Research on Deep Learning-Based Identification Methods for Geological Interface Types and Their Application in Mineral Exploration Prediction—A Case Study of the Gouli Region in Qinghai, China

Abstract

Geological interfaces are crucial elements governing deposit formation, such as silica–calcium surfaces, intrusive contact interfaces, and unconformities can serve as key symbols for mineral exploration prediction. Geological maps provide relatively detailed representations of primary geological interfaces and their interrelationships. However, in previous mineral resource predictions, the type differences in different geological interfaces were ignored, and the types of different geological interfaces vary greatly, thus affecting the validity of the mineral prediction results. Manual interpretation and analysis of geological interfaces involve substantial workloads and make it difficult to effectively apply the rich geological information depicted on geological maps to mineral exploration prediction processes. Therefore, this study proposes a model for intelligent identification of geological interface types based on deep learning. The model extracts the attribute information, such as the age and lithology of the geological bodies on both sides of the geological boundary arc, based on the digital geological map of the Gouli gold mining area in Dulan County, Qinghai Province, China. The learning dataset comprising 5900 sets of geological interface types was constructed through manual annotation of geological interfaces. The arc segment is taken as the basic element; the model adopts natural language processing technology to conduct word vector embedding processing on the text attribute information of geological bodies on both sides of the geological interface. The processed embedding vectors are fed into the convolutional neural network (CNN) for training to generate the geological interface type recognition model. This method can effectively identify the type of geological interface, and the identification accuracy can reach 96.52%. Through quantitative analysis of the spatial relationship between different types of geological interfaces and ore points, it is known that they have a good correlation in spatial distribution. Experimental results show that the proposed method can effectively improve the accuracy and efficiency of geological interface recognition, and the accuracy of mineral prediction can be improved to some extent by adding geological interface type information in the process of mineral prediction.

1. Introduction

Geological interfaces serve as records of geological evolution processes and significant geological events, while also holding great importance for locating specific minerals and resolving related geological issues [1]. Geological interfaces are closely related to the formation of mineral deposits. Certain geological interfaces serve as significant mineral-forming interfaces, such as silic-calcic surfaces and angular unconformable interfaces. Extensive studies on deposit characteristics indicate that numerous mineral deposits occur on specific geological interfaces [2]. The ore bodies of the Qixia Mountain lead–zinc polymetallic deposit in Jiangsu Province are primarily controlled by the silica–calcium surface between the Huanglong Formation limestone of the Carboniferous System and the Gaolishan Formation sandstone [3]. The ore body of the Getang gold deposit in southwest Guizhou is lenticular and stratiform and occurs on the unconformity interface between the Longtan Formation and the Maokou Formation [4]. The enrichment of ore bodies in the Xiangshan uranium deposit in southeastern China tends to occur along ring faults, interfaces, unconformities, and geological contacts intersecting with faults [5]. The pyrite layer at the Dongguashan Copper Mine in Anhui Province occurs at the silicon–calcium interface between the clastic deposits of the Wutong Formation in the Devonian System and the carbonate deposits of the Huanglong Formation in the Carboniferous System [6]. In the Canadian Athabasca Basin, unconformity-type uranium deposits are located near unconformities between Archean–Paleoproterozoic metamorphic basement rocks [7]. Mineral exploration practice has demonstrated that geological interfaces serve as key indicators for prospecting, predictive targets, and geological spaces for deposit localization, providing explicit guidance for the geological positioning of undiscovered deposits [2].

To date, most research has primarily focused on the control of geological interfaces over deposit distribution, while systematic studies on interface identification methods remain relatively scarce. Dong et al. [8] used the method of combining computer technology and GIS to identify stratigraphic contacts, which improved the efficiency of discrimination of formation contact relationships. Guo et al. [9] utilized remote sensing satellite data combined with DEM to identify linear structures (faults). Mukherjee and Roy [10] used Walsh transform technology to identify stratigraphic boundaries based on logging curves, which improved the resolution of formation boundary detection. Previous studies on geological interfaces primarily focused on common stratigraphic contacts or faults, heavily relying on data completeness and accuracy. These studies also lacked deep integration with other geological datasets, making it difficult to meet the comprehensive needs of complex geological research, such as regional geological studies and mineral exploration prediction.

