To commemorate the 10th anniversary of the founding of the Big Data and Artificial Intelligence for Geoscience Committee of the Chinese Society for Mineralogy, Petrology and Geochemistry, Earth Science Frontiers organizes this special issue to showcase the latest research advances in this field. The papers collected in this issue cover topics such as intelligent mineral exploration, 3D intelligent geological reconstruction, machine learning and deep learning, geological ontology and knowledge graphs, geological big data analysis, and mineral resource assessment and prediction. In intelligent mineral exploration, researchers proposed the GoldMiner-AI full-process intelligent system, established a “data and knowledge dual-driven” intelligent prospecting paradigm, and achieved a breakthrough in extending primary halo modeling from 2D to 3D. In 3D geological reconstruction, methods such as multi-discriminator attention generative adversarial networks, multi-modal large model-based structural geological map modeling, and adaptive finite difference method were developed. Machine learning and deep learning have made significant progress in areas such as seismic image super-resolution reconstruction, wildfire identification, and crustal thickness prediction. Research on geological ontology and knowledge graphs has constructed methods for mineral knowledge graph completion and a knowledge graph construction framework driven by tree-structured knowledge. Geological big data analysis revealed the spatiotemporal correlation between volcanic activity and dinosaur evolution, and established an analytical workflow for large fluid inclusion datasets. In mineral resource assessment and prediction, various intelligent prediction models based on AlphaEarth data, KAN networks, and ensemble learning have been developed. This special issue presents the latest achievements in the paradigm shift in Earth sciences driven by artificial intelligence and big data technologies.

To address critical bottlenecks in the intelligent transformation of geological prospecting—specifically, the lack of “full-process automation from data access to intelligent analysis” and “end-to-end systems covering data acquisition, integration, processing, anomaly identification, and intelligent prediction”—this paper systematically reviews the ongoing research efforts by the authors’ team in recent years to establish a new paradigm for prospecting using big data and artificial intelligence. We focus on the construction and application of a full-process intelligent system for prospecting tasks, named GoldMiner-AI. Built on the RuoYi-Cloud-Plus microservices architecture, this platform employs a multi-database system integrating PostGIS, Neo4j, Milvus, and MySQL to achieve unified management of diverse, multi-source geoscience data, including geology, geochemistry, geophysics, borehole data, field observations, and textual reports. For core intelligent modules, the system integrates the KAR-Graph anomaly identification framework and the MAF-Net multi-source feature fusion deep learning model. Combined with knowledge graphs and retrieval-augmented generation technology, a large language model tailored for the vertical domain of mineral exploration has been constructed, forming a complete intelligent workflow encompassing anomaly identification, target area delineation, knowledge reasoning, and intelligent Q&A. Validation results in mining districts such as the Youjiang Basin and the southern section of the Qinzhou-Hangzhou Metallogenic Belt demonstrate that: (1) the system can effectively identify the Au-As-Sb-Hg anomaly combination associated with Carlin-type gold deposits and deeply extract geochemical fingerprints related to deposit genesis; (2) through overlay analysis of multi-source layers, the system can accurately predict the spatial locations of lead-zinc mineralization zones; (3) the vertical domain large language model can significantly mitigate the “hallucination” phenomenon.

As mineral exploration increasingly targets deep-seated and concealed deposits, conventional prospecting methods and standalone machine learning models face critical limitations in generalization and geological interpretability. This study systematically delineates the evolution of the data-knowledge dual-driven intelligent paradigm and proposes a three-tier architecture: data-knowledge fusion, intelligent modeling-deconstruction, and application validation-feedback. Investigating technical pathways to open the “black box” of predictive models, this study identifies collaborative constraints—achieved through Knowledge Graph Embedding (KGE) and Graph Attention Mechanisms (GAT)—as the pivotal mechanism for deep data-knowledge synthesis. The operational logic involves a dual-constraint strategy: hard constraints derived from geological ontologies filter spatially irrelevant noise, while soft constraints via collaborative weighting enable models to adaptively capture high-probability metallogenic features. This framework establishes a closed-loop feedback system that integrates field empirical evidence with iterative model refinement. Results indicate that this dual-driven approach harmonizes expert metallogenic logic with machine computational power, significantly enhancing the interpretability and predictive precision of mineral potential mapping. This research provides a robust theoretical foundation for advancing intelligent decision-making and fostering transformative productivity growth in the global mining sector.

Primary geochemical halo exploration is a vital tool for deep mineral prospectivity assessment. However, traditional studies mostly rely on two-dimensional profile analysis, which cannot fully capture the three-dimensional structure and continuity of primary halos in complex ore-forming systems.To address this limitation, we developed a workflow for 3D primary halo modeling and proposed a compositional data analysis-based method for quantitatively extracting geochemical element associations. Using the Zaozigou gold deposit in Gansu Province as a case study, we built a 3D primary halo model. Structural analysis based on this 3D model further delineated potential targets for deep mineral exploration.The results reveal distinct zoning and continuity of geochemical element associations in 3D space. These associations exhibit gradual transitions, overlapping patterns, and spatial inheritance, reflecting the spatial evolution of the ore-forming system. Compared to traditional 2D methods, the 3D model presents the halo structure more fully, avoids complex zoning calculations, and reduces prediction uncertainty. This study extends primary geochemical halos from 2D to 3D, providing a quantitative approach focused on 3D geochemical structure for deep mineral prediction in geologically complex areas. It deepens and supplements existing exploration theory, offering guidance for deep prospecting in similar ore-forming systems.

