3.1. Lithology Prediction Through ML

A decision tree (DT) is a type of supervised learning technique that is commonly used for addressing classification issues, as well as regression assignments. The classifier is represented as a tree structure, with internal nodes representing features of the dataset, branches representing decision rules, and each leaf node indicating an outcome [49–51]. There are two distinct categories of nodes in the DT structure: decision nodes, which make decisions involving numerous branches, and leaf nodes, which represent the results without any extra branches. The characteristics of the dataset influence the process of making decisions or conducting tests [52].

One of the significant tools of ML is prediction. ML was used to predict the result on the blind well. Different types of indicated algorithms were used, and every algorithm had its own specification and accuracy. On the given data set, the lithology was predicated on the blind well. In the current research, two well log data were used. One well was X-01, and the other one was X-02. X-01 well was utilized for training data sets for the lithology prediction, and X-02 well was used for testing well for lithology prediction. This study adopts a DT classifier ML model that was trained using the lithology dataset of the Patala and Hangu formations for the X-01 well to make lithology predictions. This was demonstrated by the model’s overall accuracy, 73.20% and 84% for Patala and Hangu formations, respectively.

3.2. TOC Through Well Logging

Precisely predicting TOC values from well logs data is crucial and given the impracticality of conducting direct measurements from core/well cutting samples in numerous wells. The various methods devised over recent decades rely on correlations, numerical relationships, and easily verifiable TOC predictions [34]. This study applied different methods to estimate TOC and assess its viability, a promising correlation with core measured values. The methodology was applied to one well’s publicly available well log data to evaluate the Paleocene (Hangu and Patala formations) hydrocarbon source potential.

3.3. TOC Through ML

Chen and Guestrin [57] introduced XGB trees, a scalable tree boosting method that enhances the gradient boosting DT framework. Tree boosting is a widely used and highly effective ML technique. XGB employs classification and regression trees (CARTs) to train on data by minimizing an objective function.

3.4. RE Pyrolysis and Organic Petrography

RE pyrolysis is the most effective tool that can measure essential geochemical parameters during the source rock geochemical assessment procedure that obtain geochemical parameters which are further used to derive parameters and is the most often used method to access source rock generation potential. The highest pyrolysis yield temperature, S₁, S₂, S₃, hydrogen index (HI), oxygen index (OI), and other parameters were obtained from the RE pyrolysis [58–60]. The conventional well log data set are being considered an alternate solution that might be used to estimate the TOC values of the source rock in the absence of a well cutting/core sample [3]. The utilization of traditional well logs for the rapid evaluation of the TOC content of the source rock is a valuable technique that facilitates the identification of organic richness [61].

Globally, the petroleum industry employs RE pyrolysis to quantify, assess the quality, and determine the thermal maturity of organic materials in rock samples. When used in conjunction with TOC measurements, this screening method is the fastest and most cost-efficient for screening many samples [62, 63]. Plotting RE pyrolsate (S₂) yield versus TOC yield offers information on the kinds of insoluble OM, as well as the HI [64, 65]. In the current study, only 12 selected samples of the target formations were analyzed using the MCS CCD Z1M Zeiss microscope at a laboratory. The proper quantity of crushed sample, measuring between 0.8 and 0.2 mm, was mounted with Araldite and then allowed to dry so that it might solidify.

4.1. Structure Interpretation

Patala and Hangu formations are marked based on a synthetic seismogram on the dip line G962-SHD-316. The marked horizons are then ultimately shifted to the intersecting strike line and thus traced on the other dip lines. During this procedure, small misties are witnessed and corrected. Patala Formation comprises shales and sandstone, so its reflections are improper. For identifying Hangu, firstly, underlying strong reflections of Lockhart are marked because of prominent acoustic impedance contrast. This facilitated the picking of the Hangu reflector on available seismic lines.

The study field is in an extensively compressional regime dominated by thrust faults. The major structures formed are fault bend folds. Faults are marked on the seismic section based on disturbance in seismic behavior. Most of the marked faults are trending in the NW–SE direction. Faults are evident on the dip lines as most are north–south (NS) trending (invisible on strike lines). Seismic sections of lines G962-SHD-316 (Figure 6) and G962-SHD-313 (Figure 7) are marked with Hangu, Lockhart, and Patala horizons and faults.

Both time and depth contour maps were yielded for the Patala and Hangu formations to conform to their exact subsurface levels (Figures 8 and 9). The time and depth contour maps are generated at the contour interval of 0.25 s, and the depth contour map is generated at the interval of 50 m. The faults are present in black color, which demonstrates the structure that is present in the subsurface. Figure 8a shows the TWT contour map, and Figure 8b shows the depth contour map of the Patala Formation. Figure 9a shows the TWT contour map, and Figure 9b shows the depth contour map of the Hangu Formation.

