Accurate lithofacies identification in deep shale gas reservoirs via an optimized neural network recognition model, Qiongzhusi Formation, southern Sichuan

Abstract

The Lower Cambrian Qiongzhusi Formation is crucial for exploring deep shale gas in Sichuan, however, challenges in accurately classifying shale lithofacies have hindered its commercialization. To address this, the deep shale reservoirs of the Qiongzhusi Formation were categorized into five lithofacies, five microfacies, and two-sedimentary models utilizing thin sections, scanning electron microscopy, X-ray diffraction (XRD), and petrophysical parameters. Subsequently, various lithofacies identification methods for deep shale gas reservoirs were developed. The recognition performance of triangle and three-dimensional spatial distribution chart methods is poor. The recognition effects of neural network clustering analysis (the testing and validation datasets) are less than 80%, and the training dataset is only 82.6%. On the basis of the trigonometric features, three-dimensional spatial distribution features, and neural network clustering features of the dataset, an optimized neural network lithofacies recognition model was developed. The recognition accuracy of the testing, validation, and training datasets of the ONN model based on the DL principle yielded is greater than 80%. The model achieves a recognition accuracy (training dataset) of 89.9%, with an 85% accuracy rate for blind well lithofacies recognition. This model offers valuable guidance for the exploration and development of deep shale gas in the research area, providing a substantial reference for lithofacies identification in deep shale gas reservoirs of other regions.

Introduction

Lithofacies was initially referred as the lithological characteristics of sedimentary rocks^1,2. Shale composition is largely controlled by the interaction of physio-chemical and biological processes during and after deposition, presenting a large heterogeneity that is intimately associated with lithofacies^3,4. Therefore, compared with traditional oil and gas reservoirs, lithofacies identification is of great significance in deep shale gas exploration^5,6. Accurate identification of shale lithofacies also can directly reflect geological, geochemical, and petrophysical features of deep shale reservoir^7,8.

Prior to the 1990s, mathematical models applied to lithofacies classification involved statistical analysis, linear regression and cluster-analysis techniques⁹. Clustering behavior of logging data was used to distinguish reservoir lithofacies with the aid of a multilayered neural network¹⁰, representing potential of combining clustering analysis with neural networks applications to characterize lithofacies. Combining clustering and principal-component analysis to substantially reduce the number of features also can improve the efficiency of some multilayered neural network models¹¹. In recent decades, machine learning (ML) methods have substantially improved upon statistical and regression methods for lithofacies predictions based on well logs. This commenced three decades ago with neural networks and has diversified¹². In recent years several different types of cluster analysis and neural network methods have been applied for facies interpretation using well log data, but have met with relatively limited success^13,14. Therefore, neural network clustering analysis has recently become a popular tool for lithofacies analysis^15,16,17.

Neural network clustering analysis and DL methods are able to better exploit the non-linear, non-parametric distributions that typify well-log versus lithofacies inter actions, which cannot be achieved with standard statistical and regression methods. However, it can be useful to combine DL models with clustering techniques and other statistical analysis^18,19. In general, optimized neural networks (ONNs) with nonpolynomial smooth activation functions possess trend line approximation features^20,21. The most difficult part of using ONNs in identifications is ensuring that the monitored data flow is adequately matched to the network’s structure and features²².

For the deep shale gas reservoirs in the Qiongzhusi Formation of the Southern Sichuan, the test production of well A103 was 25.86 × 10⁴m³/d, which demonstrates the exploration potential of silty shale. Therefore, in accordance with this thinking, an optimized neural network model (ONNs) will be applied to the lithofacies identification in deep shale gas reservoirs based on the clustering characteristics of lithofacies. And accurate identification of the lithology of the thin black shale section and the thick dark gray silty shale section was carried out, opening up a new prospect of the shale gas exploration.