Since the early 21st century, growing demand for mineral resources and increasing extraction complexity have made deep mineral exploration a major trend. Concurrently, processing and analyzing massive geological data has become a critical challenge for deep mineral exploration. The Huoqiu area in Anhui Province—a strategically important region rich in minerals such as iron and copper—has seen growing emphasis on deep exploration.In this study, we used the Huoqiu area as a test site and adopted a three-dimensional mineral prospectivity modeling method based on the second-generation Graph Attention Network (GATv2). By constructing a three-dimensional geological model of the deposit, we discretized the subsurface space into a voxel graph structure. Leveraging GATv2 for dynamic attention modeling of nonlinear spatial relationships between nodes, the method effectively captures complex topological relationships and nonlinear interactions among geological units, enabling high-precision prediction of deep mineral resources.Comparison with traditional machine learning models demonstrated the superiority of GATv2 in handling complex geological data and revealing mineralization patterns. Based on the model predictions, a significant mineralization target was delineated within the deposit area, providing a scientific basis for exploration. The results indicate that GATv2 exhibits exceptional performance in 3D mineral prospectivity modeling, improving both prediction accuracy and efficiency while offering novel ideas and methods for deep mineral exploration. This study provides valuable insights for mineral exploration in analogous regions.

As shallow mineral resources become progressively depleted, exploration targets have increasingly shifted to deeper, covered, and geologically complex terrains. This shift renders the efficient processing and intelligent utilization of multi-source, large-volume geoscientific data a critical bottleneck for breakthroughs in mineral prospectivity prediction. This paper systematically reviews key big data processing technologies for intelligent mineral deposit prediction and proposes a collaborative framework driven by the integration of geological knowledge and data. The framework operates through three actionable pathways: ① Constraining data processing with metallogenic theory—using genetic models as prior constraints in multi-source data integration, quality control, and anomaly extraction—to ensure geologically reasonable outcomes from the outset; ② Guiding feature construction with deposit mechanisms—translating geological knowledge (e.g., source-transport-accumulation factors, alteration zoning, ore-controlling structures) into computable quantitative features—to replace blind algorithmic searches with explicit geological semantics and thereby endow AI models with geological meaning; ③ Constraining training sample construction with geological knowledge—defining geologically credible positive and negative samples based on metallogenic system models, correcting spatial biases from exploration history, and ensuring the geological fidelity of mineralization patterns that predictive models learn from the training data. The framework’s effectiveness is demonstrated through two representative application scenarios: regional two-dimensional prospectivity mapping of skarn-type Fe deposits in southwestern Fujian Province and ore field-scale three-dimensional deep-level prediction in the Sanshandao goldfield, Jiaodong. These cases elucidate how the knowledge-data collaborative framework is implemented in practice and where its core advantages over purely data-driven approaches lie. The paper further adopts a “core challenges—targeted countermeasures—frontier trends” structure to discuss pathways for addressing data barriers, sample imbalance, model interpretability, and physical-geological consistency. The central argument is that breakthroughs in big data processing for intelligent mineral deposit prediction depend not on the increasing complexity of predictive algorithms, but on the depth and precision with which geological knowledge is integrated throughout the entire workflow.

Intelligent mineral prospectivity prediction is an important direction in mineral resource exploration. Deep learning methods can explore intrinsic correlations within massive geoscience datasets and capture complex nonlinear relationships between multi-source ore-indicating factors and target mineral deposits. However, deep learning models for this task often feature complex structures with numerous parameters, making them difficult to integrate effectively into existing mineral resource evaluation systems or exploration tools. Furthermore, when extracting local spatial features, these models typically overlook differences in the importance of individual ore-indicating factor channels, which limits further performance improvements. To address these issues, this paper proposes a mineral prospectivity prediction method based on a lightweight convolutional neural network (CNN) optimized by a channel attention mechanism. The method employs SqueezeNet to reduce model parameters and incorporates a Squeeze-Excitation (SE) module to introduce channel-wise attention. Based on each factor’s contribution to the prediction, the model dynamically assigns different weights to individual ore-indicating factor channels. This allows the network to focus on higher-weighted channels, retaining critical mineralization-related information and thereby enhancing prediction performance. The proposed method was applied to gold prospectivity prediction in Fengxian County, Shaanxi Province. Comparative experiments—involving a traditional CNN, a SqueezeNet without attention, and two attention-optimized structures—validate that the proposed lightweight CNN significantly reduces model parameters, and that the SE module effectively improves prediction accuracy. The optimized model accurately identifies 81.8% of known gold deposits within 29.5% of the study area, demonstrating reliable and efficient mineral prospectivity prediction.

Three-dimensional geological modeling faces key challenges: limited training samples, complex geological structures, and insufficient spatial consistency and topological preservation in generated results. To address these issues, this paper proposes a three-dimensional geological modeling method based on a Multi-Discriminator Attention Generative Adversarial Network (MDAGAN). Built upon the generative adversarial framework, we introduce a multi-directional discriminator constraint mechanism that enables the generator to receive discriminative feedback simultaneously from the X, Y, and Z directions during training, thereby enhancing directional consistency and overall structural coherence. We further incorporate a Hybrid Median Spatial-Channel Attention (HMSCA) module into the discriminator. By integrating spatial and channel attention, this module enhances the model’s ability to represent multi-scale structural features and complex spatial correlations, improving training stability and structural representation under limited data conditions. To validate the proposed method, experiments are conducted on a folded binary facies model and a highly random three-dimensional pore structure model. The generated geological volumes exhibit strong agreement with the reference models in macroscopic structural morphology, spatial correlation, and three-dimensional connectivity. Quantitative analyses using variogram and directional connectivity functions further demonstrate that MDAGAN effectively preserves the spatial statistical characteristics of different geological structures. Compared with baseline models, the proposed approach achieves better performance in preserving spatial correlation, enforcing structural consistency, and fitting statistical properties. It is capable of stably reconstructing complex three-dimensional geological structures under limited training samples, providing an effective solution for deep-learning-based modeling of complex geological structures.

