Identifying Different Components of Oil and Gas Shale from Low-Field NMR Two-Dimensional Spectra Based on Deep Learning

Abstract

The detection and quantitative analysis of shale components are of great significance for comprehensively understanding the properties of shale, assessing its resource potential and promoting efficient development and utilization of resources. The low-field NMR T₁-T₂ two-dimensional spectrum can detect shale components non-destructively and effectively. Unfortunately, due to its complexity, the two-dimensional spectral results of low-field NMR are mainly analyzed using manual qualitative analysis, and accurate results of the composition cannot be obtained. Since the information contained in its two-dimensional map is determined by the morphological texture and the position in the map, commonly used image analysis networks cannot adapt. In order to solve these problems, this paper improves a novel Faster Region-based Convolutional Neural Network (Faster-RCNN). Compared with previous models, the improved Faster-RCNN has better image classification and visual key point estimation capabilities. The results show that compared with traditional methods, the deep learning method using this model can directly obtain key information such as kerogen and movable oil and gas content in rocks. The information provided in this study can help complement and improve the development of analytical methods for low-field 2D NMR spectra.

1. Introduction

With the growth in global demand for oil and gas resources, shale gas has attracted attention due to its huge resource potential. As an unconventional resource, it has the characteristics of large gas-bearing area, long production cycle and stable output, and has become a hot spot in global exploration and development [1].

Unlike conventional sandstone reservoirs, for shale gas systems, shale is both the source rock and reservoir. Therefore, finding effective reservoirs has become the core of shale gas exploration, and precise description and characterization of shale reservoirs are even more necessary [2]. The porosity and permeability of the shale gas matrix are very low. According to their origin, the pores can be divided into nanoscale organic pores, mineral intragranular pores and intergranular pores, and micron-scale micro-fractures. Organic matter pores are generated during the hydrocarbon generation process of organic matter, and the number of pores is directly related to the maturity of organic matter [3]. The occurrence modes of shale gas include free gas, adsorbed gas and dissolved gas. Free gas mainly exists in pore spaces such as inter/intra-particle pores and micro-cracks. A large amount of shale gas adheres to the surface of kerogen and clay particles in an adsorbed state, and a very small amount is dispersed in kerogen, asphalt and water in a dissolved state [4].

The measurement and characterization of shale organic matter is usually achieved using geochemical experimental methods. Organic matter content is expressed as total organic carbon (TOC), and laboratory measurements include carbon and sulfur determination, combustion, pyrolysis gas chromatography and chloroform pitch “A” determination. The maturity of organic matter can be characterized by parameters such as rock pyrolysis parameters, vitrinite reflectance, chemical composition characteristics of soluble extracts, kerogen free radical content and time-temperature index. Geochemical parameters have limitations in data discontinuity, heterogeneity and analytical laboratory workload [5].

Scanning electron microscopy (SEM) [6] and other observation techniques can observe the pore structure, minerals and contact characteristics in shale; obtain high-resolution images; and perform qualitative analysis to obtain the pore structure parameters from imaging. Currently, the mainstream field emission electron microscope (FE-SEM) [7] and focused ion beam scanning electron microscope (FIB-SEM) [8] are the most commonly used research tools for studying the nanoscale pore structure of shale. The advantage is that it can visually characterize pore morphological characteristics and cause analysis, but due to the impact of resolution and sample scale, it cannot meet the needs of large-area observations. On the basis of direct observation, many scholars combine digital rock technology to conduct three-dimensional reconstruction of shale organic matter and pore structure, and extract pore parameters such as porosity, connectivity and pore throat radius [9]. In addition, some indirect testing methods include the following: CO₂ [10] adsorption method, N₂ [11] adsorption method and high-pressure mercury injection (MIP) [12]. Non-wetting fluids such as gas or mercury are used to inject samples at different pressures to conduct adsorption–desorption experiments or capillary force experiments, obtaining rock pore size distribution, specific surface area and pore volume. These methods require destroying the rock sample, injecting test fluids and can only detect open and connected pores. Due to different testing principles, the observed pore size ranges are different. Nuclear magnetic resonance (NMR) technology has become a powerful tool for the identification and evaluation of shale organic matter and organic pores due to its non-destructive and efficient characteristics, as well as its dual detection functions in the laboratory and downhole [13].

