We have entered the era of big data and artificial intelligence. Big data or big data analytics is a new thinking for solving geoscientific problems, emphasizing the multidimensional associations among variables and allowing data to speak for themselves, thereby leading to new insights and more informative answers. Artificial intelligence is a new data mining approach with a strong non-linear modeling ability, which can deeply mine data and discover hidden patterns. The big data and artificial intelligence-driven mineral prospectivity mapping has become the high ground of global mining technology competition, reshaping the paradigm of mineral exploration. This study proposes the basic concepts and main components of big data and artificial intelligence-driven mineral prospectivity mapping, analyzes the scientific connotations, the state-of-the-art, and key scientific and technological issues in intelligent cognition, intelligent learning, and intelligent decision-making. These three key parts are essential components of intelligent mineral prospectivity mapping, and link between the Earth system and mineral system, mineral system and exploration system, and exploration system and evaluation system, respectively. In the future, big data and artificial intelligence-driven mineral prospectivity mapping should focus on the construction of prospecting big data, new geological constrained algorithms for mineral prospectivity mapping, high-performance computing in image processing, and the cultivation of innovative interdisciplinary talent.

Mineral resources are vital for national economic security and industrial development. As shallow resources become increasingly depleted, the exploration of alternative resources in the deeper parts of mines has become an inevitable option to ensure resource security. However, deep mineral prospectivity mapping faces significant challenges, including great depths, limited direct observations, and weak indirect information. There is an urgent need to overcome key technical challenges, including unclear deep ore deposit structures, obscured deep ore-controlling patterns, and significant difficulties in spatial positioning of deep ore bodies, whereas it is very difficult for traditional quantitative prediction methods for mineral resources to meet the demand for precise 3D spatial positioning of deep resources. To address these issues, this paper proposes novel theories and methods of 3D intelligent prediction of deep mineral deposits. Guided by the metallogenic system theories and data science, these theories and methods have preliminarily broken through two key scientific issues: "geological-geophysical-geochemical constraints on the 3D reconstruction of deep ore deposit structures" and "the controlling mechanism of deep 3D ore deposit structures on the spatial positioning of mineralization". It has established a methodological framework of "geological analysis - refined modeling-3D analysis-intelligent prediction", and innovatively developed a theoretical, methodological, and technical system centered on the 3D refined reconstruction of deep deposit structures, 3D geometric-material analysis of ore-forming space, and intelligent 3D positioning prediction of deep ore bodies. The core technologies include: (1) refined 3D reconstruction of deep deposit structures based on multi-source heterogeneous data assimilation and Bayesian inference; (2) intelligent extraction of 3D spatial geometric and material mineralization information using coupled simulation of multi-level structural styles and metallogenic processes; (3) intelligent 3D positioning prediction of deep ore bodies applying artificial intelligence techniques such as deep neural networks, domain adaptation, and multi-modal learning. The automation of refined deep structure reconstruction, the quantification of deep ore-controlling patterns representation, and the intellectualization of orebody positioning prediction have been realized, and significant breakthroughs have been achieved in deep ore prospecting in major mineral concentration areas in China, such as the Jiaodong Peninsula and the Jinchuan. Finally, this paper discusses the future challenges and development directions of 3D intelligent prediction of deep mineral resources from the perspectives of multi-source data assimilation for refined 3D modeling of deep structures, characterization of mineralization information based on the coupling of spatial structure and metallogenic materials, and large language model-driven 3D positioning prediction of deep ore bodies, aiming to further promote the development of in-depth intellectualization of deep mineral prospectivity mapping.

Numerical modeling provides a key approach for quantitatively analyzing the ore-forming processes, revealing ore location mechanisms, and facilitating mineral prospectivity studies for magmatic-hydrothermal deposits. In recent years, with the rapid advancement of computational geosciences, significant progress has been made in numerical modeling of ore-forming processes, which provides critical support for metallogenic prediction in multiple aspects. We summarize the fundamental theories and methods of numerical modeling, provide a comprehensive review of current research regarding advances in simulating ore-forming processes, analyzing ore location mechanisms, and facilitating metallogenic prediction. Finally, we conclude with an outlook on the future development of numerical modeling in advancing metallogenic prediction. We propose that future research should focus on advancing coupled mechanical-thermal-chemical-fluid processes modeling, developing efficient numerical methods, and promoting the intelligent integration of multi-source data. These efforts will collectively drive the evolution of mineral prospectivity modeling toward a new paradigm characterized by mechanism-data synergistic modeling.