The multi-source, multi-modal nature of geological survey data results in large data volumes and complex data relationships [11]. For a long time, mineral resource exploration primarily relied on manual interpretation and analysis of geological information, leading to low identification efficiency. With the rapid advancement of information technology, significant achievements have been made in artificial intelligence and natural language processing, especially in generative language models. These technological advancements have profoundly impacted numerous industries and become powerful tools for researchers [12]. The emergence of artificial intelligence has brought mineral exploration prediction into a new stage. That is, by leveraging the powerful and intelligent data processing and analysis advantages of computers, it helps clarify the long-accumulated geological “big data” and the geological models formed by geological experts. This provides robust auxiliary decision-making support for quantitative mineral resource prediction.

Convolutional neural network (CNN), as one of the most widely applied neural network models, can autonomously learn the features required for classification and recognition from training data [13], improving classification accuracy and efficiency without relying on manual feature selection. CNNs have mature applications in lithological identification [14], geological mapping [15], and three-dimensional geological structure inversion [16]. CNN has also been widely used in prospecting and prediction research. Shi et al. [17] used deep convolutional neural network algorithms to uncover coupled correlations among the spatial distribution of chemical elements, sedimentary facies, outcrops, faults, water systems, and manganese ore body occurrences. They delineated five mineral-promising areas in the Songtao–Huayuan study region in Guizhou, successfully achieving two-dimensional mineralization prediction with this method. He et al. [18] extracted the features of metallogenic factors by using an integrated learning synthetic convolutional neural network algorithm and a self-attention mechanism algorithm. After training, they obtained a network model with an accuracy of more than 94%, delineating four of the most favorable prospecting prediction zones in the Bawanggou gold deposit area of Shaanxi Province. Li et al. [19] extracted spatial distribution characteristics of soil geochemistry in the Daqiao area of Gansu Province by moving window method, and predicted seven favorable gold exploration target zones based on a deep convolutional neural network model. Convolutional neural network (CNN) has achieved remarkable progress in mineral exploration. Their powerful feature extraction capabilities can effectively process complex geological data, significantly improve prediction accuracy and exploration efficiency of mineral resources, which provides robust technical support and theoretical foundations for geological prospecting.

The Gouli area in Qinghai Province, China, possesses favorable geological conditions for mineralization and abundant mineral resources. After many rounds and types of geological surveys, the region has accumulated extensive exploration data. However, due to the unclear relationship between the multi-stage orogenic process and metallogenic system evolution, complex metallogenic type, and unknown metallogenic model in the Gouli area [20], it is difficult to discover new mineral resources, which limits the further development of mineral geological investigation to a certain extent. Given the outstanding advantages of one-dimensional convolutional neural networks in information analysis, this paper proposes an intelligent model for classifying and identifying geological interface types. This model automates and enhances the efficiency of identifying contact boundaries between geological bodies, significantly improving prospecting efficiency.

2. Geological Background of the Study Area

The Qinghai Gouli Mineralized Cluster Zone is situated in the eastern segment of the East Kunlun Orogenic Belt, spanning both the Kunbei and Kunnan terranes [21]. The study area possesses a Proterozoic basement and has undergone complex tectonic movements during the Paleozoic and Mesozoic eras, resulting in an extremely intricate internal structure characterized by a multi-island ocean, soft collision, and multi-cycle orogeny [22]. These features provide favorable conditions for gold and polymetallic mineralization [20]. The study area is rich in metallic mineral resources, primarily including gold, silver, lead, zinc, copper, cobalt, nickel, and iron. Among these, gold resources are exceptionally abundant, with multiple medium–large metal deposits identified locally, including the Gouluolongwa, Walega, and Annage deposits [23] (Figure 1).

The exposed strata in the region include: the Pre-Proterozoic Jinshuikou rock group and Kuhai rock group; the Neo-Proterozoic Wanbaogou group; the Lower Paleozoic Nachitai group; and the Upper Paleozoic, including the Devonian Maoniushan Formation, Carboniferous Halaguole Formation, Permian Haoteluowa Formation, Maerzheng Formation, and Gequ Formation. The Mesozoic Era encompasses the Triassic Hongshuichuan Formation, the Naocangjiangou Formation, the Elashan Formation, and the Jurassic Yaojie Formation; the Cenozoic Era includes the Neogene Guide Group and Quaternary deposits. Among these, the Jinshuikou Group of the Paleoproterozoic and the Nachitai Group of the early Paleozoic constitute the primary gold-bearing strata [25,26].