Regional geological background

Two shale series developed in the Lower Paleozoic era in the Sichuan Basin²³. Proven reserves of more than 2 × 10¹² m³ have been submitted for the Longmaxi Formation, resulting in a considerable breakthrough²⁴. The other set of shale, the Qiongzhusi Formation, has a high potential for thorough exploration (Fig. 1a). The black shale in the condensation section of the intracratonic sag has a thickness of 15–61.5 m and shows a stable distribution. The thickness of the black shale decreases toward the west of the intracratonic sag, and the lithofacies change to silty shale, clay-bearing siltstone, and other related types²⁵. The deposition of the Qiongzhusi Formation in the southern Sichuan Basin is mainly controlled by the Tongwan and Xingkai movements²⁶. Under the influence of the two tectonic movements, the Mianyang–Changning intracratonic sag was formed²⁷. During the depositional period of the Qiongzhusi Formation, an intracratonic sag paleogeomorphology with gentle west and steep east was formed in the continental shelf area under the action of the Xingkai Movement (Fig. 1b), and the marine oxygen content was closely related to the relative sea level change²⁸.

The Qiongzhusi Formation is characterized by three sets of sedimentary cycles of black and dark gray silty shales³⁰. The test production of well A103 was 25.86 × 10⁴ m³/d, indicating the potential for silty shale exploration. Therefore, to identify the exploration targets of the Qiongzhusi Formation, accurate identification of the lithology was performed for both the thin black and thick dark gray silty shale sections, paving the way for new prospects in shale gas exploration.

Fig. 1

(a) Geographical location of the research area (southern Sichuan region); (b) structural geological map of the research area²⁹.

Methodology

Experimental samples and methods

The core samples were obtained from real drilling cores of the Qiongzhusi Formation in the southern Sichuan Basin. Multiple experiments can be performed on the same sample, depending on the core sample requirements of the different experiments. Horizontal bedding developed in the core of the silty shale (Fig. 2a). Muddy veined bands, clumps, and lenses commonly develop in clay-bearing siltstone (Fig. 2b and c). The core samples were specifically sampled based on the lithofacies characteristics of the study area, primarily including full-diameter samples, small column samples, and fragmented samples.

The SEM experiment used an NS3A multifunctional scanning probe microscope to directly observe the secondary electron image of the core sample and analyze the elemental composition at a normal temperature and 45% RH. In the XRD analysis experiment, a DMAX-3 C X-ray diffraction instrument was used to analyze the composition of the clay minerals in the core samples. In the thin section analysis experiment, a DM4500P optical microscope was used to directly observe the microstructure of the core sample thin section.

In the petrophysical parameter experiment, the gas content at the drilling site was measured using a polynomial regression method based on the actual drilled core. Further, samples (small column samples) were taken from the actual drilled core, and the formation conditions in the study area were simulated using the laboratory petrophysical parameter test system to determine porosity and TOC.

Fig. 2

Part of the core photos used for targeted sampling in the research area. (a) silty shale with horizontal bedding, well B1, 3417.92 m; (b) clay-bearing siltstone, well A103, 3360 m; (c) clay-bearing siltstone with abundant veined muddy bands, well A103, 3363.5 m.

Optimized neural networks

As the name implies, ONNs comprise two main keywords: “optimization” and “artificial.” Unlike traditional neural networks, ONNs are referred to as self-configurable or optimizable because they are built to search through a space of hyperparameter search ranges, tending to select the best ONN that attains the performance evaluation for each specific problem statement³¹. Table 1 demonstrates the various optimizable configurations of the proposed ONN. The optimization results of the optimal point hyperparameters show that the number of fully connected layers is 1, and the sizes of the first layers are 1. The activation function is sigmoid, and the minimum classification error is 0.053.

Table 1 Summary of configuration of the ONN.

The selection of activation functions for the ONN model is based on the principle of path optimization. Therefore, during model operation, the activation functions mainly include ReLU, sigmoid, and tanh. The sigmoid function can be used to output hidden layer neurons, which are commonly used for binary classification. The tanh activation function, similar to the sigmoid function, uses true values but compresses them into an interval of -1–1, with the output centered on zero. The ReLU function is a piecewise linear function that compensates for the gradient vanishing problem of the sigmoid and tanh functions and is widely used in deep neural networks. The expressions of the functions are as follows:

$$\:Sigmoid\left(x\right)=\frac{1}{1+{e}^{-x}}$$

(1)

$$\:Tanh\left(x\right)=2Sigmoid\left(2x\right)-1=\frac{2}{1+{e}^{-2x}}-1$$

(2)

$$\:ReLU\left(x\right)=\left\{\begin{array}{c}x,x\ge\:0\\\:\\\:0,x<0\end{array}\right.$$

(3)

where x is the input value.