3D modeling from geological maps is a key technology for advancing digital earth development. Traditional modeling methods based on geological maps often struggle to accurately reconstruct subsurface structures using only geometric information—such as geological boundaries and attitudes. These methods typically require manual determination of stratigraphic sequences or potential field delineations, making it difficult to automatically integrate the deep geological semantics embedded in the maps into the modeling process. To address this limitation, we propose a novel 3D geological modeling framework based on a large multimodal model (LMM). First, through case retrieval based on geological structure similarity and pre-drilling data injection, the LMM is rapidly optimized to intelligently extract the stratum sequence at target positions on the geological map. Second, using geometric data from geological boundaries and attitudes, formation thickness information is automatically calculated and virtual boreholes are constructed. Finally, spatial constraint point sets for surface and subsurface strata are extracted from the virtual boreholes and geological map; stratum attributes are then classified and predicted using a neural network, and a 3D model is constructed. Experiments show that, compared with mainstream modeling methods, the proposed approach achieves automatic and intelligent extraction of virtual boreholes from complex structural geological maps and constructs three-dimensional geological models with seamless surface-subsurface integration.

Three-dimensional (3D) ore body boundary modeling primarily relys on spatial interpolation methods, such as radial basis functions (RBFs). Because the scale of the linear system is highly correlated with the number of control points, these methods are generally limited to datasets on the order of ten thousand points. Moreover, such approaches typically rely on gradient or normal constraints to maintain the stability of geometric extrapolation. This paper proposes AdaFDI, a 3D implicit boundary modeling method for ore bodies based on adaptive finite differences. Without introducing any gradient information, the method constructs complex 3D ore body boundaries from large-scale, non-uniformly distributed control point datasets comprising hundreds of thousands of points. By introducing Octree-based Convolutional Neural Networks (O-CNN), we progressively convolve and fuse spatial distribution features of data points across octree levels, merging points with similar local geometric characteristics into the same finest-level cells. This enables optimal adaptive octree mesh partitioning that accounts for spatial similarity among control points while controlling the total mesh count. Based on this adaptive octree mesh, a finite difference scheme suitable for non-uniform mesh structures is constructed. We propose difference operators for hanging nodes and constrained nodes in octree meshes, along with an adaptive smoothing weight strategy based on local data density, to suppress numerical artifacts caused by sparse data. Numerical experiments demonstrate that the proposed method can achieve complex boundary construction for non-uniform control point datasets of nearly one million points. The method outperforms traditional RBFs and uniform-grid finite difference methods in model accuracy, computational efficiency, and memory usage, exhibiting robust performance across various data distribution patterns.

Mineral exploration in the deep and peripheral areas of mines is a key content to ensure the continuous supply of resources. Accurately determining the spatial structure of mineralization is conducive to achieving breakthroughs in deep and peripheral ore exploration. Exploration data contains rich information about the spatial structure of mineralization. However, existing studies are still difficult to dig the exploration data and characterize the spatial structure of mineralization from the perspective of ore-forming process. This paper takes the Canzhuang gold deposit in the Jiaodong Peninsula as the research object. Based on three-dimensional geological modeling and quantitative extraction of ore-controlling information, a Markov Random Field (MRF) model integrating mineralization distribution and ore-controlling information is constructed to identify the gold enrichment areas. Using the Hidden Markov Model (HMM), the migration trend of ore-forming fluids implicitly contained in the data of mineralization distribution and ore-related information is inferred, the correlation of the gold enriched parts is established from the perspective of ore-forming fluid migration, and the spatial structure of gold mineralization is determined. The results show that the morphology and volume strain characteristics of the Wang’ershan Fault significantly affect the mineralization of the Canzhuang gold deposit, and the mineralization is concentrated in specific undulation and strike dilation intervals. Multiple irregular enriched parts of mineralization in the Canzhuang mining area exhibit a significant spatially heterogeneous distribution, which is consistent with the distribution of high-grade voxels. The migration channels of ore-forming fluids generally plunge towards the south-west-west, closely relating to the structural dilation zone. Based on the above findings, this paper proposes a multi-scale ore-forming model for the Canzhuang gold deposit and determines five favorable areas for deep exploration, which can provide a reference for ore exploration in the deep and peripheral parts of the mine.

Seismic image super-resolution reconstruction is a key step for improving the accuracy of geological interpretation and the reliability of reservoir prediction in hydrocarbon exploration. Its goal is to reconstruct high-frequency information and enhance the representation of fine structures such as thin beds and faults without increasing acquisition costs. Due to limitations in field acquisition conditions, as well as attenuation/absorption in subsurface media and the superposition of random and coherent noise, practical seismic data often exhibit insufficient resolution, band-limited spectra, and blurred details. These degradations reduce reflector (in-phase event) continuity and obscure fault interfaces, thereby affecting downstream workflows such as structural interpretation, sequence stratigraphic delineation, and attribute inversion. In recent years, deep learning has provided new avenues to address this problem; however, existing methods still face challenges including single-scale feature extraction, lack of physical/spectral plausibility constraints, and limited generalization capability. In this paper, we propose a Multi-Scale Dual-Discriminator Generative Adversarial Network (MSDD-GAN). Specifically, we design a generator based on Multi-Scale Residual Groups (MSRG), where parallel multi-branch pathways and cross-layer residual connections enable joint modeling of multi-scale geological features. Moreover, a space-frequency collaborative dual-discriminator mechanism is constructed: the structural discriminator evaluates spatial realism in terms of reflector continuity, fault sharpness, and event-shape consistency, while the frequency discriminator constrains the plausibility of the seismic frequency distribution and suppresses unrealistic high-frequency artifacts. Extensive experiments are conducted on both synthetic and field seismic datasets, with the representative SeisGAN method used as the primary baseline. Performance is comprehensively evaluated using PSNR, SSIM, and spectral consistency metrics. The results demonstrate that MSDD-GAN improves the reconstruction quality of thin beds and faults, enhances reflector continuity, produces clearer details, and effectively suppresses noise. Furthermore, ablation studies (removing MSRG and removing the frequency discriminator) verify the critical roles of multi-scale feature modeling and frequency-domain constraints in improving seismic image super-resolution reconstruction fidelity and stability.

