5).The electrofacies 3 classified by the neural network do not correspond with the depths of the core.However, intercalations between electrofacies 3 and 2 are the result provided by the neural network as those present in the middle portion of the core, together with some non-identified intervals.A reason for this difference in the depths is the cumulative error associated with both drilling operations and core measurements.
The drilling operation that recovers cores is performed by introducing drill bits sequentially into the borehole.The measurement of the cored interval is made based on the number of drill bits introduced in that interval.Thus, the more bits are used the more will be the error in the end of the operation.After removed from the drill bit, cores usually become bigger in length, either by the expansion promoted by the difference of pressure or by the number of fractures generated by the core drill operation.Hydration of mud-rocks containing smectites can also promote expansion of the cores.Therefore, is normal that cores have about 5 to 10% more in length than originally recovered from the borehole.Many times the core description is performed without considering this difference, expanding the real thickness of the rocks, especially muddy siltstones and claystones.On the other hand, the log tools perform measurements in situ, in which they are introduced at the bottom of the borehole and pulled up using cables.This type of measurements is also inclined toward problems, once that in small diameter boreholes as LB-137, LB-138 and LB-148, the logging is performed separately, i.e., GR, SP and RTC/RTV logs are acquired in different operations.This results in non-adjusted depths among the geophysical logs when different log acquisitions are not correctly calibrated.The borehole LB-137 exhibit at least three good examples of how this above-mentioned situation impacted the classification of sandstones by the neural network.The base of the electrofacies 1 in the GR log at approximately 295,0m is not coincident with both SP and RTC logs and the core log.Consequently, just a thin bed of electrofacies 1 was classified in the neural network log although sandstone in this interval is considerably thicker (see Figure5 -between 290,5 and 296,0m).
In the second case, between 241,0 and 253,0m, the RTC log is relatively about 80cm dislocated upward when horizontally compared with the SP log both at the top of this interval (see Figure 5).It was not possible to compare the GR log with SP and RTC in the same interval because the presence of potassium-bearing feldspars and clay minerals increases considerably the values of the natural radioactivity.However, their presences were not informed in the core log description and the positioning of the base of the sandstone based on the GR log would not be reliable.As a consequence, the total thickness of electrofacies 1 as described in the core log was classified by the neural network as intercalations of electrofacies 1, 2, 3 and also non-identified pattern.The description of this interval in the core log does not mention the presence of coal and mention silt only as a subordinate fraction of the sandstone.
The third example shows the GR log dislocated about 2,0 meters upward when compared with SP and RTC logs.
The quartzose composition as well as the presence of carbonatic cement indicates that the electrofacies 1 is defined by a low value in the GR log, a relatively lower value in the SP log and a high value in the RTC log.They can be identified as spikes in the above-mentioned logs.However, GR spike do not correspond with SP and RTC spikes in depth (see Figure 4,between 223,61 and 225,5m).Thus, the neural network recognized the electrofacies 1 as a thin bed of electrofacies 3 intercalated within electrofacies 2. This is easily explained because the combination of values of three geophysical logs was not sufficient to characterize the electrofacies 1 as previously defined during the training and validation phases.

Conclusions
Neural networks are a powerful and fast technique for identifying patterns and classifying rocks from well log data, especially in areas with a great number of drilled boreholes.Based on the results obtained from the study area it was possible to determine a set of important elements involving both NNs and geological aspects capable to influence the results.
The neural model used was the feedforward.Different tested parameters, such as number of hidden neurons and the number of variables as input has revealed that two intermediary layers, four variables of input and the sigmoidal logistic activation function [0;1] in all layers provided the optimal configuration.The best performance was achieved combining the above-mentioned characteristics with the Levenberg-Marquardt and Resilient backpropagation algorithms for training and validation purposes.
The accuracy of approximately 80% in the training and validation phase cannot be understood as a guarantee of high quality results.This was noted analyzing the confusion matrices of the LB-137, LB-138 and LB-148 boreholes.Electrofacies 2 had accuracy up to 88% whereas electrofacies 1 ranged from 9,8 and 51,8% and electrofacies 3 from 0 to 57,5%.A number of different factors explain the discrepancy of accuracy among the different electrofacies and their correspondent rocks, such as the difference of the scale of description with the resolution of the log acquisition, the frequent presence of silt and clay into the sandstones and coals, the abundance of potassium feldspar into the siltstones and fine-grained sandstones, the wide range of composition of the coals, the quality of the log acquisitions, the correct calibration of the geophysical logs acquired in different runs and the limited availability of appropriated logs for rock classification.
Electrofacies 2 showed the highest accuracies because they displayed the major variations of values in the logs whereas electrofacies 1 and 3 were defined by a more restricted combination of the log values.Additionally, problems in the depth calibration of different logs contributed substantially for reducing the accuracy of sandstones and coals, especially in the boreholes LB-137 and LB-148.Despite the high complexity, this technique is a powerful resource to help geologists analyzing and classifying a massive number of boreholes bearing several logs.The fastness in obtaining interpreted data in a digital format is an additional advantage.However, the inspection of a log analyst will be always necessary to avoid problems related with both log acquisition and calibration.

Figure 3 .
Figure 3.Comparison between the descriptive litholog of the borehole LB-138-RS of CPRM (left) and the classification performed by the neural network (right)

Figure 4 .
Figure 4. Comparison between the descriptive litholog of the borehole LB-148-RS of CPRM (left) and the classification done by the neural network (right)

Figure 5 .
Figure 5.Comparison between the descriptive litholog of the borehole LB-137-RS of CPRM (left) and the classification done by the neural network (right)

Table 3 .
Confusion Matrix of the Borehole LB-137-RS