Mineral prospectivity mapping (MPM) driven by big data and artificial intelligence (AI) represents a cutting-edge approach in mineral exploration. However, exisiting methods are typically constrained by the inability to achieve stable cross-regional applications due to the limited generalization ability, poor transferability, and insufficient interpretability. Foundation models, based on the "pretrain and fine-tune" paradigm, have demonstrated excellent cross-task transfer and strong generalization ability in the fields such as natural language processing and computer vision, offering a promising path to overcome the aforementioned bottlenecks. The development of foundation models for MPM has significant potential to revolutionize traditional models and improve exploration efficiency, representing a new research direction for intelligent MPM. This study systematically reviews the development and construction processes of foundation models, focusing on the state-of-the-art technical characteristics of large language models, visual foundation models, and multimodal foundation models. This study also summarizes the limitations of existing foundation models for MPM, and explores the construction process of MPM foundation models from a perspective of language-based, visual-based and multimodal-based foundation models, and discusses the challenges of developing MPM foundation models, providing a reference for development of MPM foundation models.

With the advancement of big data and artificial intelligence technologies, mineralization prediction is undergoing a series of technological innovations, transitioning from data-sparse to data-intensive approaches. This shift is expected to become a new "technology engine" for breakthroughs in mineral exploration and discovery. Despite the substantial accumulation of heterogeneous multisource geological, geophysical, geochemical, and remote sensing data, as well as rich geological reports and literature, it remains a critical research challenge to effectively integrate and deeply mine these valuable data resources to further optimize mineralization prediction indicator systems and construct high-quality mineralization prediction datasets. To address the challenge, this paper proposes integrating multilayered and multidimensional mineralization prediction knowledge across the Earth system, metallogenic system, exploration system, and prediction-evaluation system through large models and knowledge graph technologies. A multilayered, multi-system-coupled mineralization prediction knowledge graph will be constructed, and intelligent construction of mineralization prediction indicator systems will be achieved through knowledge graph mining. Based on big data and artificial intelligence technologies, a multilayered method system for intelligent construction of mineralization prediction data will be developed. This system will focus on intelligent mining of geological exploration data, automated scientific data extraction and spatiotemporal analysis, and intelligent data inversion and simulation. Such advancements are expected to strengthen the deep coupling between prediction data and indicators, enhancing the accuracy and reliability of prediction results, and providing more robust technical support for mineral exploration and discovery breakthroughs.

Mineral exploration is a fundamental task for safeguarding national resource security and the stability of industrial supply chain. As a core step of mineral exploration, mineral prospectivity mapping (MPM) has undergone transformative development, driven by big data and artificial intelligence, emerging as a prominent research field within Earth science and accumulating a substantial volume of literature. In this study, we employ bibliometric methods to analyze and discuss the research status, developmental trajectory, and hotspot evolution of MPM over the past five decades, based on a dataset of 935 relevant publications from three flagship journals of the International Association for Mathematical Geosciences, spanning from 1969 to 2025. Bibliometric statistics on authors, institutions, and countries reveal that Carranza E.J.M. and Zuo Renguang are the most highly productive and highly cited scholars in the field as listed as author and first/corresponding author, respectively. China is the largest contributor of publications in this field, and the China University of Geosciences (Wuhan) ranks first in global institutions in both publication volume and total citation. The analysis of collaboration networks indicates a strong regional orientation, lacking a high-level and regular international cooperative research network. The evolution of MPM, based on the hotspot analysis of keywords, is divided into three distinct stages, namely the foundation stage (1969-1990), the expansion stage (1991-2010), and the boom stage (2011-2025). The thematic focus and developmental trajectory of each stage are determined by the prevailing technologies and algorithms of the era. The foundation stage, focusing on mineral resource assessment, was dominated by geostatistics (variogram and Kriging). The rise and widespread application of GIS technology during the expansion stage facilitated the shift of MPM into the mainstream scientific task. In the boom stage, the prevalence of machine learning algorithms led to the dominance of intelligent MPM in the thematic tasks. Recent research hotspots and trends indicate a shift from relying solely on high-performance predictive models towards in-depth exploration and optimization of the internal mechanisms of intelligent models. The focus is on leveraging cutting-edge AI technologies to address inherent challenges such as the black-box nature of decision processes and sample scarcity. Although advanced deep learning algorithms have gained significant traction, classic shallow learning algorithms, such as support vector machine, which exhibits great performance in processing high-dimensional data and nonlinear problems, and random forest characterized by its strong resistance to overfitting, remain popular choices among scholars in this field during the boom stage due to their high suitability for few-shot MPM tasks. By leveraging quantitative statistical and visualization tools, this study provides a macro and comprehensive perspective for understanding the development of MPM, and offers critical insights into future research directions of intelligent prediction in MPM.