NMR can provide reservoir parameters such as porosity, permeability and irreducible water saturation, and observe fluid distribution status [14]. It is one of the most important means for reservoir evaluation of complex oil and gas reservoirs. At present, the application of this technology in conventional reservoirs is very mature and effective. However, in the application process of shale gas reservoirs, it is affected by factors such as complex mineral composition, special pore structure, rich organic matter, ultra-low permeability and nanoscale pores. It faces many challenges such as low detection resolution, low signal-to-noise ratio and inapplicable interpretation models [15,16]. However, with the development of nuclear magnetic resonance equipment for many years, the minimum echo interval of nuclear magnetic resonance logging tools has reached 0.2 ms, and the echo interval of laboratory desktop nuclear magnetic resonance core analyzers (Magritek, Oxford Instruments and other companies) has reached 0.06 ms, which can detect the shortest T₂ value of the core at 0.01 ms. Thus, the fluid signal in the nanoscale pores of shale can be obtained.

As a result, domestic and foreign scholars have used experimental methods to conduct a large amount of research work on the NMR response characteristics of shale. On this basis, they have gained a deeper and updated understanding of the NMR relaxation mechanism of shale. In the early stages of shale research, researchers compared NMR T₂ spectra with pore sizes measured by mercury intrusion, adsorption, etc., and divided the pores into small pores and large pores based on the T₂ spectrum relaxation time. They believed that short relaxation groups represented small pores. The main components are organic pores (Kausik, Cao, Tinni, Richard, Sigal) [17,18,19,20,21], and the organic pore signal is positively correlated with the total organic carbon content (Chen, Kausik, Habina) [22,23,24]. The organic pore surface relaxation rate is about 40–50 μm/s, which is obtained by converting the T₂ spectrum into a pore size distribution. Washburn (2014) [25] calculated the surface relaxation rate through shale micromineral analysis and paramagnetic substance content, and found that the surface relaxation rate of shale organic pores should not be so high. Therefore, industry experts have questioned the pore size distribution of NMR T₂ spectrum conversion in organic-rich shale. Rylander (2013) [26], Cao (2013) [27] and others tried to explain it from the perspective of organic matter wettability, but found that even considering the wettability characteristics of organic matter oil, they could not fully explain the high surface relaxivity of organic pores. R. Kausik (2017) [28], Korb et al. (2018) [29] and others used high-field NMR measurement methods to study shale characteristics and isolate kerogen and asphalt signals. With the improvement of instrument measurement accuracy and the increase in research on the nuclear magnetic resonance properties of organic matter itself, researchers have found that the short relaxation of the T₂ spectrum contains signals generated by the shale organic matter itself, and that the surface relaxation of organic matter pores and inorganic pores are caused by paramagnetic substances. The surface relaxation is different, which is the reason why the capillary pressure curve or adsorption curve cannot correspond to the nuclear magnetic resonance T₂ spectrum. Washburn (2013) [30], Hugh Daigle (2014) [31] and others conducted theoretical research on the surface relaxation of organic pores and believed that the surface relaxation in organic pores depends on the homonuclear couple between the fluid and the hydrogen protons in the surface. This extreme coupling is different from the reason why relaxation occurs in inorganic pores (inter-granular pores, intra-granular pores), therefore it is difficult to detect this part of the relaxation information by conventional measurement methods (CPMG, inversion recovery or saturation recovery method). In recent years, the application of two-dimensional NMR has become increasingly widespread. Two-dimensional T₁-T₂ NMR is also starting to be used in the medical field for non-invasive detection of diabetes and other conditions [32]. Washburn and Birdwell (2013) [33] tried to introduce the solid-state NMR method into shale measurement, enhance the relaxation signal of organic matter kerogen, etc., analyze the relaxation mechanism of the organic matter pore surface and the organic matter itself, and confirm the existence of the homonuclear nature of shale organic matter and organic matter pores, including the dipole coupling phenomenon. This research results updated people’s understanding of the NMR relaxation mechanism of shale. Since 2014, Xiao Lizhi, Jia Zijian, et al. [34,35] of China University of Petroleum (Beijing) have carried out some basic theoretical and experimental research on shale NMR theory and technology, confirming the existence of a special surface relaxation mechanism of shale organic matter. Song Yiqiao of Harvard University (2019) [36] systematically introduced the advantages and future development trends of NMR technology in shale oil and gas. Fleury (2016) [37], R. Kausik (2019) [38], Tan Maojin (2015~2020) [39,40,41,42,43] and other experts and scholars have tried to apply NMR T₁ through a large number of shale relaxation experiments. T₂ distribution distinguishes different components of shale, including organic matter, gas, water, etc., thereby obtaining typical NMR response values of different components. However, different components of shale are affected by other factors such as fluid saturation state, organic matter maturity, porosity, etc., and their exact positions and shapes on the T₁-T₂ map are quite different. It is not possible to directly apply this response characteristic map for shale exploration [44]. Taking the maturity of shale organic matter kerogen as an example, as kerogen matures, it will lose fatty chains (the part with high hydrogen content), aromatic rings form clusters, solid organic matter becomes harder, and its T₁ relaxation time will become longer. The T₂ relaxation time of solid organic matter is almost constant. Therefore, as the maturity of kerogen increases, the T₁/T₂ ratio gradually decreases, and NMR can indirectly reveal the maturity stage of organic matter [45].