To establish an ONN classifier based on various experimental study results, it is first necessary to extract predictive variables and responses and process the data into appropriate shapes to train the model. The main output response characteristics are siliceous shale, carbonaceous shale, silty shale, clay-bearing siltstone, and sand-bearing shale. An ONN classification model was established based on the DL theory (Fig. 3). At the input terminal, the data pool primarily comprises XRD experimental and petrophysical parameter experimental analysis results. After data input and data operation processing, the ONN input primarily comprises silicon content, calcium content, clay content, gas content, TOC, and porosity.

Fig. 3

ONN classification model for lithofacies identification of deep shale gas reservoirs.

Results

Analysis results of experiments

The thin section photos of C1 revealed that the quartz particles were developed and evenly distributed (Fig. 4a). Silt lamina was observed in a thin section of A103 of clay-bearing siltstone. The silt content was relatively high (Fig. 4b).

Fig. 4

Thin section photo of well C1 (3133.82 m); siliceous shale with a uniform distribution of quartz particles and well-developed intergranular pores (a). A thin section photo of well A103 (3320.74 m); clay-bearing siltstone with silt lamina (b).

The scanning electron micrograph of Well Z2 showed that particle-shaped potassium feldspar and quartz were intermittently distributed. Further, intragranular pores were observed to develop within the quartz particles (Fig. 5a). Figure 5b shows the corresponding energy spectrum analysis chart. The distribution characteristics of O, Si, and K confirm the existence of quartz and potassium feldspar.

Fig. 5

Scanning electron microscopy image of well Z2 (4600.25 m); particle-shaped potassium feldspar and quartz are intermittently distributed and quartz develops intragranular pores (a). Energy spectral image of well Z2 (4600.25 m), and spectral confirmation of quartz and potassium feldspar (b).

The violin plots were used to analyze the experimental results of the XRD and petrophysical parameter experiments. The white circle represents the median of each parameter data, and the gray black area represents the range from the lower quartile to the upper quartile (Fig. 6). The siliceous content of the core samples was relatively high, with a wide top and a sharp bottom, indicating that the data distribution exhibits a right-skewed peak feature. The distribution of calcium content data followed a gyroscopic pattern, indicating a left-skewed peak characteristic. The clay content data exhibited a gourd-like characteristic, showing a normal distribution (Fig. 6a). The experimental results of the petrophysical parameter experiment revealed a top-pointed and bottom-wide distribution with a left-skewed peak characteristic, and the correlation among the three parameters was good (Fig. 6b).

Fig. 6

Violin diagram of the analysis results of XRD experiments (a) and rock physics parameter experiments (b).

Lithofacies characteristics

The Qiongzhusi shale formation of southern Sichuan is divided into 5 lithofacies based on 476.7-m-long cores and various experimental results. Siliceous shale is mainly developed in three sets of thin black shale of the Qiongzhusi Formation with horizontal bedding, where organic matter is enriched with TOC from 1.03 to 3.16% (Fig. 7a). The mineral composition mainly comprises siliceous minerals (quartz and feldspar), followed by clay and some carbonate minerals. Quartz is mostly symbiotic with organic matter in the form of microcrystals, which have characteristics of biogenic origin (Fig. 7b).

Carbonaceous shale is mainly developed at the bottom of the Qiongzhusi Formation in Changning, Ziyang, and northern Guizhou and is characterized by black organic shale and partial graphitization of organic matter (Fig. 7c). The average content of quartz is 31.88%, the average content of clay is 47.91%, and a layered structure is developed. Organic matter is distributed in clusters and bands, with some organic matter filling the intergranular pores of moldy pyrite (Fig. 7d and e).

Fig. 7

(a) Core photo of A1 well (3304.31 m), black siliceous shale with horizontal bedding. (b) Thin section photo of well A1 (3297.43 m), quartz particles with a maximum particle size of 0.12 mm, as shown by the green line. (c) Core photo of Z2 well (4694.14 m), carbonaceous shale with mirror scratches and high carbon content. (d) Thin section photo of well Z2 (4694.27 m); carbonaceous shale with a layered structure and distribution of clustered and banded organic matter. (e) Scanning electron microscopy image of well Z2 (4693.82 m), carbonaceous shale with organic matter-filled intergranular pores of moldy pyrite.