The intensification of global warming has increased the frequency and severity of wildfires. Traditional fire detection methods suffer from high false alarm rates, insufficient real-time performance, and lack of comprehensive detection capabilities. To achieve accurate and near-real-time fire identification, we proposed a Dynamic Multi-scale Hybrid Attention Network (DMHAN) and a multi-dimensional data augmentation scheme. To address wildfire sample scarcity and class imbalance, we first generated synthetic fire samples based on the heat conduction equation and radiative transfer models, providing high-quality data for training. Next, we constructed the dual-branch DMHAN model. This model integrates a Dynamic Multi-scale Spatial Attention Convolution (DMSAC) module for adaptive spatial feature extraction and a Gated Temporal Feature Fusion (GTFF) module for stage-adaptive temporal feature fusion. Finally, an adaptive spatio-temporal feature fusion strategy balances the contributions of spatial details and temporal dynamics. Using Himawari-8/9 satellite data, we evaluated the model on three typical wildfire events in Southwest China (Yuxi, Liangshan, Bijie). The DMHAN model achieved an average Fire Accuracy (FA) exceeding 90% and an Overall Accuracy (OA) over 97%. Its False Alarm Rate (FAR) and Overall False Rate (OFR) were significantly lower than those of the JAXA WLF L2 product and comparative models (LSTM, Transformer, MSSTF). The model demonstrated high robustness across different combustion stages and multi-scale fire scenarios. Ablation studies further verified the effectiveness of each core module. This work provides an efficient technical solution for comprehensive fire monitoring, balancing high accuracy with strong generalization.

Crustal thickness is a key parameter for understanding continental tectonic evolution. However, its quantitative prediction in marginal orogenic belts has long been hampered by data scarcity. Although artificial intelligence methods utilizing geochemical data have advanced, the fundamental influence of the inherent “closure effect” in compositional data on model performance is often overlooked. To address this, we developed a novel method integrating isometric log-ratio (ILR) transformation with a random forest algorithm to mitigate data bias and enhance prediction accuracy. The resulting ILR-RF model, trained on a global petrological-geochemical database, achieved excellent test accuracy and significantly outperformed mainstream comparative models. Applying the model, we reconstructed the Mesozoic crustal evolution history of the Lesser Xing’an-Zhangguangcai Range and eastern Jilin-Heilongjiang region in Northeast Asia. The reconstruction reveals a dynamic process of crustal thickening followed by thinning: thickness peaked at ~60 km in the Early Triassic (~250 Ma), thinned to ~45 km by the Late Triassic (~220 Ma), thickened again during the Middle Jurassic (~170 Ma), and finally stabilized around 43 km in the Early Cretaceous (~110 Ma). Notably, the significant Early Cretaceous thinning event correlates strongly with the peak period of large-scale gold mineralization in the region. This finding provides key constraints for understanding Mesozoic tectonic dynamics along the East Asian continental margin and offers a novel perspective for exploring the deep-seated drivers of major mineralization events.

The evaluation of mineral resource utilizability ranges from micro-level assessments of individual mines to macro-level evaluations of single mineral types nationwide. It considers multiple dimensions—geology, mining, processing, metallurgy, infrastructure, economy, market, law, environment, community, and policy—to scientifically assess the total amount of economically extractable resources from identified mineral deposits. The data were sourced from the national mineral resources reserves database, covering over 190 vanadium mining areas—including active, under construction, suspended, closed, and unused sites—and were randomly split into training, validation, and test sets in a 4∶1∶1 ratio. Based on this, we selected 35 indicators across five categories—geology, technology, economy, policy and law, and external conditions—as features, with “utilizability” as the target variable. Categorical features, such as deposit type, were preprocessed, and an intelligent evaluation model for vanadium resource utilizability was constructed. The results show that the CNN model effectively captures local correlations and complex nonlinear interactions among the indicators. Key indicators influencing vanadium resource utilizability in China include resource reserve scale, beneficiation difficulty, ore-bearing strata, power supply conditions, mining methods, ore composition, transportation distance, and exploration type. The assessment results indicate that currently unavailable resources are concentrated mainly in stone coal-type vanadium deposits within Cambrian black shale and in mining areas located in remote western regions with weak infrastructure. Compared with the traditional Delphi method, this approach offers significant advantages in indicator identification accuracy, evaluation efficiency, and model upgradability.

Active, transitional, and relict rock glaciers differ markedly in water content, paleoclimate information storage capacity, and hazard risk levels, making systematic classification essential for understanding climate evolution and assessing water resources in cold high-altitude regions. Currently, geomorphological analyses based on optical remote sensing can distinguish rock glacier types but lack quantitative indicators, whereas kinematic analyses derived from radar remote sensing enable quantitative classification but still produce a substantial number of undefined types. Integrating both optical and radar observations helps address these limitations by compensating for the insufficient quantitative capacity of optical data and the incomplete classification derived from radar data alone. Therefore, a representative sample dataset was established by jointly using Sentinel-2 imagery and LOS deformation rate data, and a high-accuracy three-class rock glacier model and a broadly applicable classification workflow were developed by combining a deep learning framework with ensemble learning strategies. The model achieved F1-scores higher than 0.90 on the training, testing, and validation sets, demonstrating strong generalization capability. Systematic classification carried out across five representative regions of the Tibetan Plateau further verified the practical applicability of the model. Feature importance analysis showed that the deep learning framework effectively fuses radar- and optical-derived information, leveraging their complementary strengths for rock glacier classification. This method provides a new technical pathway for high-precision rock glacier classification in the Tibetan Plateau and other high-mountain regions.