In recent years, with the rise of artificial intelligence, especially the advent of the era of big data and deep learning, machine learning, as the main method of experimental artificial intelligence, can design algorithms so that computers can learn certain patterns from large amounts of data. It has outstanding performance in the fields of mining and image recognition [46,47,48,49,50,51]. Tamoto [52] discusses the use of supervised machine learning models to predict nuclear magnetic resonance porosity well logs in a carbonate reservoir. The two-dimensional NMR image contains information on many components such as organic matter, organic pores, oil, gas, water, etc. Each component has different NMR response mechanisms and different response results. As the shale organic matter maturity, organic matter content and other component information changes, the NMR spectrum may change dramatically [53]. Manual identification has problems such as being time-consuming, using large calculations and having low accuracy. The introduction of machine learning methods can improve NMR spectra identification. Data processing efficiency can be achieved by introducing geochemical results and nuclear magnetic resonance results to establish a convolutional neural network model and adding appropriate expert intervention. This can improve the accuracy of interpretation and analysis of shale components. Therefore, using machine learning methods to perform data processing on two-dimensional NMR spectra should be able to achieve quantitative identification of special components of shale. This is the first time that AI has been applied to the analysis of two-dimensional NMR results of rocks.

2. Theory

2.1. NMR Theory

The T₂ relaxation is widely used to assess porosity, pore size distributions, fluid content and wettability in both core analysis and borehole evaluation. The T₂ distribution is acquired using the magnetization decay curves and Laplace inversion algorithm.

However, for shale, due to the complexity of its composition, using a surface relaxation model based on paramagnetic impurities is problematic. Current low-field NMR technology can detect hydrogen signals in organic components within the core. Therefore, the NMR response in this case is influenced not only by pore fluids but also by the core matrix. Organic matter directly affects the NMR signal in two ways: firstly, it contains hydrogen atoms that can be directly measured, and secondly, it indirectly affects the NMR signal by influencing the relaxation time of the fluid in contact with the pore surface in organic pores. In both cases, the impact of organic matter depends on its maturity. Considering the direct interaction between atoms, maturity is proportional to the mobility of hydrogen atoms. Therefore, organic matter characteristics from gas, oil, or immature organic matter samples will be different. Unlike conventional pores, clay content and distribution in shale reservoirs add another aspect of difficulty to fluid identification. Similarly, clay minerals contain hydrogen atoms, and their NMR response affects the response of attached fluids in many different ways.

The theoretical model established by Bloembergen in 1961 on the relationship between NMR relaxation time and molecular Brownian motion provides a functional relationship between the relaxation rate and the rotational correlation time τ_c of the molecule [54].

$$\begin{array}{l}{\displaystyle \frac{1}{T_1}}={\displaystyle \frac{3}{10}}{\displaystyle \frac{{\gamma }^4{\hslash }^2}{b^6}}[{\displaystyle \frac{{\tau }_c}{1+{\omega }_0^2{\tau }_c^2}}+{\displaystyle \frac{4{\tau }_c}{1+4{\omega }_0^2{\tau }_c^2}}]\\ {\displaystyle \frac{1}{T_2}}={\displaystyle \frac{3}{20}}{\displaystyle \frac{{\gamma }^4{\hslash }^2}{b^6}}[3{\tau }_c+{\displaystyle \frac{5{\tau }_c}{1+{\omega }_0^2{\tau }_c^2}}+{\displaystyle \frac{2{\tau }_c}{1+4{\omega }_0^2{\tau }_c^2}}]\end{array}$$