The horizontal bedding of silty shale is developed, with many off-white clay and calcite lamina layers, bands, and lenses (Fig. 8a). The mineral composition is primarily siliceous minerals, with some carbonates and mica. Owing to the dilution of organic matter by the massive injection of terrestrial silt, the organic matter content is low, with an average of less than 0.5%. The clay content is less than 10–30%. The content of quartz and feldspar silt exceeds 50%, and it is well sorted, featuring subangular and subrounded roundness (Fig. 8b). Mica is directionally arranged amidst the quartz silt particles, which can undergo bending deformation caused by strong compaction (Fig. 8c).

Clay-bearing siltstone often develops calcareous siltstone lenses and nodules with irregular distribution (Fig. 8d). The clay comprises 10–35%, whereas the quartz silt content is about 70%. The calcareous siltstone lenses and bands show the characteristics of high carbonate mineral content. The calcareous cemented and clay laminae are veined and staggered (Fig. 8e). The rock particles have a linear contact relationship, and dissolution pores develop on the surface and edges (Fig. 8f).

Sand-bearing shale is mainly developed at the top of the Qiongzhusi Formation, which has high carbonate mineral content and is mixed with sand and silt (Fig. 8g). When the distal clay is injected, the organic matter is not enriched. The average TOC content is 0.3%, the clay mineral content is about 50%, and the siliceous mineral content is 25–35%. The silt particle content is 10–25%, which primarily comprised heteromorphic siliceous minerals. The grain size content of the sand particles is approximately 50%, primarily exhibiting a laminated structure (Fig. 8h) and occasionally featuring surface solution pores in feldspar (Fig. 8i).

Fig. 8

(a) Core photo of A1well (3484.35 m), silty shale with calcareous lamina. (b) Thin section photo of well A1 (3484.85 m), calcareous siltstone band. (c) SEM image of A1 well (3500.93 m), point-line contact among silt particles. (d) Core photo of A103 well (3370 m), clay-bearing siltstone with abundant calcareous lenses. (e) Thin section photo of well A103 (3394.37 m), clay-bearing siltstone with calcareous concretion. (f) SEM image of A103 well (3384.03 m), particle line contact, with dissolution pores developed on the surface and edges. (g) Core photo of B1 well (3394.12 m), sand-bearing shale. (h) Thin section photo of well B1 (3382.01 m), sand strip. (i) SEM image of B1 well (3392.56 m), surface solution pores of felspar.

Characteristics of sedimentary facies

Deep-water shelf

The deep-water continental shelf refers to the anoxic deep-water sedimentary area below the storm wave base that has the weakest hydrodynamic conditions and the lowest energy^32,33. Based on the characteristics of mineral composition, organic matter content, and maturity, it can be further divided into siliceous and carbonaceous deep-water shelves (DWSs).

The microfacies of the siliceous deep-water continental shelf are mainly siliceous shale (Fig. 9a), which is developed in the Jingyan–Qianwei area on the west side of the intracratonic sag. The 10-cm-thick gravel-bearing medium-coarse sandstone formed by transgression erosion is usually observed at the bottom, which shows abrupt contact with the underlying gray siliceous shale (Fig. 9b). The carbonaceous DWS microfacies are dominated by organic-rich carbonaceous shale, which is developed in the Ziyang and Changning areas in the intracratonic sag (Fig. 9c and d). The sedimentary water body is deep, and the organic carbon content ranges from 3.2 to 4.5% (Fig. 9e).

Fig. 9

Characteristics of sedimentary structures within DWS facies. (a) Core photo of A103 well (3584.74 m), siliceous shale with horizontal bedding and pyrite. (b) Core photo of A103 well (3425 m), transgression erosion. (c) Core photo of Z2 well (4694.04 m), carbonaceous shale. (d) Core photo of Z2 well (4693.94 m), carbonaceous shale with mirror scratches, and pyrite particles. (e) SEM image of Z2 well (4693.98 m), numerous organic matter develops in the intergranular pores.

Semi-deep-water shelf

The semi-deep-water shelf (SDWS) refers to the sedimentary area near the redox depth below the average wave base, with relatively shallow water, silt-rich sediments, relatively coarse grain size, enhanced hydrodynamics, and intermittent turbulence³⁴. Based on the lithofacies characteristics, it can be further divided into three sedimentary microfacies: sandy SDWS, sand-muddy SDWS, and semi-deep-water turbidite sand.