The completeness of a mineral knowledge graph critically determines the effectiveness of its downstream applications. To improve the integrity of such graphs， we propose a knowledge graph completion method based on an entity-aware pre-trained language model. We first collect and integrate the latest mineral data to expand the existing knowledge base， converting newly added entities into graph nodes. To infer missing relationships， we fine-tune the language model on mineral domain data to enhance its understanding and reasoning capabilities. Structured prompts are automatically generated via frequent pattern mining and combined with retrieval-augmented generation （RAG）， which introduces relevant contextual information to improve accuracy and reliability in knowledge triple prediction tasks. The proposed method was implemented in Python and evaluated against existing baseline models. Experimental results show that our approach outperforms the baselines on Hits@5 and Hits@10 metrics， while increasing the total number of relations in the completed mineral knowledge graph by 8.4%. These results verify the effectiveness of the proposed method and provide a feasible technical pathway and tool support for improving mineral knowledge systems.

The efficient integration and in-depth utilization of geological knowledge have long been hindered by challenges such as multi-source heterogeneity， inconsistent terminology， and fragmentation of textual and graphical information. To address these challenges， we propose a tree-structured knowledge-driven method for constructing geological knowledge graphs. The method first establishes a multimodal parsing framework to integrate semantic information from texts， images， and tables. Next， it introduces an explicit tree-structured knowledge framework as a domain prior， guiding large language models （LLMs） to perform hierarchical knowledge extraction following stratigraphic logic—for example， “section-stratum-lithology.” Furthermore， an intelligent entity alignment mechanism integrating edit distance and semantic vectors is applied； combined with literature metadata， this enables cross-document entity fusion. The proposed method was validated through a case study on the Cambrian-Ordovician strata in the Yichang area， Hubei Province. The results demonstrate that the method effectively parses complex geological semantics， achieves structured extraction of multi-granularity knowledge and precise alignment of cross-modal information， and constructs a traceable， dynamically updatable regional stratigraphic knowledge graph. This provides a systematic technical framework for the digital reconstruction and intelligent application of geological knowledge.

To address the challenges posed by the large volume， diversified formats， and complex structure of geological data， this study develops a domain-specific question-answering （QA） system based on the knowledge graph of tin deposits in Yunnan Province. The system integrates two core modules： semantic parsing and knowledge representation. A BERT-BiLSTM-CRF model is employed for geological entity recognition， while TextCNN is used for relation classification. For answer retrieval， the TransD knowledge representation model is adopted to encode entities and relations. A manually designed domain-oriented question template set， large-scale entity annotations， and a domain-specific negative sampling strategy are constructed to improve model performance. Experiment results show that the BERT-BiLSTM-CRF model achieves an F1-score of 0.898 for entity recognition， TextCNN reaches an accuracy of 0.932 in relation classification， and TransD demonstrates the best link prediction performance among the tested models. Based on these technologies， a Web-based geological QA system is implemented， supporting both knowledge graph browsing and natural-language querying. This study provides a technical reference for constructing a comprehensive QA system for other mineral types in the future.

Volcanic activity has long been regarded as one of the principal geological drivers of dinosaur evolution and extinction. Elucidating the spatiotemporal relationships between volcanism and dinosaur evolutionary dynamics is therefore essential for advancing our understanding of adaptive evolutionary strategies， dispersal patterns， and clustered extinction events in deep time. However， systematic investigations into the spatial coupling between volcanic activity and dinosaur evolution remain insufficient. In particular， despite the rapid expansion of geoscientific big data， there is still a lack of integrative studies that combine multi-source databases with quantitative spatial analytical approaches， which has constrained a comprehensive assessment of their co-evolutionary dynamics. Within a data-driven analytical framework， this study integrates dinosaur fossil records from the Paleobiology Database （PBDB） and the Deepbone Database （DBD）， together with volcanic rock data from the Geochemistry of Rocks of the Oceans and Continents （GEOROC）. By combining paleogeographic reconstructions with kernel density estimation and K-D tree nearest-neighbor analysis， we systematically characterize the spatiotemporal associations between dinosaur evolution and volcanic activity from the Late Triassic to the Late Cretaceous. The results indicate that under the overall warm climatic background of the Mesozoic， dinosaur distributions exhibit a pronounced latitudinal gradient， with a diversity peak concentrated in mid-latitude regions. Major eruptive events associated with the Central Atlantic Magmatic Province and the Karoo-Ferrar Large Igneous Province exerted significant influences on dinosaur evolutionary trajectories， displaying strong temporal and spatial correlations. Spatial distance analyses further reveal that the distribution of distances between dinosaur fossil occurrences and volcanic rocks follows an overall decreasing trend and exhibits a bimodal structure， with a primary peak at 0-200 km and a secondary peak at approximately 450 km. This spatial pattern likely reflects the combined effects of short-term environmental disturbances induced by volcanic eruptions， long-term ecological benefits associated with volcanic material input， and paleogeographic configurations shaped by tectonic processes.

Fluid inclusion research provides vital information on the temperature， salinity， density， pressure， and composition of ore-forming fluids， serving as an indispensable tool for revealing mineralization processes and metallogenic mechanisms. Previously， such studies predominantly focused on a limited number of typical deposits with small-scale datasets and relatively simple analytical methods， which restricted the extraction and interpretation of deep-level information. The Nanling metallogenic belt， characterized by the widespread development of Mesozoic quartz-vein-type tungsten deposits， is an ideal region for investigating the large-scale evolution of ore-forming fluids. This paper systematically constructs a large dataset of primary fluid inclusions for quartz-vein-type tungsten deposits in the Nanling region. Employing a data-driven approach， we conduct in-depth data mining of the fluid inclusion information. The Wasserstein distance is utilized to quantify the similarity of microthermometric data among different ore clusters， which is then combined with Number-Size （N-S） fractal analysis and Gaussian Mixture Model （GMM） clustering to delineate the fluid evolution stages. The results indicate that the quartz-vein-type tungsten deposits in the Nanling region exhibit multi-stage mineralization characteristics. Fluid inclusions within the ore-bearing quartz veins are predominantly liquid-rich two-phase inclusions， with H₂O dominating the liquid phase. The homogenization temperatures range from 109.6 to 397.1 ℃， and the salinity varies from 0.18% to 20.75% NaCl_eq （mass fraction）. Based on the Wasserstein distance， the deposits can be classified into four categories， illustrating four distinct evolutionary pathways of the ore-forming fluids. Furthermore， the variations in the Bayesian Information Criterion （BIC） for both the N-S fractal model and GMM clustering collectively demonstrate that hydrothermal activities in each ore cluster are not continuous and gradual， but rather characterized by significant multi-stage features. The fluid evolution processes at varying scales display pronounced self-similarity within local ore clusters， confirming that the intrinsic physicochemical mechanisms governing deep fluid evolution are highly consistent on a macro-regional scale. Through mathematical geological modeling and big data analysis， this study overcomes the limitations of previous single-deposit research， attempting to provide a quantitative perspective for evaluating the similarity of ore-forming hydrothermal fluids and identifying multi-stage fluid characteristics.