(1)

where γ is the gyromagnetic ratio, ћ is the reduced Planck’s constant, ω₀ = 2πf, f is the Larmor frequency of ¹H and b is the distance between two adjacent ¹H atoms on the same compound molecule. The relationship between the relaxation time and τ_c, as calculated using Equation (1), is illustrated in Figure 1. According to the figure, as τ_c increases, T₁ initially decreases and then increases, while T₂ continuously decreases. When τ_cω₀ << 1, T₁/T₂ ≈ 1, which corresponds to the characteristics of most light fluids. When τ_cω₀ = 1, T₁ reaches its minimum value. When τ_cω₀ ≥ 1, T₁/T₂ > 1, and the ratio increases with increasing τ_c. This generally corresponds to the region where solid and semi-solid organic materials are located.

Washburn [25,30], Fleury [37], R. Kausik [38], and other experts and scholars have conducted numerous shale relaxation experiments, attempting to use NMR T₁-T₂ distribution to differentiate shale components such as organic matter, gas and water (Figure 2). Their goal is to obtain typical NMR response values for different components. However, due to the influence of factors such as fluid saturation state, organic matter maturity and porosity on the different shale components, the exact positions and shapes on the T₁-T₂ map vary significantly. Therefore, these response characteristic maps cannot be directly and simply applied for shale interpretation.

The different protons in the T₁-T₂ map can be associated with the following sources:

Hydroxyl: These are typically considered to be part of the OH groups in the clay structure or the edges of clay platelets. This signal is always at the resolution limit, below 0.1 ms, and can only be detected with high-precision NMR instruments.

Kerogen: Depending on the maturity, these can overlap with hydroxyl. They are best detected in dry samples as their hydrogen index is relatively low compared to water.

Water: This signal is typically located on or near the T₁ = T₂ line, even for very small pore sizes, such as the interlayer spaces in clay.

2.2. Image Acquisition

The shale material in this experiment comes from Fuling and Liaohe, China. ¹H-NMR spectroscopy was performed using instruments produced by China Numax Technology. The frequency was 23 MHz and the magnet temperature was 32 °C. The NMR T₁-T₂ spectrum was obtained using inversion recovery Carr–Purcell–Meiboom–Gill (IR–CPMG) sequence. Echo time TE was 0.06 ms. The repetition time was 1 s. The number of echoes was 1000.

Shale NMR T₁-T₂ two-dimensional planar spectroscopy is a technique used to analyze the characteristics of shale oil and natural gas reservoirs. Such 2D spectrograms combine information from NMR T₁ relaxation times and T₂ relaxation times to provide detailed insights into reservoir pore structure and fluid type. In the T₁-T₂ two-dimensional plane spectrum, the horizontal axis represents the T₂ relaxation time, and the vertical axis represents the T₁ relaxation time. Each data point represents a measurement sample and its location corresponds to the T₁ and T₂ values of that sample. Representing signal strength or frequency through changes in color or grayscale can reveal relationships between different rock components and fluid types. The two-dimensional spectrum of 3 samples is shown in Figure 3 and Figure 4. During the training process, we added more than one hundred samples of data. By conducting T₁-T₂ experiments and complementary geochemical analyses on core samples in different states (dried, centrifuged, oil-saturated, water-saturated, etc.), the signal properties and their occurrence states can be determined. By comparing the signal changes under different states, we determined what components of the shale the signal represents.

We collected 62 shale samples from Fuling and Liaohe as our dataset, which was divided into training, validation and test sets. The division was performed using a random allocation method. For one hundred data points, we randomly selected 50% of the 2D spectra as the training set, 20% as the validation set and 30% as the test set.

2.3. Network Structure

Deep learning is a subfield within machine learning that focuses on simulating the workings of the human brain using multi-layered neural networks. This approach allows computers to learn based on large amounts of data, automatically discovering useful features and patterns in the data.

Convolutional neural networks are a common form of deep learning. They consist of multiple layers; each layer contains several neurons. Neurons are connected through weights and undergo nonlinear transformation through activation functions. The features learned in each layer are abstract representations of the features in the previous layer. By stacking multiple layers, neural networks can learn increasingly complex features to solve complex problems. Deep learning models have been widely used in many fields, and models in the field of target detection and recognition have also become very mature. Among many models, we chose to use the Faster-RCNN model as a method for identifying NMR T₁-T₂ two-dimensional spectra. There are several reasons for choosing this model.