The sand-muddy SDWS microfacies consist mainly of gray-black clay-bearing siltstone and silty shale interbedded with numerous gray calcareous beds. Silty shale also exhibits obvious bedding and horizontal bedding development (Fig. 10a and b). Flaky and lenticular bedding is occasionally observed. The microfacies of argillaceous SDWS are mainly sand-bearing shale, clay-bearing siltstone, and silty shale, with low organic carbon content (Fig. 10c and d).

Fig. 10

Characteristics of sedimentary structures within sand-muddy semi-deep-water shelf and argillaceous semi-deep-water shelf. (a) SEM image of A103 well (3375.9 m), mica with layered structure. (b) Core photo of A103 well (3394.2 m) with horizontal bedding. (c) Core photo of A103 well (3275.99 m) with calcareous sandstone band. (d) SEM image of A103 well (3265.93 m), an extremely fine crystalline calcite exhibits inlaid contact and develops dissolution pores.

The semi-deep-water turbidite sand microfacies primarily comprise clay-bearing siltstone and sand-bearing shale, both of which are episodic events. The local hydrodynamic force is strong during the strong regression period, causing the near-source calcareous silt current to be transported into relatively deep water to form irregular lens bodies (Fig. 11a) and a band. Simultaneously, soft-sediment deformation can be observed: convoluted bedding (Fig. 11b), flame structure (Fig. 11c), and oblique bedding formed under strong hydrodynamic conditions (Fig. 11d).

Fig. 11

Characteristics of sedimentary structures within semi-deep-water turbidite. (a) Core photo of A103 well (3372.58 m) with calcareous siltstone lens. (b) Core photo of A103 well (3574.93 m) with convolute bedding. (c) Core photo of A103 well (3539.32 m), flame structure. (d) Core photo of A103 well (3580.94 m), oblique bedding.

Sedimentary model

Based on the aforementioned sedimentary facies and lithofacies analysis, the sedimentary model of the Qiongzhusi Formation in southern Sichuan is established under the background of the Mianyang–Changning intracratonic sag.

During the depositional period of the deep continental shelf, the water was deeper and the upwelling provided a various nutrient. The primary productivity is high, the origin supply is less, the oxygen-deficient zone is shallow and wide, and black siliceous and carbonaceous shales are mainly deposited. Organisms flourish and organic matter is enriched. Because the intracratonic sag was filled during the process of late sedimentary evolution, the sea level slowly dropped, and the deep-water continental shelf gradually evolved into the shallow-water continental shelf (Fig. 12a).

During SDWS deposition, the water level fluctuated frequently and was occasionally affected by storms and waves. The strong surface water is fully exchanged with the bottom water, and the oxygen content is high so that the anoxic zone of the seafloor cannot be formed. Although it is unsuitable for organic matter preservation, it easily forms a silt grain-supported texture (Fig. 12b).

Fig. 12

Deep-water shelf sedimentary model (a) and semi-deep-water sedimentary model (b) of the Qiongzhusi Formation in southern Sichuan.

Discussion

Based on lithofacies characteristics, sedimentary facies characteristics, and sedimentary model of deep shale gas reservoirs in the research area, lithofacies identification was conducted using the methods of triangle diagram, three-dimensional spatial distribution, neural network clustering, and ONN. This indicates a research progression from qualitative to quantitative and from inaccurate to accurate.

Triangle and three-dimensional spatial distribution chart methods

The trigonometric method (silicon content, calcium content, and clay content of XRD experiments) was used to distinguish the lithofacies of deep shale gas reservoirs in the study area. Carbonaceous shale has a clear boundary with the other four types of lithofacies. Siliceous shale is mainly distributed in the yellow area. Silty shale, clay-bearing siltstone, and sand-bearing shale have similar distribution characteristics, making it difficult to accurately identify them using the triangle diagram (Fig. 13a).

A three-dimensional spatial distribution chart of the data is drawn based on the experimental results of the physical parameters of the rock (gas content, TOC, and porosity). The three-dimensional spatial distribution between siliceous and carbonaceous shales has clear boundaries and can be distinguished from the other three lithofacies. The three-dimensional spatial distribution characteristics of silty shale, clay-bearing siltstone, and sand-bearing shale are relatively similar, and the recognition effect is poor (Fig. 13b).

Fig. 13

Triangle analysis of XRD experimental results (a) and three-dimensional correlation analysis of petrophysical parameter experimental results (b).