Three-dimensional reconstruction of fluid pathways is critical to understanding ore genesis and guiding exploration since hydrothermal fluid migration pathways exert a fundamental control on the transport, concentration, and precipitation of ore-forming fluids. However, robust three-dimensional reconstruction of fluid pathways at the deposit scale remains challenging due to complex structural overprinting, sparse sampling, and limited quantitative tools. Here, we present an integrated knowledge and data-driven framework to reconstruct the 3D hydrothermal fluid migration pathways of the Xiadian gold deposit in the Jiaodong Peninsula. Geological indicators related to fluid flow were extracted from drill-hole and mine-level datasets and incorporated into a spatial probability model using a Graph Convolutional Network (GCN). A Markov chain model was subsequently applied to quantitatively trace three-dimensional migration trajectories. The GCN demonstrates strong predictive performance under small-sample conditions (AUC=0.956 9), delineating high-probability fluid pathways that are consistent with established metallogenic models. The reconstructed pathways indicate that ore-forming fluids originated at depth and migrated upward along the Zhaoping Fault Zone, exhibiting a branching and diffusive architecture. Major fluid conduits are primarily controlled by deep-seated structural variations of the main fault, whereas dense terminal branch networks are dominated by secondary faults and fracture systems, reflecting a synergistic structural control on fluid migration and mineral precipitation. The results confirm the existence of a conceptual model of "transport along major conduits and precipitation within terminal branches" in Xiadian gold deposit, providing new insights into the coupling between tectonics, fluid flow, and mineralization. On this basis, two prospective targets for deep exploration within the Xiadian deposit are identified.

Three-dimensional mineral prospectivity modeling serves as a crucial technical approach in the exploration of deep concealed mineral resources. In recent years, deep learning methods represented by convolutional neural networks have achieved some progress in integrating 3D predictive information; however, constrained by the local receptive fields of CNNs, it remains difficult to extract long-range dependencies and global correlations between 3D predictive factors and mineralization occurrences, which limits the prediction accuracy for deep concealed ore bodies. To address these issues, this study develops a 3D-ViT model based on the Vision Transformer (ViT) architecture, tailored for 3D geological data. The model employs a 3D voxel-patch embedding module and decoupled 3D positional encoding to explicitly preserve the structural information of geological bodies. By leveraging a multi-head self-attention mechanism, a global perceptual field is constructed to model cross-scale spatial relationships between multiple predictive factors-such as intrusions, strata, and structures- and mineralization evidence. In a case study of the Shizishan ore field in Anhui Province, the model successfully predicted the main known ore bodies, achieving an AUC of 0.96. It demonstrated strong predictive capability and precision with accuracy, recall, and F1-score above those of 3D-CNN and traditional machine learning models. Based on the prediction results, four prospective target areas were delineated in the deep part of the Shizishan ore field, verifying the method's effectiveness and reliability in detecting concealed mineralization under complex geological settings. This study not only extends the application of ViT to three-dimensional geoscientific data but also provides a novel method with global perception for intelligent prediction of deep mineral resources, holding significant potential for practical exploration applications.

This study proposes a three-dimensional mineral prospectivity modeling method that integrates causal inference with Graph Attention Networks (GAT) to improve the accuracy and efficiency of deep concealed skarn-type copper deposits prediction in complex geological settings. Using the Anqing area of the Middle-Lower Yangtze Metallogenic Belt as a case, a high-precision 3D geological model involving strata, intrusions, faults, and ore bodies was constructed based on geological maps, borehole data, and geophysical information through a hybrid explicit-implicit modeling approach. On this basis, the RESIT causal inference algorithm, which is built upon non-Gaussian assumptions, was employed to analyze 62 ore-controlling factors. A causal graph was established, and 14 key controlling variables were identified. Subsequently, a 3D prediction dataset incorporating spatial adjacency relationships was developed, and the causal structure was introduced into the GAT model for mineralization probability prediction. Comparative experiments demonstrate that the proposed method outperforms commonly used approaches-including Random Forest, Support Vector Machine, Graph Convolutional Networks, and 3D Convolutional Neural Networks-in terms of accuracy, AUC, and success rate curves. Based on the predictions, four deep high-potential target zones were delineated, which are closely associated with diorite intrusions and Triassic carbonate contact zones. The results indicate that integrating causal inference with deep graph learning not only enhances prediction performance but also improves the geological interpretability of the model, providing a promising technical pathway for deep mineral exploration.