Chloride （Cl^-） is a key indicator of soil salinization in coastal wetlands and plays a crucial role in identifying ecological thresholds for understanding water-salt environment evolution. Focusing on the Liaohe Estuary wetland， this study integrates principal component analysis （PCA）， K-means clustering， and the interquartile range （IQR） method to systematically analyze Cl^- covariation patterns and identify its ecological thresholds， based on 340 samples and 28 soil physicochemical parameters. PCA results indicate that Cl^- exhibits stable spatial covariation with water-salt factors such as soil moisture content， and its spatial variation is primarily controlled by local moisture conditions， reflecting the influence of evaporation-leaching processes on salt redistribution. K-means clustering divides the samples into three groups with significantly different water-salt characteristics， corresponding to varying degrees of salinization. Using the IQR rule， reasonable variation ranges and potential outliers are identified， and key ecological thresholds are determined. This data-driven approach avoids reliance on predefined empirical thresholds or subjective classifications. The results show that the low-to-moderate risk threshold for Cl^- is 28.93 mg/kg， the moderate-to-high risk threshold is 144.66 mg/kg， and localized anomalous enrichment reaching approximately 405.06 mg/kg occurs in high-salinity samples. When combined with vegetation distribution characteristics， Cl^- concentrations in low-salinity vegetation sites are generally below 28.93 mg/kg， whereas halophytic vegetation and reed sites under high salinity stress mostly exceed 144.66 mg/kg， indicating strong consistency between the identified thresholds and observed ecological responses. This study establishes an ecological threshold identification framework driven by artificial intelligence and big data analysis， providing a scientific basis for the intelligent quantification of key geochemical indicators and for ecological management in complex environments.

Titanium—hailed as the “Third Metal” for its unique physicochemical properties—finds extensive applications in aerospace， marine engineering， and biomedicine. In China， weathering-crust ilmenite placers are a crucial titanium resource type. The Funing area in southeastern Yunnan holds immense potential for such deposits， yet exploration progress remains slow. Remote sensing geology—an intelligent prospecting method increasingly applied in recent years—offers significant advantages for exploring supergene weathering-crust deposits. However， previous remote sensing studies have often faced challenges in effectively fusing multi-source， heterogeneous data and precisely analyzing ore-controlling conditions. To address these challenges， we selected the Funing ilmenite placer district as the study area. Utilizing AlphaEarth data—released globally by Google in 2025 with multi-source data fusion， 10-meter resolution， and 64 dimensions—we analyzed ore-controlling conditions by integrating multi-dimensional data， including geology， geochemistry， and geomorphology. Consequently， we constructed a remote sensing intelligent prospecting framework based on machine learning modeling. Comparative experiments were conducted using Random Forest （RF）， Support Vector Machine （SVM）， and Multilayer Perceptron （MLP） models to optimize metallogenic prediction for ilmenite placers. The results indicate that AlphaEarth data significantly enhanced the identification of mineralization anomalies and clarified their geological implications. Among the models tested， MLP demonstrated superior performance in prediction accuracy and stability， exhibiting exceptional capability in identifying mineralization anomalies within complex geological backgrounds. The metallogenic favorability map generated by MLP showed high consistency with known deposits. Validation through success rate curves and field verification confirmed high prediction accuracy （AUC=0.94）. This study provides a scientific basis for ilmenite placer exploration in the Funing area and offers a technical framework for intelligent prediction of similar deposits in analogous regions.

Machine learning is widely applied to mineral prospectivity mapping （MPM）； however， mainstream supervised methods are often constrained by limited positive samples， subjective negative sample selection， and the “black-box” nature of predictions. To address these issues， this study introduces the Kolmogorov-Arnold Network （KAN）—a model characterized by inherent interpretability and high parameter efficiency—into MPM for the first time. Using the Zhaishang-Mawu gold district in Gansu Province as a test case， we constructed a 30-meter resolution multidimensional dataset integrating regional geochemistry and multi-source remote sensing data， encompassing morphostructural， hydrothermal alteration， and geochemical anomaly features. A pseudo-label iterative expansion mechanism was designed to alleviate overfitting risks associated with small-sample learning. The complex metallogenic setting and challenges in targeting concealed mineralization in this area provide an ideal validation scenario. Results show that the KAN model achieved an AUC of 0.82 in the training area and correctly identified all known deposits within 60-meter buffer zones in an independent blind test area， significantly outperforming a Random Forest benchmark. Interpretability analysis revealed the relative contributions of geochemical anomalies （44.2%）， remote sensing alteration information （38.6%）， and morphostructural features （17.2%）， thereby collectively mapping the complete mineralization process. This study demonstrates that KAN not only enhances prediction accuracy and model interpretability but also provides a technically viable pathway combining high accuracy with process transparency for mineral exploration in areas with weak exploration signals and similarly complex geological settings.