Neural network clustering analysis

Neural network cluster analysis is used to classify a substantial dataset into meaningful subsets so that the statistics in each subset have identical properties and are distinct from each other³⁵. In previous studies, neural network clustering methods have been applied to reservoir quality classification, tight sandstone lithofacies classification, and seismic facies classification^36,37. Cross-validation is a statistical tool that is used to ensure the accuracy of modeling and prediction procedures in the statistical modeling³⁸. Cross-validation can be adopted in three different ways: Random Subsampling, K-Fold Cross-Validation, and Leave-one-out Cross-Validation^39,40. Leave-one-out Cross-Validation is performed through splitting the original dataset into three parts: the train, validation, and test^41,42.

Figure 14a shows the neural network clustering neighboring point weight analysis for lithology identification of deep shale gas reservoirs in the study area, with classification results divided into five categories based on weight color (siliceous shale, carbonaceous shale, silty shale, clay-bearing siltstone, and sand-bearing shale).

The recognition accuracy of the testing and validation datasets respectively are 75.4% and 78.3%, and only the recognition accuracy of siliceous shales is good (Fig. 14b and c). The recognition effect of neural network clustering analysis (training datasets) is 82.6%, with a recognition rate of over 90% for siliceous and carbonaceous shales. The recognition effect on silty shale and clay-bearing siltstone is good, with recognition rates of 82.4% and 80%, respectively. The recognition effect of sand-bearing shale is poor, with a recognition rate of less than 50% (Fig. 14d).

Fig. 14

Neural network clustering neighbor weight distance graph (a) and confusion matrix of the testing (b), validation (c), training datasets (d).

Optimizable neural network method

In the process of the ONN model for lithological identification, the L2-norm (Euclidean norm) is used to normalize the dataset, and Fig. 15a shows parallel coordinates of the lithological identification feature parameters. The distribution characteristics of the characteristic parameters have considerable differences for siliceous and carbonaceous shales. The characteristic parameter distribution characteristics of silty shale and sand-bearing shale are relatively similar, and they have certain similarities with clay-bearing siltstone. The input data for the model are experimental data with strong reliability that are closely related to shale lithofacies. However, not all wells in the study area have this data, and the reliability of the data also requires manual screening.

The ONN deep shale gas reservoir lithology identification model underwent 30 iterations. In the 22nd iteration, the optimal point hyperparameter and the minimum error hyperparameter appeared, achieving the best recognition effect and efficiency (Fig. 15b).

Fig. 15

Parallel coordinates of lithological identification feature parameters (a) and minimum classification error graph for optimizing neural network models (b).

Figure 16a, b, and c respectively show the confusion matrix results of the testing (10%), validation (25%), and training (65%) datasets for the optimized neural network model. The optimized neural network significantly improves the recognition performance of deep shale lithofacies. The recognition accuracy of the testing and validation datasets respectively are 82.6% and 84.1%, which is better than the neural network clustering model (Fig. 16a and b).

The recognition effect of the ONN deep shale gas reservoir lithology recognition model (training datasets) is 89.9%, while siliceous shale, carbonaceous shale, silty shale, and clay-bearing siltstone all have a recognition effect greater than 85%. Because sand-bearing shale and clay-bearing siltstone have very comparable lithological features, their recognition effect is relatively poor (Fig. 16c).

The false positive rates of the working points are all close to 0 for the ROC curves of the five lithofacies. The AUCs of siliceous shale, carbonaceous shale, silty shale, and clay-bearing siltstone are all greater than 0.9, and the AUC of the sand-bearing shale is 0.8076 (Fig. 16d). This indicates that the reliability of the model is high and the recognition effect is good.

Fig. 16

Confusion matrix of the testing (a), validation (b), training datasets (c) and receiver operating characteristic curve (d) of optimized neural network models.

Workflow and Identification results

Workflow

Research workflow can provide the best research ideas and help understand the novelty of the entire research work^43,44,45,46. To the best of our knowledge, the accuracy of ONNs in diagnostic procedures is substantially greater than that of multilayer perceptron, radial basis function, and convolutional neural networks. This research pioneered the integration of neural network clustering learning and Deep learning (optimized neural networks and stochastic search reverse design algorithms simultaneously) to make the lithofacies identification of deep shale gas reservoirs more time efficient and effective. The workflow is shown in Fig. 17. The input data for the workflow includes gas content, TOC, brittle minerals, clay minerals, and porosity that are closely related to shale lithofacies, all of which are experimental data with strong reliability. The workflow running process not only reflects the progressiveness of the DL method, but also considers the characteristics of clustering analysis. The output of the workflow is constantly iterating, and the recognition efficiency and accuracy both obtain the optimal path. This workflow is customized for shale lithofacies identification and is not applicable to sandstone and carbonate rocks.