To address the challenges of deep mineral body prediction, this study proposes a three-dimensional mineral prospectivity prediction method that integrates mineralization process numerical simulation and machine learning, using the Maodeng copper-tin deposit in Inner Mongolia as a case study. The Flac^3D is used for numerical simulation of the mineralization process to obtain key physical field parameters that control mineralization, such as stress, temperature, and fluid pressure. These physical results, combined with geological data, are then used in the XGBoost machine learning model for three-dimensional quantitative mineral prospectivity prediction. It is demonstrated that the method successfully simulated the stress field, temperature field, and fluid migration process of the mining area. The AUC value of the XGBoost model reached 99.26%, showing excellent predictive ability. SHAP analysis reveals that shear stress, pore pressure, and temperature are the main factors affecting the distribution of mineral bodies. The predicted results highly correlate with the known mineral bodies, providing a reliable basis for mineral body prediction and ultimately identifying two mineral prospecting target areas. The study demonstrates that the combination of mineralization process numerical simulation and machine learning prediction methods can effectively improve the accuracy of deep mineral resource predictions, offering new technical insights into mineral resources assessment in similar regions.

The construction of fine-scale 3D models of deep structures remains challenging due to the lack of direct exploration data and the high uncertainty associated with geophysical prospecting inferred data. To address these issues, utilizing prior knowledge to mitigate the limitations of scarce exploration data and uncertain geophysical data is a valid idea. In this work, a 3D refined modelling method for deep fault named Deep Learning Aided Kalman Filter (DLAKF) was proposed. Based on the concept of Kalman filtering, the 3D modelling of the deep fault from shallow to deep was regarded as a "temporal sequence prediction" problem involving both system disturbances and observation errors: (1) A state equation for the Kalman filter was constructed to predict deep fault positions. The prior knowledge constraints of shallow fault locations and occurrence were integrated in the equation. (2) A deep spatial attention convolutional network, embedded with prior knowledge constraints, was designed. The observation equation for Kalman filter was construct based on the outputs of the deep neural network. By calculating the Kalman Gain, the positions predicted by the state equation and the observation equation were dynamically fused and the optimal estimation of deep fault location was achieved. The proposed method was applied to construct a 3D detailed model of deep fault at the Xiadian gold deposit. DLAKF successfully constructed the detailed model down to 3000 meters. The average horizontal error between the constructed model and drilling holes was 6.17 meters. The accuracy was improved by 91% to 93% compared to existing implicit modelling methods, which means the detailed model constructed by DLAKF reflected the detailed geometry of the deep fault structures more accurately. Based on the reconstructed 3D deep fault model, four prospective mineralization areas were identified in the deep sections of the Xiadian deposit, providing valuable guidance for deep resource exploration.

Enhancing the transparency of deep geological structures at the ore-field scale is critical for subsurface mineral exploration and prospectivity modeling, and three-dimensional (3D) lithological modeling serves as a critical technology for this objective. However, existing ore-field-scale modeling workflows rely on explicit modeling approaches with relatively low efficiency, which can hardly meet the demands of multi-stage mineral exploration and real-time mining. Consequently, high-precision and high-efficiency implicit 3D lithological modeling methods are urgently needed. To address this issue, a Hybrid Attentional Mechanism deep learning model (HAM) is constructed on the basis of the 3D Convolutional Neural Network (3D CNN), integrating the Convolutional Block Attention Module (CBAM) and the Self-Attention Module (SAM). Based on this algorithm, deep representations within multi-source geological and geophysical data are mined to determine the boundaries of geological bodies required for modeling, thereby achieving a 3D lithological implicit modeling method capable of capturing both local details and long-range dependencies. To validate the effectiveness of the proposed hybrid attentional mechanism model, the Jiaojia gold field in the Jiaodong Peninsula was selected as the study area, and comparative and ablation experiments were conducted. Relative to baseline models‒Random Forest (RF) and a vanilla 3D-CNN, HAM markedly improves the macro-averaged accuracy, precision, recall, macro-averaged F1 score and confusion matrix of ore-field-scale implicit 3D lithological modeling, with direct implications for subsurface mineral exploration and mining operations.