Fault structures are highly developed in the study area, with the Kunzhong Fault and Kunnan Fault being the largest and serving as the regional main structural framework, influencing the distribution of strata and igneous rocks [21] (Figure 1). Influenced by the Kunzhong Fault and Kunnan Fault, multiple sets of NWW-, NW-, and NE-trending faults have developed in the study area [27]. Among these, the NWW-trending faults are the most significant secondary faults. Represented by the Xiangride–Delong Fault, it is closely associated with mineralization, with the most gold deposits distributed along both sides of this fault [21,27]. Magmatic activity in this region is intense and prolonged, persisting from the Proterozoic through the Mesozoic. Various volcanic and intrusive rocks are widely distributed throughout the area [21], with the Variscan and Indosinian periods exhibiting the most frequent and vigorous magmatic activity. This formed the largest and most extensive meso-acidic intrusive bodies within the region [26], which are markedly controlled by faults and predominantly distributed along fault zones [22].

3. Intelligent Identification Methods and Results for Geological Interface Types

3.1. Classification of Geological Interface Types

A total of 34 mineral deposits have been identified within the study area, including 19 gold deposits, 3 molybdenum occurrences, 6 iron occurrences, and 6 copper occurrences. Based on geological data and mineral exploration reports from the study area, the exposed strata in the Walega Gold Mine district are dominated by the Jinshuikou Group, characterized by intense magmatic activity. The intrusion of Variscan granodiorite not only supplied hydrothermal fluids for mineralization but also played a crucial controlling role in gold mineralization [28]. The Ewate iron–copper mineralization belt occurs near the contact zone between granodiorite and porphyritic granite. Known molybdenum deposits are primarily distributed in the contact zones of various geological bodies from the Variscan to Indosinian periods, closely associated with Late Triassic medium–acidic volcanic–magmatic mineralization [29]. The primary source layer for the Yerigeng gold–copper–iron deposit is the Early Carboniferous Halaguole Formation, where elemental background values significantly exceed crustal abundance values. Concurrently, Variscan Period intermediate acidic intrusive rocks are the primary drivers of gold and nickel mineralization [30]. The Shanglonggang iron deposit area exposes Carboniferous carbonate and sandstone–shale formations. Ore bodies occur in faults of limestone and carbonaceous slate and are strictly controlled by structures. The strata exposed in the Dulenggou copper–cobalt mine area are mainly the Wanbaogou rock group of the Mesoproterozoic and Neoproterozoic. The copper–cobalt ore bodies discovered are all produced in andesite or in the contact position between andesite and schist [31]. All gold ore bodies discovered in the Dareer Gold Prospecting Area occur within NNE-trending structural alteration zones, while silver ore bodies are found in near NW-trending zones [32]. The Delong Gold Mine is located within a regional NW-NWW-trending gold-bearing structural alteration zone that has undergone multiple episodes of activity [33].

Based on the geological characteristics and mineral exploration practices in the Gouli area, the geological interfaces in the study area are classified into 4 major categories and 9 subcategories (

Table 1. Geological interface classification table.

Major Categories	Subcategories	Labels
Structural interfaces	Fault	0
Stratigraphic interfaces	Conformable	1
	Parallel unconformable	2
	Angular unconformable	3
Intrusive contact interfaces	/	4
Special lithological interfaces	Limestone and siliceous rocks	5
	Limestone and moderately acidic intrusive rocks	6
	Volcanic–sedimentary interface	7
	Eruption coverage	8

), including the following: (1) structural interfaces—fault; (2) stratigraphic interfaces—conformable, parallel unconformable, and angular unconformable; (3) intrusive contact interfaces; and (4) special lithological interfaces—limestone interfaces (contacts with siliceous rocks and intermediate–acidic intrusive rocks) and volcanic rock interfaces (volcanic–sedimentary interfaces and eruption cover layers). That is, the classification labels for the training and test datasets.

3.2. Intelligent Identification Method for Geological Interface Types

The process of geological interface identification based on one-dimensional convolutional neural networks primarily involves the following six steps (Figure 2):

• Data Preparation and Processing

Systematically collect 1:100,000 digital geological maps of the study area, encompassing vector data (geological boundaries, structural lines) and attribute databases (stratigraphic units, rock mass descriptions). Use MapGIS6.7 for data quality inspection, resolving issues such as intersecting geological boundaries and missing attributes through topological checks. This process ultimately generates a standardized, high-precision digital geological map dataset.

2.

Extraction of Geological Interface Data

Based on the topological structure of the digital geological map, Python 3.10.13’s Shapely library is used to automatically identify and segment geological boundaries, treating each arc segment as an independent analysis unit. The ID numbers and attribute information of geological bodies on both sides of each arc segment are extracted separately.