Mineral prospectivity mapping is a critical component of regional mineral exploration； however， under complex conditions characterized by limited samples and heterogeneous geological data， the prediction stability and generalization ability of traditional approaches and single models remain insufficient. To address this issue， this study focuses on the western Chagai mineral belt in Pakistan and integrates hyperspectral remote sensing， geological， geophysical， and geochemical data to develop a mineral prospectivity mapping framework based on an ensemble-learning-driven adaptive feature-spatial fusion network. The proposed method employs three-dimensional convolutions to jointly extract pixel-level geological attributes and spatial neighborhood information， incorporates an adaptive attention mechanism to enable collaborative feature learning， and leverages ensemble learning to mitigate the effects of sample selection bias， class imbalance， and noise. The results show that the generated prospectivity maps exhibit a high spatial consistency with known deposits in both the training area and independent test areas， while field validation further identifies mineralization and alteration features related to porphyry copper systems within multiple predicted high-potential zones， demonstrating strong exploration significance. Feature importance analysis reveals that Cu grade contributes most strongly and shows a significant positive correlation with mineralization probability， whereas Ag exhibits a negative correlation， consistent with elemental migration and zoning patterns in porphyry copper mineralization. At the spectral level， high weights are assigned to shortwave infrared diagnostic bands associated with phyllic， advanced argillic， and propylitic alteration， reflecting the spatial differentiation of alteration zones controlled by magmatic-hydrothermal processes. These results indicate that the proposed approach achieves robust predictive performance and provides a reliable and transferable framework for intelligent mineral prospectivity mapping in complex geological settings.

Mineral resources constitute a critical material foundation for socioeconomic development. Their assessment and prospectivity modeling provide a scientific basis for sustainable resource management. Traditional evaluation methods are often costly， time-consuming， and constrained by limited data-processing capabilities. Conversely， computational approaches frequently rely on fixed patterns and cannot adequately incorporate expert knowledge. Moreover， the utilization of multi-source heterogeneous data， particularly unstructured text data， remains low. To address these limitations， this study formalizes expert knowledge and analytical methods into prior rules and proposes a novel framework： the Mineral Generative Prior Transformation Model （MineralGPT）. The MineralGPT framework is driven by the rule-based description， storage， and analysis of prior knowledge， supporting diverse tasks such as data processing， metallogenic information extraction， prospectivity modeling， and content optimization. A case study optimizing gold-polymetallic exploration targets in the Xiaoshan-Xiongershan area demonstrates the model's efficacy. Experiments on a term-weighting-based target optimization model within MineralGPT show that results， even when supported by limited data， achieve high consistency with expert evaluations. Compared to large language models （LLMs） like ChatGPT， which require massive data and computing power， MineralGPT offers advantages including lower cost， fewer constraints， and high customizability.By integrating rule-based prior knowledge， MineralGPT provides a novel methodology for mineral resource assessment and prediction. Furthermore， it offers a new perspective for developing next-generation artificial intelligence that synergistically combines rule-based and learning-based paradigms.

Hot dry rock (HDR) represents one of the primary forms of geothermal energy storage and is classified as a green, low-carbon, and sustainable energy source. With vast development potential, HDR occupies a strategically significant position in global energy transitions. However, extreme conditions in deep, high-temperature, high-pressure, and high-hardness environments pose multidisciplinary technical challenges for HDR exploration, drilling, and reservoir stimulation. Over the past half-century, global HDR development has progressed through incremental advancements, with the U.S. Project Red achieving a revolutionary breakthrough via “horizontal wells+staged fracturing” technology, marking the onset of commercial exploitation. In China, the main distribution patterns of HDR resources have been preliminarily clarified, accompanied by the establishment of a “source-reservoir-caprock” enrichment theory and a stepwise evaluation framework integrating “strategic area selection-zone prioritization-target assessment.” Innovations include the development of specialized polycrystalline diamond compact (PDC) drill bits, drilling fluid active cooling technologies, high-temperature-resistant rotary steering systems, and work-fluid formulations to enhance rock-breaking efficiency and wellbore thermal management. A novel “intermittent injection+cyclic variable displacement” fracturing technique has been introduced to construct efficient thermal exchange fracture networks in reservoirs. Field experiments have been conducted in the Qinghai Gonghe Basin, Hainan Fushan Depression, and Hebei Matouying areas. Future efforts will prioritize advancements in reservoir-formation theory and intelligent, efficient development technologies to establish a “geothermal+multi-energy synergy” development model. A national-level pilot-scale platform will be established to foster emerging disciplines in deep-earth science and engineering, thereby accelerating large-scale commercial HDR exploitation and providing strategic support for China’s energy transition and carbon neutrality goals.

Three-dimensional (3D) geological modeling is crucial for urban underground space development but faces challenges of data sparsity, limited automation, and interpolation methods that often ignore stratigraphic influence, compromising accuracy. This paper presents an automated workflow for 3D geological modeling using multi-source data. To address data scarcity, diverse geological profiles, seismic sections, and contour data were integrated to generate virtual boreholes, which supplemented the actual borehole dataset. Contour data were used to constrain the virtual borehole locations. Subsequently, to enhance modeling efficiency, a hierarchical approach was developed and implemented for the explicit, automated construction of the 3D geological structural model. Finally, to improve attribute modeling accuracy, this structural model served as the framework for generating various attribute models (e.g., single-layer, multi-layer). Crucially, stratigraphic constraints were incorporated into standard interpolation methods(Inverse Distance Weighting and Kriging) to increase their geological relevance and precision. Thus, this study addresses the key challenges of data sparsity and low interpolation accuracy by presenting an automated, strata-constrained modeling workflow, providing a practical solution for urban underground space development in Shanghai.