Porphyry deposits are widely distributed in the world, and their mineralization mechanisms are of significant scientific research value. Numerical simulation is an important approach to quantitatively and continuously analyze the formation of ore deposits, revealing the temporal-spatial distribution of mineralization/alteration and migration-evolution of ore-forming fluids. This study established a simple geometric model of porphyry deposits, and conducted multi-field coupled (heat-transfer, fluid-flow, chemical reaction, diffusion) numerical simulation of its formation. Results show that using simple geometric models in metallogenic simulation research is feasible, which has certain enlightening significance for searching for deep prospecting targets and explaining the genesis of large and ultra-large porphyry deposits. This method can not only be used to calculate the spatial distribution of mineralization and achieve deep prospecting prediction, but also to infer the tectonic environment during the mineralization period through different forms of mineralization distribution, thereby enabling a deeper study of issues such as the ancient mineralization environment. In addition, simple models are characterized by low computational cost, minimal human influence, and high credibility, so they can play an important role in the study of some specific metallogenic theoretical issues.

Current applications of Large Language Models (LLMs) in geological prospecting face challenges including insufficient domain expertise, data privacy concerns, and model hallucinations. Furthermore, there remains a lack of efficient and rapid knowledge recommendation methods for LLMs in this field. This study proposes a KG-RAG (Knowledge Graph-Embedded Retrieval-Augmented Generation) framework that automates the extraction and structured representation of geological prospecting knowledge under the constraints of a geological ontology, leveraging large LLMs as tools. It further employs multi-hop retrieval algorithms within the knowledge graph to enhance the depth and breadth of retrieved content, thereby constructing an intelligent question-answering model for geological prospecting. Experimental results demonstrate that KG-RAG achieves scores of 0.807 (Precision), 0.833 (Recall), and 0.819 (F1-score) in knowledge graph construction tasks. Compared to direct knowledge extraction using the baseline LLM (GLM4-9B), KG-RAG delivers improvements of approximately 50% (Precision), 8% (Recall), and 29% (F1-score), respectively. In question-answering tasks, KG-RAG achieves 0.917 (Recall) and 0.88 (Precision), outperforming document vector-embedded retrieval-augmented generation methods by approximately 24% (Recall) and 22% (Precision), respectively. KG-RAG exhibits superior performance in both knowledge graph construction and intelligent question-answering. It effectively collects and represents geological prospecting and mineral exploration knowledge, providing a valuable reference to geologists for the combined application of LLMs and knowledge graphs.

The new generation of artificial intelligence technologies, represented by Large Language Models (LLMs), provides new opportunities for the structured representation and intelligent reasoning of geological knowledge. To address the challenges posed by the complexity of geoscientific knowledge systems, as well as the semantic fragmentation, limited reusability, and poor visualizability of unstructured texts, this study proposes a unified strategy for constructing a knowledge graph that integrates deposit genesis and prospecting indicators, taking rare metal deposits in South China as a study object. Based on the DeepSeek R1-32B large language model and prompt engineering, a knowledge graph covering key rare metal elements such as Li, Be, Nb, and Ta, is automatically extracted and constructed. The knowledge graph construction and its extensibility analysis indicate that rare metal mineralization in South China is closely associated with Indosinian and Yanshanian magmatic activities, characterized by significant high-degree fractionation and magmatic-hydrothermal processes. Rare metal elements exhibit a combinatorial anomaly of Li-Be-Nb-Ta-W-Sn. It is concluded that the knowledge graph constructed using LLMs reveals the multi-stage metallogenic mechanisms of rare metals in South China, clarifies the intrinsic relationships among geochemical anomalies, structural controls, and alteration zoning of rare metal deposits, and provides an intelligent research framework for the exploration of rare metals in South China and adjacent regions.