3.

Construction of Training and Validation Datasets

For the ID numbers and attribute information of geological bodies on both sides of the arc segments, interface types are annotated based on the geological information on either side. The label data consists of a series of geological interface types, as defined in Section 3.1. The final dataset includes geological boundary IDs, textual attributes of geological bodies on both sides, and mineralization geological interface type labels (e.g., “Late Triassic porphyritic diorite–Carboniferous gray medium-thick bedded limestone–limestone and moderately acidic intrusive rocks”, and “Late Triassic gray-purple rhyolitic breccia tuff–Paleoproterozoic light pinkish biotite gneiss–eruption coverage”).

In the study, three researchers classified the geological interfaces (with a total of eight categories, each geological interface having only one interface type). To verify the consistency of the evaluators’ annotations on the dataset, we use Fleiss’ Kappa [34] coefficient for quantitative analysis. This article will adopt Equations (1)–(3) to calculate the Fleiss’ Kappa coefficient.

$$P_o={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N{\displaystyle \frac{1}{n(n-1)}}{\sum }_{k=1}^K{n}_{ik}(n_{ik}-1)$$

(1)

$$P_e={\sum }_{k=1}^K{\left[{\displaystyle \frac{1}{N_n}}{\sum }_{i=1}^N{n}_{ik}\right]}^2$$

(2)

$$kf={\displaystyle \frac{P_o-P_e}{1-P_e}}$$

(3)

where $N$ is the number of samples, $n$ is the number of evaluators, $K$ is the number of categories, and $n_{ik}$ is the number of times the i-th sample is classified into the k-th category.

After calculation, the Fleiss’ Kappa coefficient is 0.901. According to the classification standard proposed by Landis & Koch [35], this coefficient corresponds to almost perfect agreement, confirming that the labeled data has good reliability and can serve as the core data support for subsequent model training, experimental analysis, and conclusion derivation.

4.

Semantic Vectorization of Datasets

Convert unstructured text data into numerical features processable by convolutional neural networks. Based on the Sentence Transformer text embedding model [36], fine-tune it to adapt to geological terminology. Vectorize geological body attribute texts to generate embedding vectors, preserving semantic information. For interface type labels, a “text label-numeric code” mapping table is established (e.g., “fault” → 0, “angular unconformity” → 3). Discrete categorical labels are converted into a continuous vector space using one-hot encoding [37]. This data format enables the convolution kernel to slide along the sequence dimension, facilitating the model’s learning of inter-category differences.

5.

Construction of the CNN1D Network Model

This paper designs and implements a CNN-based model for geological interface recognition. The model consists of 3 convolutional layers, 3 pooling layers, and 1 fully connected layer.

The input layer receives preprocessed embedding vectors, feeding them into the network for feature extraction and learning.

Convolution layers extract textual features by processing data in parallel using convolutional kernels of varying sizes to capture multi-scale semantic information. The convolution operation is defined as in Equation (4).

$$y[i]=ReLU(\sum _{j=0}^{j-1}x\left[i+j\right]\times w\left[j\right]+b)$$

(4)

where $x\left[i\right]$ is the input sequence, $w\left[j\right]$ is the convolution kernel weights, $b$ is the bias term, and $ReLU\left(x\right) = max (0, x)$.

The essence of a pooling layer is downsampling. After multiple convolutions, the data generates a large number of parameters, which not only significantly increases the difficulty of network training but also easily leads to overfitting. Therefore, placing a pooling layer after the convolutional layer compresses the data, reduces its dimension, and decreases the number of parameters. The pooling layer uses max pooling to reduce the output dimension of the convolutional layer by a fixed stride. It retains the maximum feature value within each window, enhancing feature robustness. The max pooling operation is defined as in Equation (5).

$$y\left[i\right]=max\left(x\right(2i),x(2i+1\left)\right)$$

(5)

where $x\left[i\right]$ is the input sequence, $y\left[i\right]$ is the pooled output.

The dropout layer is introduced to reduce the number of parameters, accelerate model training, and mitigate overfitting. The hidden dropout rate is set to 0.3.

Through multiple convolutional and pooling operations, dimensionality reduction is applied to the features extracted by the network model. Finally, a fully connected layer is established to unfold the feature matrix into a one-dimensional vector, which is then mapped to the geological interface type sample space. The fully connected layer operation is implemented by Equation (6).

$$y=\mathrm{R}\mathrm{e}\mathrm{L}\mathrm{U}\left(W\times x+b\right)$$

(6)

where $W$ is the weight matrix, $x$ is the input vector, $b$ is the bias vector, and $ReLU\left(x\right) = max(0, x)$.