For data recognition of NMR two-dimensional spectra, both the shape and position of the signals need to be considered. Conventional image recognition methods are not suitable for this. However, Faster-RCNN can address these issues. The principle of this model for object detection is as follows: an image is input and processed through a backbone feature extraction network (DeepConvNet in the diagram), which is a convolutional model, to obtain convolutional feature maps. These feature maps are then processed by the region proposal network (RPN) to generate proposal box regions (Figure 5). These proposal boxes are scaled to a fixed size and then fed into two fully connected layers.

Faster-RCNN, as a representative of the two-stage network structure model, offers higher detection accuracy than classic detection algorithms. Its technology is relatively mature and stable, and it surpasses previous models in image classification and visual key point estimation capabilities.

The network structure of Faster-RCNN includes the following key parts. (1) Convolutional layers: These are used to extract features of the input image. Pretrained convolutional neural networks (such as Visual Geometry Group, Residual Network, etc.) are usually used as feature extractors. (2) Region proposal network (RPN): This is responsible for generating candidate target region proposals. RPN slides a small window at each position and outputs multiple bounding boxes and their corresponding probabilities for each window as candidate target areas. (3) ROI pooling layer: This maps candidate areas of different sizes to fixed-size feature maps for subsequent classification and regression operations. (4) Classification network: This classifies the extracted candidate areas and determines the target category to which they belong. (5) Bounding box regression network: This fine-tunes the bounding box of the candidate area to more accurately frame the location of the target. As shown in Figure 5, the overall workflow is as follows: first, extract the features of the input image through the convolution layer; then, RPN generates a candidate target area frame; next, the ROI pooling layer maps the candidate area to a fixed-size feature map; then, the classification network classifies each candidate area into target categories; finally, the bounding box regression network fine-tunes the position of each candidate area to obtain the final target detection result.

2.4. Algorithm Process

The deep learning method of convolutional neural networks is used to analyze and model organic matter-weighted and organic pore-weighted T₁-T₂ maps. This approach enables the rapid acquisition of information such as organic matter maturity, organic porosity, fluid components and shale content, facilitating identification and quantitative calculations. Initially, the shale T₁-T₂ distribution maps are obtained as inputs for the training data. The data are preprocessed according to the characteristics of the network and the input data. Manually labeled organic pore-weighted T₁-T₂ distributions, organic matter-weighted T₁-T₂ distributions and signal areas of kerogen, oil, water, hydroxyl, etc., are used as samples and data labels. The Faster-RCNN model is then trained using the prepared dataset. The accuracy is assessed to determine if it meets the requirements; if not, the error is calculated and the weights in the Faster-RCNN model are updated, with training continuing until the accuracy meets the requirements. This results in a network model capable of predicting components and content based on different T₁-T₂ distribution maps. Finally, the trained convolutional network is utilized to apply the T₁-T₂ maps for identifying shale components. The technical route of this technology module is illustrated in Figure 6.

3. Experimental Results and Analysis

In this section, we will analyze the parameters of the training results from the 1st, 5th, 10th and 15th iterations. The training graphs are labeled with sections such as kerogen, adsorbed oil, free water, adsorbed water, hydroxyl substances and other objects. The following sections will provide explanations for some of the resulting parameters and demonstrate the impact of the model on these outputs.

3.1. Curve

3.1.1. P_Curve

A P_curve is a graph of the relationship between accuracy and confidence, which represents the accuracy of each category recognition when the confidence is set to a certain value. Theoretically, the greater the confidence, the greater the accuracy. The value of confidence when the accuracy reaches 1 for the 1st, 5th, 8th, 10th and 15th training is as shown in the table. From

Table 1. Accuracy and confidence at different training iterations.

Number of Trainings	Confidence	Precision
1	0.936	1
5	0.939	1
8	0.962	1
10	0.943	1
15	0.950	1

, the value of confidence for the 8th training is the highest.

3.1.2. R_Curve

An R_curve is a graph of the relationship between recall and confidence, which represents the recall probability of each category when the confidence is set to a certain value. Theoretically, when the confidence level is smaller, the category detection is more comprehensive. The value of the recall rate when the confidence reaches 0 for the 1st, 5th, 8th, 10th and 15th training times is shown in

Table 2. The recall rate and confidence at different training iterations.