4.2. Lithology Identifications

Various previous studies have examined lithofacies and investigated how lithofacies affect diagenesis and reservoir heterogeneity in shale gas plays. Due to its complex diagenesis, rapid lithofacies, and reservoir quality variation, unconventional resource plays make studying lithofacies for diagenesis control more difficult and important [66]. Shale reservoirs exhibit variations in reservoir quality (petrophysical and geomechanical properties) due to variability in their mineral and OM compositions, diverse lithofacies, and sedimentary–diagenetic environment. The reservoir quality of shale directly influences shale gas extraction and development activities. So, the classification of shale lithofacies types and the identification of the optimal shale lithofacies with the highest reservoir quality are for targeting the “sweetspots” [11, 67, 68].

Li et al. [69] conducted a detailed study on lithofacies and organofacies of marine and lacustrine shale (organic-rich matter) and found that a robust association exists between lithofacies and organofacies in the sedimentary rocks. Lithofacies and organofacies play significant roles in unconventional petroleum exploration. The heterogeneity in lithofacies affects the identification of “sweetspots” in shale plays as distinct lithofacies most probably align with varying organofacies. The differences in thermal maturity of organofacies influence the variations in hydrocarbon occurrence states and pose a significant challenge to the identification of sweetspots. Jin et al. [70] performed a comprehensive study on the classification of shale oil reservoirs in China. The shale can be classified as shale oil reservoirs based on its sedimentary structure.

The standard practice to precisely identify lithology and describe reservoir characteristics is from well core observations. However, the limited availability of core samples poses significant challenges for geoscientists precluding a thorough characterization of lithologies. Therefore, well log data are quite beneficial as it provides continuity and ease of acquisition to determine different parameters [71]. Identification of different lithologies from well log data might be difficult due to the presence of a variety of minerals. Additionally, the well log data may be influenced by several geological and drilling conditions [72]. To overcome these issues related to lithology identification, artificial intelligence-based approaches have been employed to improve accuracy as well as efficiency [73].

At present, the rapid development of ML approaches could aid in getting rigorous assessment of lithologic characterization from well log data. These techniques can classify lithology in the absence of core data after being trained with the available core data. Lithology classification is utilized for depth measurement in model training by employing a combination of well log and core data. Main input logs for lithology identification include GR, CNC, RD, and ZDEN [74]. In the current study, these logs are utilized to predict the lithology of the Hangu Formation and Patala Formation by training the core data lithology on X-01, while the testing was performed on X-02 well.

In investigating lithological anomalies, cross-plots are pivotal tools for visually discerning potential hydrocarbon reservoirs through the graphical representation of interrelationships between various well log responses. In the context of the Patala and Hangu formations, four distinct cross-plots have been generated, each elucidating meaningful associations among different variables. These plots include the relationship between RD and density, GR and porosity, density and GR, and GR and RD. These cross-plots, commonly called z-plots, employ GR as the primary indicator variable, with lithological facies serving as the secondary indicator. Notably, the juxtaposition of GR and lithology in these plots underscores the utility of GR as a lithological indicator. Lithology delineations within each formation, particularly shale, marl, and limestone in the Patala Formation, and sandstone and shale in the Hangu Formation, are distinctly discernible as shown in Figures 10 and 11. This characterization is particularly salient in analyzing the X-02, widely regarded as a blind well.

The lithology in the current study has been identified based on observations, descriptions of the core samples, and core pictures. The following are the lithology categories, (1) sandstone, (2) shale, (3) limestone, and (4) marl, which is visible in the Patala and Hangu formations that characterize this type of lithology. The dominant lithology observed in the Patala Formation is limestone, shale, and marl, and Hangu Formation; only sandstone and shale were encountered, as shown in Figure 12a,b. The result of the classifier report for both Patala and Hangu formations is shown in Figure 13a,b. The F1 score for the Patala Formation is 73.20%, and for the Hangu Formation, it is 84%, representing the good training of the data set. Figure 14a,b demonstrates the lithology of both testing and training well, showing the lithology prediction in X-02 on Hangu and Patala formations. The prediction of lithology in the Hangu Formation is reasonable compared to that of the Patala Formation.

Figure 15a,b shows the training X-01 and testing X-02 well for the lithology prediction in the Hangu Formation. Four different log tracks are shown in the plot: the first track represents the GR log, the second track represents the neutron log, the third track represents the density log, the fourth demonstrates the RD, and the last track shows the lithology prediction. The prediction of lithology results in the Hangu Formation is better as compared to the Patala Formation because of the availability of a log curve. Figure 16a,b shows the training X-01 and testing X-02 well for the lithology prediction in the Patala Formation. The color in the lithology track shows the variation in the lithology.