The output layer uses the Softmax activation function (Equation (7)) to produce the final geological interface classification results.

$$y\left[i\right]={\displaystyle \frac{exp\left(z\left[i\right]\right)}{{\sum }_{j=1}^{j}exp\left(z\left[j\right]\right)}}$$

(7)

where $y\left[i\right]$ represents the probability of category $i$, $z\left[i\right]$ denotes the raw output value of the final layer, $j$ is the total number of categories in the output dimension, and $exp$ is the exponential function.

6.

Model Training and Parameter Optimization:

To ensure model generalization capability, a stratified sampling strategy is used to proportionally divide different interface types into a 70% training set and a 30% validation set. Data augmentation techniques (such as random text perturbation and SMOTE (Synthetic Minority Over-Sampling Technique)) are applied to expand sample diversity. The model employs L2-regularization and early stopping to mitigate the risk of overfitting. Different parameters are used to train the model, and the optimal model is selected.

3.3. Test Results

This study uses the CNN1D model to identify and classify arc segments within the Gouli research area. During model training, the Adam optimizer is used for parameter updates, with the loss function set to categorical cross-entropy. Training is conducted for 100 iterations with a batch size of 32. The study area contains 9 interface types and 1375 geological interface units. Over 5900 geological interface relationships are constructed, with 70% allocated for training and 30% for validation. Using the aforementioned parameters and dataset, the model underwent training and validation. After 100 iterations, the model demonstrated high overall prediction accuracy on the validation set, achieving an overall accuracy of 96.52% across all interface types. Figure 3 illustrates the curve of loss function values and accuracy during model training.

3.4. Distribution Characteristics of Various Geological Interfaces and Their Influence on Mineralization

Geological interfaces within the study area are classified and mapped onto geological maps (Figure 4), with distribution characteristics analyzed for each interface. To clearly depict the spatial distribution of other geological interfaces within the study area, the boundaries of Quaternary sediments are excluded. The contact interfaces between limestone and siliceous rock, as well as between limestone and moderately acidic intrusive rock, are comprehensively classified as limestone interfaces. The interface between volcanic rocks and sedimentary rocks, and the eruption coverage area, are comprehensively classified as the volcanic rock interface.

To quantitatively analyze the spatial relationship between geological interfaces and deposits, this study establishes three-tiered distance boundaries centered on proven mineral deposits: 500 m, 1 km, and 2 km, forming influence domains of varying extents (

Table 2. Statistics of boundary types.

Major Categories	Subcategories	Quantity	Deposit
			500 m	1 km	2 km
Structural interfaces	fault	2259	15	8	8
Stratigraphic interfaces	conformable	7	0	0	0
	parallel unconformable	8	0	0	0
	angular unconformable	1253	0	0	0
Intrusive contact interfaces	/	1756	18	6	5
Special lithological interfaces	limestone and siliceous rocks	329	0	2	4
	limestone and moderately acidic intrusive rocks	24	1	0	2
	volcanic–sedimentary interface	46	0	1	2
	eruption coverage	252	1	6	2

). Building upon this, and considering the specific geological characteristics of the deposits, systematically analyze the influence intensity of various geological interfaces on mineralization. The intensity of influence is usually calculated as a function of the distance between geological elements. From the influence intensity, we can analyze the relationship between various geological interfaces and existing gold deposits, thereby revealing the geological layers and ore-controlling faults closely related to gold deposits. This paper uses an exponential function (Equation (8)) to describe the decay process of influence intensity as the distance from geological factors increases:

$$\mathrm{I}=\mathrm{a}{\mathrm{e}}^{-\mathrm{b}\mathrm{x}}$$

(8)

where $a$ represents the magnitude of influence (usually $a$ = 1); $b$ represents the attenuation coefficient, with higher values indicating faster decay rates; $x$ represents the distance from geological features; and 1 represents the intensity of influence, with higher values suggesting a closer potential relationship to mineralization.