Fig. 17

Integral research principles and flowchart.

Identification results

The lithofacies of the Qiongzhusi Formation were identified using different methods in the Z2 well (blind well). Figure 18 shows comprehensive histogram of lithofacies identification. The agreement rate is calculated between the identification results of the three methods and the lithology profile (Eq. 4).

$$\:LAR=\left({\sum\:}_{i=1}^{n}\frac{Thickness\left(ALI\right)}{Thickness\left(Identification\right)}\right)\%$$

(4)

where LAR is the agreement rate with the lithology profile, i is the lithofacies identification section number, thickness (ALI) is the accurately identified thickness with the lithology profile, and thickness (Identification) is the identified thickness.

The agreement rate is calculated between the identification results of the three methods and the sedimentary microfacies profile (Eq. 5).

$$\:SMAR=\left({\sum\:}_{i=1}^{n}\frac{Thickness\left(ASMI\right)}{Thickness\left(Identification\right)}\right)\%$$

(5)

where SMAR is the agreement rate with the sedimentary microfacies profile; thickness (ASMI) is the accurately identified thickness with the sedimentary microfacies profile.

The qualitative identification results of the triangle and three-dimensional spatial distribution diagrams have a consistency rate of 75.6% with the lithology profile and 80.2% with the sedimentary microfacies profile. The agreement rate is 79.5% between the neural network clustering analysis results and the lithological profile, and the agreement rate with the sedimentary microfacies profile is 82.4%. The agreement rate between the recognition results of the ONN model and the lithology profile is 85.3%, and the agreement rate with the sedimentary microfacies profile is 87.6% (Fig. 18). Therefore, the deep shale gas reservoir lithofacies identification model of the ONN has improved the agreement rate between the recognition results and the actual situation (the lithological profile and the sedimentary microfacies profile).

Fig. 18

Comprehensive histogram of Z2 well lithology classification results.

The ONN model is a stable, reliable, and effective lithofacies identification method for deep shale gas reservoirs in continental shelf facies, which improves the accuracy of lithofacies identification and is more conducive to the efficient development of deep shale gas reservoirs in the study area. It supports the efficient exploration and development of Well Z2, which has recently achieved a daily production of over 100 × 10⁴ m³/d. Simultaneously, it provides a new approach for lithofacies identification of deep shale gas reservoirs and lays the foundation for exploring more DL methods for different fields and stages of oil and gas field development in the future.

Conclusions

Through accurate lithofacies identification based on an ONN recognition model in deep shale gas reservoirs of the Qiongzhusi Formation, southern Sichuan, the following conclusions are drawn.

The deep shale gas reservoir of the Qiongzhusi Formation is classified into five lithofacies. The organic matter in siliceous shale is enriched, and the mineral composition is primarily biogenic siliceous minerals. The Qiongzhusi Formation is formed within the depositional environments of continental shelf facies, and the siliceous and carbonaceous DWSs are more conducive to the production, enrichment, and preservation of shale gas.

The two-sedimentary model is established. The DWS sedimentation model has a deeper water level, and the rising water flow contains various nutrients. Water level fluctuations are frequent in the SDWS sedimentation model. The gas generated by favorable lithofacies in the DWS sedimentation model can be enriched and preserved in the sandy texture generated by the SDWS sedimentation model.

Various lithofacies identification methods for deep shale gas reservoirs were developed. The recognition performance of triangle and three-dimensional spatial distribution chart methods is poor, only able to distinguish between siliceous shale and carbonaceous shale. The recognition effects of neural network clustering analysis are less than 85%, with relatively good recognition performance for siliceous and carbonaceous shales.

The recognition accuracy of the testing, validation, and training datasets of the ONN model based on the DL principle yielded is better than the neural network clustering model. A confusion matrix result of training datasets is 89.9%, with an 85% coincidence rate with the actual case. The shale lithofacies identification of this model is more accurate, reliable and efficient.