To address the challenges in effectively utilizing massive unstructured data within geological exploration and the issues of hallucination and lack of specialized logic in general Large Language Models (LLMs), we propose an intelligent knowledge mining framework for vertical domains by integrating Knowledge Graph (KG) and Retrieval-Augmented Generation (RAG). This framework is validated through a comparison case study of Carlin-type gold deposits in the Southwest Guizhou, China, and in Nevada, USA. Firstly, a RAG-based intelligent question-answering system was constructed using a locally deployed DeepSeek-32B model. Through vector retrieval and generative reading comprehension, the system achieved precise traceability of professional knowledge and highly reliable Question & Answer (Q&A). Secondly, leveraging Supervised Fine-Tuning (SFT) techniques on the LLM, we developed a cross-regional metallogenic knowledge graph systematically covering stratigraphy, structure, alteration minerals, and ore-controlling factors based on hundreds of multi-sources, and heterogeneous geological documents. The results demonstrate that the proposed system significantly outperforms GPT-4o in terms of objective accuracy. For subjective content generation, it exhibits high faithfulness, full traceability and effectively mitigate the hallucination. Analyses based on graph topology not only quantitatively reveal the macroscopic similarities and differences of Au mineralization between the two regions but also quantify the cascading indicative pathways-from orebody entities and alteration assemblages to geochemical element anomalies, confirming the system's capability to discover implicit clues for mineral exploration. This study realizes the intelligent transformation and in-depth mining of knowledge from unstructured text to structured representations, offering a novel technical pathway to address the dilemma of "data-rich yet knowledge-poor" prevalent in the geoscience domain.

To address the challenges faced by general-purpose large language models in mineral exploration, including scarcity of domain corpora, insufficient coverage of domain terminology and register adaptation, and pronounced factual hallucinations. We constructed a mineral-exploration corpus of approximately 25 million tokens and, on this basis, proposed a curriculum-based continual pre-training strategy, which organizes training data into three stages: terminology, mechanisms, and cases. Coupled with gradual unfreezing of Transformer blocks and learning-rate scheduling, we conducted continual pre-training of Qwen3-1.7B to achieve stage-wise domain adaptation, resulting in a mineral-exploration-oriented LLM, Geo-MineLLM. During inference, we integrated a Hybrid RAG framework, leveraging hybrid retrieval and evidence-constrained generation to enhance factual consistency. Human evaluation indicates that Geo-MineLLM substantially improves domain question-answering performance relative to the base model and larger-parameter models within the same family. With Hybrid RAG enabled, overall domain QA performance approaches that of GPT-4.1. The proposed training-inference integrated framework provides a lightweight pathway for building mineral-exploration LLMs and enabling reliable domain-specific question answering.

This study aims to deepen the understanding of metallogenic regularities and evaluate the prospecting potential of gold deposits in the Jiangnan Orogen. Focusing on the gold deposits within and adjacent to the Jiangnan Orogen, technologies related to the knowledge graph were introduced. A domain knowledge schema was developed using a top-down approach, and the metallogeny-exploration knowledge graph of gold deposits was constructed by integrating deep learning and Large Language Models (LLM). Community detection and Jaccard similarity evaluation were used to analyze the clustering characteristics of the gold deposits. The knowledge schema contains 28 geological entity types and 10 semantic relationship types. The resulting knowledge graph encompasses 60 representative gold deposits in the region, containing 2 212 geological entities and 5 497 semantic relationships. Community detection successfully extracted key ore-controlling factor combinations and metallogenic regularities, such as "alteration-mineral-strata". Jaccard similarity analysis indicates that the Shuikoushan and Huangjindong gold deposits have high similarities to global large-to-giant deposits, revealing significant prospecting potential in their deep-seated zones and peripheral areas.

While advances in satellite hyperspectral technology and machine learning have significantly boosted its application in mineral prospectivity modeling, conventional data-driven approaches often fall short by neglecting the essential geological information that controls mineralization processes. To bridge the gap, this study develops a novel methodology that integrates hyperspectral imagery with critical geological determinants-specifically pluton boundaries and fault systems, establishing a 39-channel comprehensive data set featuring hyperspectral geological information, and introduces an enhanced Graph Convolutional Network (GCN) model. Architectural improvements include the incorporation of residual connections and the systematic application of batch normalization across both residual modules and convolutional layers, substantially stabilizing and accelerating the training process. Validation using ZY-1 02D hyperspectral data from the Dahongliutan area demonstrates that our refined GCN model achieves superior accuracy in identifying mineralized granitic pegmatites. Quantitative evaluations confirm substantial performance gains, with accuracy improvements of 7, 22, and 27 percentage points over the baseline GCN, Convolutional Neural Network, and Support Vector Machine models, respectively. This work establishes an effective and automated framework for high-precision prediction of lithium- and beryllium-mineralized granitic pegmatites via hyperspectral remote sensing.