Number of Trainings	Confidence	Recall
1	0	0.33
5	0	0.90
8	0	0.86
10	0	0.92
15	0	0.90

. The recall rate value of the 10th training is the highest, and the performance of detecting all targets is the best.

3.1.3. F1_Curve

The F1_curve shows the changes in F1_score. An F1_score is an indicator for measuring classification problems. It is the harmonic mean of precision and recall. The F1_score ranges from 0 to 1, which takes into account the accuracy and recall of a specific classification. Therefore, the closer the F1_score is to 1, the better the classification effect. The confidence values corresponding to the best F1 scores for the 1st, 5th, 8th, 10th and 15th training are shown in the table. It can be seen from

Table 3. The F1 score and confidence at different training iterations.

Number of Trainings	Confidence	F1_Score
1	0.786	0.33
5	0.841	0.90
8	0.843	0.85
10	0.816	0.92
15	0.820	0.89

that after the 10th training, the F1 score obtained is optimal and the degree of confidence is 0.816.

3.2. Training Result Indicators

Faster-RCNN is a two-stage classification regression prediction model, and its loss function consists of two parts: region proposal network (RPN) loss and bounding box regression loss. The total loss is also the sum of these two. During the training process, both losses are back-propagated to update the parameters of the model, allowing it to learn to generate accurate region proposals and improve bounding box predictions, resulting in more precise detection results for the model.

In addition, during the model training process, several metrics for object detection are calculated to evaluate the effectiveness of the model. A smaller mean value of the object detection loss function indicates a stronger object detection capability. A smaller mean value of the classification loss function indicates more accurate classification ability. Precision represents the proportion of true positive predictions among the predicted positive samples, while recall represents the probability of the model predicting a positive sample correctly among the actual positive samples. During the model training process, it is also important to monitor the performance of the validation set, including the validation set’s bounding box loss (Val box), mean object detection loss function (Val objectness) and mean classification loss function (Val classification).

The mAP@0.5 represents the average mAP with a threshold greater than 0.5, expressed by the area enclosed by precision and recall as two-axis plots. m represents the average, and the number after @ represents the threshold for determining whether the Intersection over Union (IoU) is a positive or negative sample. mAP@0.5:0.95 (mAP@[0.5:0.95]) represents the average mAP at different IoU thresholds (from 0.5 to 0.95, step size 0.05). The larger the values of mAP@0.5 and mAP@0.5:0.95, the better. The values of various result indicators of Faster-RCNN for the 1st, 5th, 8th, 10th and 15th training are as shown in

Table 4. The various performance metrics of Faster-RCNN at different training iterations.

TN	Box	O	C	P	R	VB	VO	VC	mAP@0.5	mAP@0.5:0.95
1	0.029	0.029	0.005	0.33	0.33	0.02	0.01	0.0020	0.18	0.06
5	0.017	0.022	0.002	0.90	0.90	0.01	0.01	0.0008	0.86	0.37
8	0.017	0.022	0.002	0.85	0.86	0.01	0.01	0.0008	0.81	0.33
10	0.017	0.021	0.002	0.91	0.92	0.01	0.01	0.0008	0.90	0.34
15	0.017	0.021	0.002	0.89	0.90	0.01	0.01	0.0008	0.86	0.38

. It can be seen from Table 4 that during the 10th training of the model, the model’s various detection result indicators for shale are generally the best, but the value of the indicator mAP@0.5:0.95 is not the highest, and the 15th model training result is the best. In Table 4, TN represents the number of training times; Box represents the regression value of the target coordinate frame during the training process; O represents the mean value of the target detection loss function; C represents the mean value of the classification loss function; P represents precision; R represents recall rate; VB represents the bounding box loss of the verification set; VO represents the mean target detection loss function of the verification set and VC represents the mean classification loss function of the verification set.

3.3. Confusion Matrix

The confusion matrix is a summary of the prediction results of the classification problem. It can be easily used to see whether the machine confuses two categories. The confusion matrix can intuitively show the types of errors in the classification model, which helps to overcome the limitations of relying solely on classification accuracy. Each column of the confusion matrix represents the model’s prediction of the class, and each row represents the actual probability of the class.