The study area has undergone multiple episodes of tectonic movements, magmatic activity, and weathering erosion. This has resulted in a long stratigraphic time span, significant variations in thickness, frequent stratigraphic gaps, and complex rock assemblages. Stratigraphic contacts are predominantly unconformable or fault. Near Gouli town in the central part of the study area, the Halaguole Formation and the Haoteluowa Formation are exposed, exhibiting the parallel unconformity relationship. In the southwest, the Hongshuichuan Formation and Naocangjiangou Formation, both marine deposits, and there is a conformable contact between them. Quaternary deposits in the study area are relatively well-developed and widely distributed, covering approximately 20% of the study area. They are complex in origin and diverse in type, predominantly consisting of fluvial–alluvial, residual slope, and glacial sediments. Moreover, the angular unconformities in the study area primarily developed between the Quaternary sediments and the underlying other strata, exhibiting distinct spatial variations in their distribution. In the eastern part, they form lacustrine plains or intermontane basins unconformably overlying other strata, while in the west, Quaternary deposits are primarily distributed along river valleys [25]. According to statistics (Table 2), the association between the above three interfaces and mineralization is relatively weak, so these three interfaces are not included in the discussion of mineralization influence for the time being.

The study area has been influenced by multiple orogenic events, characterized by frequent and prolonged magmatic activity spanning from the Archean to the Mesozoic era, with Caledonian and Indosinian magmatism playing dominant roles. Intrusive rocks are widely distributed with extensive outcrops, occupying the vast majority of the study area and forming the majestic East Kunlun Magmatic Belt. The magmatic rocks within the area are diverse and complex, with ultramafic, mafic, intermediate, and acidic intrusive rocks all well-developed (Figure 5a). The spatial distribution of intrusive contacts and deposits within the study area exhibits distinct characteristics. The intrusive contact distribution map (Figure 5b) reveals that most deposits show close spatial association with intrusive bodies. Statistical analysis (Table 2) indicates that 18 deposits lie within 500 m, 6 within 1 km, and 5 within 2 km of intrusive contacts. The intrusive contact interfaces influence intensity map (Figure 5b) indicates that the most gold, silver, molybdenum, and cobalt deposits developed near zones of moderate mineralization influence intensity. Zones of high mineralization influence are primarily distributed in the Caledonian band of Indosinian intrusive rocks. Areas exhibiting strong influence intensity are mainly concentrated near Indosinian intrusive rocks.

Carbonate rock formations are primarily distributed in the southern part of the study area, exposing the carbonate sections of the Halaguole Formation, the Haoteluowa Formation, the Gequ Formation, and the Naocangjiangou Formation limestone sections. Despite intense tectonic activity and frequent magmatic processes within the study area, which have formed igneous rock assemblages of different ages and types, volcanic rocks are poorly developed and sparsely exposed. Only limited outcrops of volcanic rocks are present, including sections of the Lower Carboniferous Halaguole Formation, volcanic complexes of the Triassic Hongshuichuan Formation, and lava sections of the Elashan Formation. Based on the interface distribution map (Figure 6a), the locations of a few mineralized points show spatial association with intrusive bodies. Since Figure 6 involves multiple interfaces, duplicate counts of mineral occurrences exist. Detailed data is provided in Table 2. The lithological interface influence intensity map (Figure 6b) reveals that zones with higher mineralization influence intensity are primarily distributed at the interfaces between Carboniferous limestone and Triassic volcanic rocks. Areas with strong influence intensity are primarily distributed near the interfaces between Carboniferous limestone and siliceous rocks, as well as acidic intrusive rocks. Limestone and volcanic rock bodies often possess favorable mineralization potential. Combined with tectonic movements and hydrothermal activity, they frequently form large-scale deposits, making lithological interfaces crucial indicators for mineral exploration.

The study area exhibits highly complex tectonics with strongly developed fault structures, influenced by multiple episodes of ocean–continent, arc–continent, and continent–continent collision orogeny and superimposed tectonic activity during the Caledonian, Variscan, and Indosinian periods. Fault structures can be grouped into three sets based on their strike direction: NWW-EW, NW, and NE. The EW-trending fault structures are large-scale and extensively extended, while the NW-trending structures are predominantly secondary. Based on existing data, the EW-trending structure represents the primary ore-conducting and ore-controlling structure within the area, while the secondary NW-trending structure also functions as one of the major ore-controlling structures in the region. The distribution map of faults in the study area (Figure 7a) reveals significant spatial relations between the most discovered deposits and fault structures, with deposits predominantly concentrated along fault zones and their adjacent areas. Statistics (Table 2) indicate that there are 15 deposits within 500 m, 8 deposits within 1 km, and 8 deposits within 2 km. The fracture structure influence map (Figure 7b) indicates that areas with moderate mineralization influence are primarily distributed along fault zones, whereas regions with stronger mineralization influence are mainly concentrated at the central positions of fault zones and at intersections of multiple faults. The areas with high fault influence intensity are primarily developed along the NW-trending faults in the southern Gouli and at the intersection of NW- and NE-trending faults in the northern Gouli. The most gold, silver, and molybdenum deposits occur in intersections, at both ends of the fault, and along the fault. Copper deposits predominantly develop near NE-trending faults.