To achieve satellite-based hyperspectral mapping of lithium spatial distribution to support mineral exploration efforts, a lithium content inversion model was developed using a multi-algorithm approach, based on spectral data, XRD analysis, lithium content measurements, and hyperspectral imagery from the ZY1-02D site. The quantitative inversion model was established using the Ensemble Transformer Neural Network (ETNN) regression algorithm. This model was integrated with Spatial Spectral Endmember Extraction-Spectral Angle Mapper (SSEE-SAM) for mineralized outcrop identification and supplemented with the CORrelation Alignment (CORAL) algorithm for spectral domain correction. The quantitative inversion model was then applied to generate a spatial distribution map of lithium content. Training set R²=0.93, RPD=3.91, RMSE=110.13; validation set R²=0.89, RPD=3.08, RMSE=183.04, indicate high accuracy and strong fitting capability; field test set R²=0.75, RPD=2.02, RMSE=263.86, demonstrate robust generalization ability. Correlation coefficients indicate that lithium exhibits the strongest association with montmorillonite and chlorite. Importance analysis reveals the 2 132-2 350 nm wavelength band as critical for the inversion model. This study establishes a comprehensive inversion methodology linking spectral data to lithium content, providing technical support for exploration of clay-type lithium deposits in the Tongchuan region, Central Yunnan basin, and Central Guizhou basin.

Establishing a detection model that can take into account the multi-element geochemical spatial-spectral characteristics and effectively fit the complex distribution of data is the key to identification of abnormal areas. In response to the challenge of extracting geochemical prospecting anomalies in the high-altitude, deep-cutting, and shallow-coverage areas of the Eastern Kunlun Mountains in Xinjiang, this study proposes a Spatial-Spectral Feature and Global Spatial Correlation Network (SSGSNet). Based on ResNet residual blocks, the spatial-spectral feature branch is integrates a dual-attention module to extract local spatial-spectral features, with the spatial correlation branch using patch embedding and self-attention mechanisms to mine global spatial correlation features. Incorporating tectonic data improves the accuracy of geochemical prospecting, and SHAP values explain the critical role of faults within the model. Experimental results show that the AUC value of the SSGSNet model reaches 0.945 3, significantly outperforming the ResNet and ViT single models as well as the conventional spatial-spectral dual-branch model. Field verification shows that gold mineralization phenomena of varying degrees were found in four high-anomaly areas, including Yaoxi and Bashiganike, which confirms that the model can effectively solve the problem of extracting complex background geochemical anomaly information, providing reliable technical support and target area guidance for mineral exploration in covered areas.

The newly identified cryptic explosive breccia-type fluorite deposits in the western Guizhou fluorite ore concentration area possess significant prospecting potential. However, the brecciated textures, hydrothermal alteration, and other characteristics of this type of fluorite deposit are easily confused with those of other hydrothermal breccia-type deposits or intensely structurally altered vein-type deposits. Therefore, accurately distinguishing between cryptic explosive breccia-type fluorite deposits and basin brine-related hydrothermal filling-type fluorite deposits in the study area is one of the key scientific challenges for achieving breakthroughs in fluorite prospecting in Guizhou Province. This paper conducts a comparative study of Support Vector Machine (SVM) and Random Forest machine learning classification models using systematically collected rare earth element (REE) data from three genetic types of fluorite deposits: cryptic explosive breccia-type, magmatic hydrothermal-related filling-type, and basin brine-related hydrothermal filling-type, which is combined with comprehensive analysis, including statistical analysis based on Principal Component Analysis (PCA), dimensionality reduction visualization, and quantitative evaluation using an REE separation scoring system. The results indicate that the discriminant model constructed by SVM exhibits significantly higher accuracy and stability compared to Random Forest, enabling more effective discrimination among these three genetic types of fluorite deposits. Furthermore, it identifies a refined candidate pool of key elements that can be used to distinguish them. Newly constructed discriminant diagrams (Tb/Dy vs Sm/Yb, δCe vs Sm/Yb, δCe vs Sm/Tm, δEu vs Sm/Lu) have been developed, which effectively differentiate among cryptic explosive breccia-type, magmatic hydrothermal-related hydrothermal filling-type, and basin brine-related hydrothermal filling-type fluorite deposits.