In this work, the confusion matrix is a summary of the predicted results of the shale classification problem. Using it, it is easy to see if the model confuses two (more) shale categories. The confusion matrix visualizes the types of errors in the classification model and helps overcome the limitations of relying on classification accuracy. Each column of the confusion matrix represents the model’s prediction for a particular shale component class, and each row represents the actual probability of that class. The confusion matrix for the first training result is shown in Figure 7. In the figure, “Background FN” indicates that the model misclassified a category as background, and “Background FP” indicates that the model misclassified background as a category. In order to have a better view of the display, this confusion matrix plot is normalized in the row direction, and also in order to be able to see the recognition and misrecognition rates of each category more intuitively, the values in each row of the confusion matrix are normalized here by dividing by the total number of the corresponding category, expressed as a percentage. The blue gradient bar on the right side indicates the size of the probability obtained above, with darker colors representing higher probabilities. Therefore, the closer the confusion matrix plot is to the dark blue diagonal line from the top left to the bottom right, the better the model’s recognition performance is. As can be seen in Figure 7, the recognition performance of the model for each shale constituent material class in the confusion matrix is very accurate. The only shortcoming is confusing the recognition of background and other materials. It can be seen that after the first training, the classification and recognition ability of the model is already very good. The numbers in the upper right corner of the graph indicate the scaled values multiplied and added to the values on the vertical axis, e.g., 1 × 10⁻⁷ + 0.999999.

The confusion matrix of the fifth training result is shown in Figure 8. It can be seen from Figure 8 that compared with the first model training result, the recognition performance of various substances in the confusion matrix has reached the best, and the classification effect is the best. It fully demonstrates the superior material classification and recognition performance of Faster-RCNN.

The confusion matrix of the 10th training result is shown in Figure 9. From Figure 9, we can know that the model’s ability to distinguish the background and free water in the T₁-T₂ spectrum is weakened. Compared with the 5th model training result, free water has a higher probability of being identified as background.

The confusion matrix of the 15th training result is shown in Figure 10. From Figure 10, it can be seen that compared with the effect of the 10th training, the model has the same effect as the 5th training, and has returned to the best recognition. Except for the confusion matrix renderings produced by the 1st and 10th model training, the confusion matrix training effects of the remaining rounds are relatively excellent. Judging from the confusion matrix effects of the above several model trainings, the model’s recognition and classification performance of shale components do not improve with more training times, but there is only a certain round of training that reaches the best performance.

3.4. Test Result Indicators

We validated the model using sample data collected from different types and instruments than the test set. First, we used shales’ two-dimensional NMR data collected with the same instrument as the training images for prediction (Figure 11). The predictions were all correct, with a majority of them having confidence scores above 0.95. Then, we used shale two-dimensional NMR spectra obtained with different instruments and software for prediction (Figure 11). The validation results were still correct, with the majority of them having confidence scores above 0.8.

In Figure 11, the model successfully segmented and identified different components of the shale. The names of the components are labeled above the identification boxes, and the numbers indicate the confidence level of the identification results. The test results show that the optimal weights can separate the high peak value regions from the background, and the identification of different components is effective. These components include hydroxyl groups, adsorbed water, free water and adsorbed oil. According to NMR theory, there should be no cases where T₁ < T₂; signals in the T₁ < T₂ region are likely artifacts caused by the NMR inversion. To verify the reliability and adaptability of the model, we also tested shale T₁-T₂ maps obtained from different instruments and inversion software (Figure 12). The results were favorable.

However, we can see in the results that when recognizing the 2D spectra from different software, the identification rate of different components in the spectra did not exceed 90 due to slight differences in color labeling. Additionally, due to variations in rock structure and composition among shales from different regions and formations, a large number of samples from various areas is needed for rock physics and geochemical experiments, using the results as a dataset for training.

4. Discussion and Conclusions

We propose a deep learning neural network method for predicting shale composition based on NMR spectra. Compared with traditional techniques, deep learning methods have the advantages of high throughput, accurate predictions and do not require any knowledge accumulation by operators.

Under the optimal weight test, the prediction accuracy of kerogen, structured water, free water, free oil, adsorbed water and adsorbed oil based on the Faster-RCNN model identification of T₁-T₂ map information is more than 0.9. The above results show that the use of the Faster-RCNN model can assist in the detection and quantitative analysis of shale components.

Moreover, experiments have found that the ratio and signal shape of T₁/T₂ are related to the maturity of kerogen. In the future, a large number of two-dimensional NMR spectra of samples with known kerogen maturity levels will be used for training, and the introduction of machine learning algorithms is expected to solve the identification problem of shale organic matter faster and more accurately.