4. Intelligent Prediction Methods Based on Geological Interface Type Information

4.1. Principles of Intelligent Mineral Exploration Prediction

Intelligent prospecting prediction is to “learn” the geological, geophysical, and geochemical characteristics of known deposits through a CNN algorithm, and then use these characteristics to “compare” unknown areas to judge whether they have metallogenic potential. Multiple sources of information, such as geology, geophysical exploration, geochemical exploration, and remote sensing, reflect the material composition, formation, and evolution characteristics of spatial locations from different dimensions. Together, they constitute the multi-dimensional information basis for deposit prediction [38,39]. If a certain type of geospatial feature is present at a certain spatial location, it is possible that deposits may also be present in other areas with the same type of geospatial feature [40,41,42]. This data-driven model, compared with the traditional way of relying on the experience of geological experts, can more comprehensively and objectively mine metallogenic laws.

Intelligent mineral exploration prediction uses deep learning models to analyze these multi-source geological data. It automatically learns the correlations between known mineral deposits and their geological characteristics and constructs a model that can predict the location of deposits. This modern exploration approach integrates multi-source data with deep learning technology, which can effectively improve exploration efficiency and success rates and provide robust support for mineral resource prospecting.

4.2. Mineral Exploration Prediction Results

The mineral exploration prediction method based on the CNN2D model includes the following steps.

(1) Grid the geological and geochemical data of the study area, converting them into regular grid data to enable the integration of multi-source geological data.

(2) Mark windows containing gold deposits as positive samples. Select windows lacking mineralization conditions or randomly choose some windows to mark as negative samples. Use sliding window data augmentation to expand the training dataset and increase its scale.

(3) Input the generated training dataset into the CNN2D model. This model consists of an input layer, three convolutional layers, three pooling layers, one fully connected layer, and an output layer. The input layer receives regular grid data, and the convolutional layer captures the spatial patterns of the data to fully extract the features of different ore deposits. Each convolutional layer uses the ReLU activation function (Equation (9)) to introduce nonlinear relationships, enabling the neural network to better learn and represent complex patterns in the data. To reduce the computational load of the data without losing important features of the data, a pooling layer is added to downsample the feature maps obtained by the convolutional layer. The fully connected layer flattens the feature maps output by the model into one-dimensional vectors and then maps them to the final output layer. The output layer is the last layer of the CNN, responsible for converting high-level features into specific prediction results.

$$ReLU\left(x\right)=\mathrm{m}\mathrm{a}\mathrm{x}(0,x)$$

(9)

where $x$ is the value input to the activation function; $max(0, x)$ returns the greater value between $0$ and $x$.

(4) An amount of 80% of the sample data is used for model training, and 20% for model validation. The model is trained and verified with the prepared training dataset, and the optimal model is selected through the verification set.

(5) Forecast the working area by the sliding window method, delineate the favorable area for gold prospecting, and further analyze the reliability of the prediction results according to the mineral geological data of the working area.

Based on the above-mentioned intelligent mineralization prediction methods, this study employs the mineral exploration prediction model based on two-dimensional convolutional neural networks. Select the window size 24 × 24 grid units and set the ratio of positive to negative samples to 1:7. Input geochemical and geological interface data into the model to predict the gold element and delineate the favorable prediction zones. Finally, four favorable target areas for further exploration of gold deposits are selected (Figure 8).

The geological conditions of each prediction area are as follows:

Prediction Zone Au-01 is located in the eastern segment of the East Kunlun Mineralization Belt, within the East Kunlun Central Structural Belt unit. The zone primarily exposes the Jinshuikou group and Haoteluowa Formation. The structural orientation is NNW, consistent with the Kunlun Central Fault Zone’s trend. Despite significant Quaternary overburden, the presence of mineral deposits on both eastern and western sides, coupled with the discovery of multiple-phase concealed intrusive bodies during exploration, provides favorable conditions for gold–copper deposit formation.