Complex geological structure modeling is of significant importance in fields such as resources exploration, underground engineering design, and geological hazard prediction. Generative Adversarial Networks (GANs) have demonstrated strong nonlinear modeling capabilities and pattern transfer abilities in geological modeling. However, when dealing with complex geological constraints and the reconstruction of fine structures, they still face challenges in modeling accuracy, structural connectivity, and modeling efficiency. To address these issues, this paper proposes a GAN-based geological modeling method incorporating multi-scale feature fusion and deep separable convolutions. A multi-scale feature fusion module enhances the expression of geological structure details and overall consistency, while deep separable convolutions reduce model parameters and computational costs, improving modeling efficiency. Additionally, a conditional feature adaptive fusion and progressive resolution generation strategy enhances the model's sensitivity to conditional data. To validate the method's effectiveness, typical models including two-dimensional river phases, multi-attribute ice wedges, and three-dimensional fold structures were selected. Systematic evaluations were conducted across spatial variability, connectivity, attribute consistency, and conditional point reconstruction accuracy. Comparative analyses were performed against multi-point statistical methods (e.g., QS) and an improved generative adversarial network (e.g., CWGAN-GP). The results show that at resolutions of 64×64 and 64×64×64, the MS-SWD indicators of the generated models for the two-dimensional and three-dimensional datasets are 0.016, 0.025, 0.007 9, and 0.008 7 respectively, which are significantly lower than those of the comparison methods. At the same time, the average connected region size of the generated models is closest to that of the reference model (300.59 pixels for the two-dimensional river data and 17 814.17 pixels for the three-dimensional fold data). In terms of overall accuracy, the accuracy rate and MSE indicators of the proposed method are superior to those of the comparison method (73.24%, 69.48% and 0.024, 0.047 respectively), and the advantages in efficiency and parameter quantity are proved through efficiency analysis and ablation experiments. The experiments show that the proposed method is suitable for efficient modeling tasks of complex non-stationary geological bodies since it significantly improves the modeling efficiency while ensuring reasonable and high fidelity, endowed with broad engineering application prospects.

To address the mode collapse and artifacts encountered in deep generative models under hard data constraints, this study proposes a Stratified Autoregressive Generation (SAG) method, aiming to develop a robust reservoir characterization approach. The method utilizes an offline-trained Transformer architecture as a conditional distribution estimator to replace the computationally expensive online "search and count" process of MPS. A three-level, coarse-to-fine strategy is adopted to define global structures at large scales first and subsequently propagates constraints to finer scales, thereby avoiding the quadratic computational complexity on large grids. Multiple sets of experiments as well as multidimensional scaling and variogram analyses indicate that the generated realizations possess diversity and accurately reproduce the global statistics and spatial continuity of the training data. Quantitative assessment using histogram intersection further confirms high local pattern fidelity without artifacts. Uncertainty assessment reveals that uncertainty increases outward from hard data points, showing a convergence pattern consistent with geological laws. The results indicate that the proposed method maintains spatial continuity and realization diversity under varying amounts of hard data constraints, which achieves the accurate characterization of complex reservoir structures and properties.

The global exploration and development of deep geothermal resources is at a critical stage, transitioning from experiment to application. Artificial intelligence, particularly deep learning, has demonstrated transformative potential in big data analysis, pattern recognition, and nonlinear problem solution, offering new pathways to address challenges hindering efficient and precise exploration of deep geothermal resources. It is significantly important to promote the integration of deep learning with traditional geothermal exploration processes to enhance China's competitiveness in the development and utilization of deep geothermal resources. This paper focuses on the integration of deep learning data processing, modeling, and prediction with deep geothermal resource exploration (including geothermal geological surveys, geophysical exploration, and geochemical exploration, etc.). It systematically reviews and summarizes key technological methods in geothermal resource exploration, deep learning techniques, and the critical advancements and research outcomes that empower deep geothermal exploration. This study demonstrates the improvement of efficiency, accuracy, and precision brought by deep-learning-based geothermal resources exploration methods compared to traditional methods. Finally, the paper discusses the core challenges with cutting-edge technologies faced by deep geothermal exploration. In future, intelligent deep geothermal resources exploration urgently needs to focus on multiple modal data fusion, interpretable and trustworthy artificial intelligence, and construction of intelligent computing foundations and large models, which ultimately, will enable a leap from "experience-driven" to "knowledge-driven" and then to "intelligent-driven", providing core technological support for the digital and intelligent development of the geothermal energy resources industry.