4.3. TOC Through Well Cutting

All samples’ X-01 well cuttings have a TOC ranging from 0.64 to 3.20 wt.% (Figure 17) which indicates fair to very good hydrocarbon potential. According to the classification of Bacon et al. [75], Hangu Formation TOC present values are in the fair to very good range. In comparison, the TOC values for Patala Formation are also fair to very good.

4.4. Wireline Log Analysis

The consistent availability of data over the whole zone of interest makes TOC estimates based on well logs quite common [76]. Figures 18 and 19 illustrate the computed TOC content in X-01 in Paleocene formations using the bulk density, GR, ΔlogR, and multivariant techniques. Figures 18 and 19 show that the observed TOC shown with a red color triangle and the predicted TOC shown in the curve with the help of well log data demonstrate the best fit matching in the log plot for Hangu and the Patala formations. The Hangu Formation has six observed sample data in which two points of TOC approximately match the log curve of TOC prediction, while in the case of Patala, the maximum points are tied with the log curve.

The TOC values estimated through conventional well logs normally differ from laboratory measures [77, 78]. The various well log data-based approaches, including single and composite logs methods, are routinely employed to estimate TOC values. As indicated in Table 3, the correlation between the observed TOC from well cuttings and the predicted TOC (based on the well logs) was assessed in the current study. Table 3 presents the detailed correlation between actual and predicted TOC for the Hangu and Patala formations. The correlation means that the formation Passey method gives the best results compared to other TOC prediction methods.

4.5. TOC Through ML

The TOC content prediction using ML techniques helps a lot to overcome the deficiencies of the conventional quantitative regression approaches used to predict TOC content [58, 79, 80]. Several ML algorithms, including RF, NN, SVR, Bayesian regression, DT, and XGB, have been employed to predict TOC content for source rock characterization. In contrast to actual regression models, these algorithms can handle exceedingly complex relationships among independent and dependent variables to forecast unidentified values [56]. TOC content estimated from borehole data using ML method can immediately recognize unidentified relationships between the TOC features and the well logs [10, 23, 81]. ML approach is considered more accurate compared with old empirical regression techniques in estimating TOC content as it is more rigorous [79]. In this study, the effectiveness of RF, DT, XGB, and NN in TOC content forecasting is compared, and the potential of ML models in specific settings is also determined.

In petroleum systems, the prediction of TOC based on ML techniques is a relatively new research area. Various studies reveal that the use of ML approaches along with traditional mineralogical and geochemical techniques can sufficiently increase the precision of TOC predictions. Generally, ML techniques as compared to traditional statistical or deterministic models provide detailed information about the complicated relationships between TOC and other source rock properties [82]. The RF method provides the best results (R² = 0.915) among the other ML models including SVR and XGB to predict the TOC in organic-rich shale reported in Sun et al. [56]. Khan et al. [83] also applied the same ML models to determine TOC values in shale play basins in Asia and North America and found the best results from the RF approach with a strong correlation (R² = 0.85). Similar results were obtained to predict the TOC in Devonian Duvernay shale using RF, SVR, and DT methods, and the RF technique provided the optimum results with correlation coefficients (R) between 0.93 and 0.99 [84].

Shan et al. [85] employed a deep spatial–sequential graph convolutional network to predict the TOC with R² = 0.87 in the Sichuan Basin. Nyakilla et al. [28] employed SVM and Gaussian process regression to predict the TOC and concluded that Gaussian process regression provides the best results with R² = 0.95. In the present study, four ML models including RF, DT, XGB, and NN techniques were utilized to predict TOC content for Hangu and Patala formations. Subsequently, the results obtained from these models are validated by predicting unknown TOC content and compared to empirical regression methods, multiple linear regression, and ΔlogR. The ML algorithms can reduce the cost and improve the efficiency of laboratory testing in predicting the accuracy of TOC content among different forms and locations.

The research involved calculating correlation coefficients between all input parameters and core and log TOC output parameters. Figures 20 and 21 present pair plots, visually examining the correlations between input and output parameters. In Figure 20, the pair plot demonstrates the plot before removing the outlier or irrelevant data, and Figure 21 represents the pair plot before removing the outlier. It is necessary to remove the outlier for the betterment of results. These graphical representations offer insights into the relationships and trends among various input and output parameters. Some parameters exhibit strong interconnections, while others demonstrate a moderate level of correlation.