Prediction Zone Au-02 is located in the eastern segment of the East Kunlun Mineralization Belt, within the transitional unit between the East Kunlun Central and East Kunlun North structural belts. A copper–molybdenum deposit exists nearby. The area features NW-, NNW-, and nearly EW-trending faults that intersect here. Stratigraphically dominated by the Jinshuikou Group, it exhibits intense magmatic activity.

Prediction Area Au-03 is located in the eastern segment of the East Kunlun Mineralization Belt, within the transition zone between the East Kunlun Central and East Kunlun North structural belts. This area is located about 10 km northeast of the Delong gold deposit, and NW-trending faults are developed in this area. The strata are dominated by the Kuhai group and Halaguole Formation, and the intrusive rocks are quartz monzonite of Early Permian.

Prediction Area Au-04 is located in the eastern segment of the East Kunlun Mineralization Belt, within the transitional unit between the East Kunlun Central and East Kunlun North structural belts, representing an active continental margin arc environment. The area primarily exposes the Jinshuikou Group, characterized by intense magmatic activity dominated by Middle–Late Triassic magmatism. Fault structures serve as key mineral-controlling features within the zone, representing favorable zones for Au enrichment and mineralization.

5. Discussion

5.1. Model Evaluation

To further validate the effectiveness of the CNN1D model, we compared it with a random forest model. Using the same data type and parameters as CNN1D, the random forest-based model achieved an accuracy of 93.11% (Figure 9). However, it failed to identify boundaries of smaller geological bodies and exhibited errors in recognizing specific lithological interfaces (Figure 10). In contrast, the CNN1D model achieved an accuracy of 96.52% (Figure 3), with boundary distributions consistent with actual geological distributions (Figure 4).

5.2. Improvement of Prediction Results After Incorporating Geological Interface Information

In order to compare the impact of adding geological interface data to the prediction results, we incorporated geochemical data and geochemical data combined with geological interface data into the mineral exploration prediction model.

Figure 11 displays the prediction results using only geochemical data. The model demonstrated excellent predictive performance with an accuracy rate of 93.49%. The prediction of mineral points in the study area can reach 89.5%, which can effectively identify potential mineralization areas and initially screen out areas worthy of attention for mineral exploration. However, the prediction zone (12.7% of the total study area) is large and scattered, making it difficult to form a systematic and large-scale exploration plan. This limitation somewhat restricted the predictive results’ practical value for guiding mineral exploration work.

Figure 12 shows the prediction results obtained by combining geological interface data with geochemical data. Compared with Figure 11, which uses geochemical data alone, this approach demonstrates significantly improved prediction performance, achieving an accuracy rate of 98.21% and covering 6.9% of the prediction area (

Table 3. Prediction results of different data.

Datasets	Accuracy	Predicted Area	Predicted Deposits
Geochemical data	93.49%	12.7%	89.5% (17/19)
Geochemical data + geological interface data	98.21%	6.9%	100% (19/19)

). The distribution of the prediction results is more refined and concentrated. There are no large prediction areas, and many small prediction areas in Figure 11 have also been screened out. Furthermore, the model precisely and completely covers known gold deposit locations in this prediction. After incorporating geological interface data, the model better constrains the expression of mineralization characteristics, yielding predictions that more closely align with actual geological patterns and mineral distribution features. Both the accuracy and practicality of predictions have been significantly enhanced, enabling more efficient screening of key areas for mineral exploration and facilitating subsequent targeted exploration efforts.

6. Conclusions

• This study classifies nine types of geological interfaces, and more than 5900 datasets of geological interfaces are constructed. Natural language processing techniques are applied to perform word embedding on the dataset. A geological interface type recognition model is developed using CNN1D, which identifies geological interfaces within the study area with an accuracy rate of 96.52%.
• In mapping the geological interfaces of the study area to the geological map according to classification, Gouli exhibits frequent tectonic movements and magmatic activity, with well-developed faults and widespread intrusive rocks. Volcanic rocks are poorly developed and rarely exposed. Carbonate rock formations are primarily distributed in the southern part of the study area. Combined with geological data from the mining area and conducting quantitative analysis of the distances between geological interfaces and mineral deposits, a strong spatial correlation between geological interfaces and mineral deposits has been identified.
• Comparing mineral exploration prediction maps reveals that incorporating identified geological interface data into the universal mineral exploration model achieves a prediction accuracy of 98.21%. The predicted area covers 6.9% of the region, exhibiting finer and more concentrated distribution. This provides scientifically grounded guidance with greater practical value for mineral resource exploration and development.