Figure 22 shows the box plot of the TOC estimated by the well log data before and after removing outliers. The primary rationale for outlier elimination is that outliers are exceptional values that differ significantly from the remaining data, hence distorting conclusions and complicating statistical investigations. The removal of outliers improves the reliability and precision of the results by limiting their influence. Outliers are removed from the Z-score treatment method, and this technique is applied to the calculated TOC by different methods, as shown in Figure 22.

A comprehensive performance evaluation was conducted to compare the efficacy of the four artificial intelligence algorithms employed in this study, namely, DT, RF, XGB, and NN. Figure 23 demonstrates the scatter plots illustrating the correlation of actual versus predicted TOC generated based on different algorithms during training data sets. Figure 24 represents the testing correlation of all the algorithms used in studies; blue color shows the original data set, and other colors demonstrate the predicted values. Based on the scatter plot, the XGB, DT, and RF score better than NN because NN data are more scattered than other data sets.

Table 4demonstrates that hyperparameter tuning is the procedure of modifying key variables for each kind of ML algorithm to get the highest accuracy for forecasting. The RF model variables, such as n estimators (the number of trees), max depth (the greatest depth for every tree), and min samples split (the smallest amount of data needed to separate a node), were modified to discover the best values. The choice to increase the tree effectiveness of the model was improved by modifying the max depth and min sample split. Hyperparameters in the XGB model, such as learning rate (step size shrinkage), n estimators, and max depth, were adjusted to balance learning speed and accuracy. To improve the learning abilities of the NN (MLP regressor), the hyperparameter’s hidden layer sizes (structure of the NN layers), activation (activation function for the neurons), and learning rate init (initial learning rate) were changed. The linear regression method did not necessitate hyperparameter adjustment because it employs a straightforward linear method. The optimum variables for every model were selected using grid search CV, and the model’s effectiveness was assessed using mean squared error (MSE). Table 5 represents the training and test results of different ML algorithms used in the current study. The algorithm XGB and DT give the best training, while the RF and XGB provide the best testing results.

4.6. Kerogen Type and Its Microscopy Characteristics

Various sorts of kerogen can yield different kinds of hydrocarbons. Type I kerogen, which integrates marine organic material; type II kerogen, which mostly generates gas but also produces oil; and type III kerogen, which produces gas in the majority, all result in the production of oil [86]. A cross-plot of the HI versus OI was used to find out the kind of kerogen using the Van Krevelen diagram. This cross-plot, which employs the atomic ratios of hydrogen, carbon, and oxygen, was created by Van Krevelen. Later, Tissot [87] used the HI and OI values instead of these atomic ratios.

Samples from Paleocene formations are given in an HI versus OI plot (Figure 25). All Paleocene Age formations’ kerogen types are depicted in the plot. Hangu Formation with an HI value of under 200 is classified as type III. Types II to III kerogen are the Patala Formation with HI values between less than 500.

The kerogen type known as vitrinite derives its composition from organic materials found on land. The amount of incident light returned from the polished surface of the sample is used in this procedure. Compared to the other macerals, the optical characteristics of the vitrinite maceral group alter more gradually as their grade advances. The sample’s thermal maturity is determined using this method. The geochemical changes in the macerals of vitrinite, which are positively correlated with the reflectance value, are also influenced by the geothermal history of the sedimentary basin [88]. Vitrinite reflectance reveals the stage of OM transformation and provides insight into the type of hydrocarbon the sample generated, as seen in Figure 26.

The R_o standards of 1.711%, 0.907%, and 0.589% were used to calibrate the microscope. The Hangu Formation, which belongs to the early mature oil phase category, has a vitrinite reflectance value of 0.78–0.90 (Table 6). The Patala Formation is similar in the oil window thermal maturity stage, according to its value; however, it is shallower than the other formation and has R_o values of 0.83–0.94. The R_o information demonstrates that all formations can generate oil.

Several kerogen kinds were identified on polished samples, and a Zeiss microscope (MCS CCD Z1M) was used for white and fluorescent light analysis. According to a petrographic study, vitrinite is the primary maceral in Paleocene formations. The second principal maceral is inertinite, and the third maceral is solid bitumen. Pyrite is the main mineral in Paleocene formations. The Hangu and Patala formations are well-characterized by the presence of distinct maceral kinds, including pyrite, inertinite, and vitrinite (Figures 27 and 28).

4.7. Genetic Potential

However, the kind of hydrocarbons formed during pyrolysis cannot be foreseen using this method [90]. The cross-plot TOC versus GP values were utilized to assess the source rock’s quality and its potential for hydrocarbon production in Paleocene formations. With the exception of a limited number of samples of Hangu and Patala formations, the GP of Hangu and Patala formations is fair to very good, as shown in Figure 29.