The Jurassic formation in the Mosuowan uplift, a significant deep-buried oil-bearing series in the central Junggar Basin, has undergone a complex hydrocarbon accumulation process due to multi-phase tectonic activities. To clarify the coupling mechanism between diagenetic evolution and hydrocarbon accumulation in the Jurassic sandstone reservoirs, this study employs an integrated, multi-scale approach, combining detailed reservoir petrography, in-situ laser ablation calcite U-Pb dating, and fluid inclusion analysis. Integrated with burial and thermal history modeling, we quantitatively characterize the dynamic coupling relationship between reservoir diagenesis and hydrocarbon charging. The research establishes a hydrocarbon accumulation model of “early charging and late adjustment” for the Jurassic reservoirs in the Mosuowan Uplift, identifying three distinct charging episodes: The first episode occurred during the Early Cretaceous (132.5±1.3 Ma), involving the charging of low-maturity crude oil, as evidenced by yellow fluorescent oil inclusions within calcite cement. The U-Pb dating results show a strong consistency with the timing (133 Ma) projected from the minimum homogenization temperatures of coeval aqueous inclusions. The second episode, during the late Oligocene (approximately 35 Ma), involved the adjustment and charging of mature oil, indicated by blue fluorescent oil inclusions in quartz microfractures. The third episode consisted of natural gas charging during the Neogene. However, large-scale gas invasion into the Jurassic reservoir was prevented by the sealing effect of an overpressure barrier. This study not only advances the understanding of deep hydrocarbon accumulation mechanisms in the Junggar Basin but also provides an important theoretical basis and practical guidance for exploring and developing hydrocarbon reservoirs in similar complex tectonic settings.

There are numerous BIF iron deposits located within the greenstone belts of the Luxi Terrane, including the Huibaoling, Xiaoyanzhuang, and Gouxi iron deposits. The Cangyi iron ore belt is situated within the greenstone belt. Previous studies have indicated that the greenstone belt and the BIF iron deposits were formed concurrently, with mineralization, magmatism and sedimentation being interrelated processes. The greenstone belt formed during two stages of magmatism: 2.7-2.6 and 2.5 Ga. However, research on the timing of formation and genesis of iron deposits remains relatively poorly understood. In this study, we present detailed geochronological and geochemical data of amphibolite, amphibole schist as well as leptynite, which are the host rocks of the Cangyi iron ore belt. The study of this research is to determine the diagenetic age and magmatic evolution processes, as well as to explore the tectonic background and metallogenic environment. Apatite U-Pb dating indicates that the protolith age of the amphibolite is 2522 Ma, the metamorphic age of the greenschist-amphibolite facies is 2411 Ma, and the metamorphic age of the amphibolite-granulite facies is 1874 Ma. Amphibolite and amphibole schist exhibit high contents of Al₂O₃ and FeO, while having low contents of Na₂O, K₂O, and TiO₂. In addition, there is a positive anomaly in Eu, and light rare earth elements show variable degree of enrichment relative to heavy rare earth elements, suggesting that their protolith are basalts. In contrast, leptynite displays high contents of SiO₂ and Al₂O₃, low content of FeO, negligible Eu anomaly, and are rich in light rare earth elements while being poor in heavy rare earth elements. This composition suggests that their protolith are dacites. Amphibolite and amphibole schist exhibit negative anomalies of Ce, Nb and Ti, which are akin to those found in the source region of depleted mantle magma. Nb, La and Nd show a positive correlation with Zr, while V demonstrates a negative correlation with Zr. This suggests that fractional crystallization occurs during magma evolution. High Nb/Th and Th/Yb ratios indicate the incorporation of crustal materials during this process. The Nb/Ta ratio of leptynite ranges from 9.7 to 12.6, which is similar to that of crustal source. Amphibolite, amphibole schist and leptynite display characteristics of arc magmatism, suggesting that they formed in a subduction-collision process at a convergent margin.

The Hongge layered mafic-ultramafic intrusion in the Panxi region hosts world-class Fe-Ti-V oxide deposits, yet its genesis remains debated. To address this debate, we conducted an in-situ Sr isotopic and trace element study of clinopyroxenes to investigate the magma emplacement, evolution, and ore-forming processes. The intrusion is divided into lower, middle, and upper zones based on mineral assemblages and textures. Petrographic observations indicate that clinopyroxene is an early-crystallizing phase, making its Sr isotopic composition a reliable tracer for the parental magma. Clinopyroxenes from all three zones show a narrow range of initial ⁸⁷Sr/⁸⁶Sr ratios: 0.7053-0.7062 (lower zone), 0.7055-0.7064 (middle zone), and 0.7055-0.7064 (upper zone). Compared to the whole-rock samples from the lower zone (0.7057-0.7076), the clinopyroxenes have lower values, indicating that the parental magma was enriched in radiogenic Sr prior to emplacement, with the elevated whole-rock ratios likely resulting from later crustal fluid contamination. The lack of correlation between Cr and Ni in clinopyroxenes and whole-rock V in the lower zone suggests that early Fe-Ti oxide crystallization was controlled by parental magma composition rather than clinopyroxene fractionation. In contrast, positive correlations in the middle and upper zones indicate that Fe-Ti oxides crystallized copiously alongside clinopyroxene. The main ore-bearing lower and middle zones show minimal variation and high average Sc content in clinopyroxenes (59.2 ppm), significantly higher than that in the ore-poor upper zone (38.7 ppm). This suggests frequent replenishment by Fe-Ti-rich magma in the former, but weak replenishment in the latter. Furthermore, clinopyroxene Zr content implies the Hongge magma chamber was part of a conduit system. Thus, the frequent replenishment of Fe-Ti-rich magma was the key factor in forming the thick ore bodies of the middle zone.

The crystallization fractionation during the magmatic evolution stage of alkaline granite-hosted rare metal-rare earth elements (REE) deposits is closely related to the enrichment of REE and rare metals. The transitional relationship between primary zirconium-containing minerals and hydrothermal zircon reveals the evolutionary process of REE from magmatic to hydrothermal systems. Zircon-quartz pseudomorphs are the main form of occurrence of hydrothermal zircon, wherein the pseudomorphic assemblage completely replaces the primary mineral while preserving its original crystal morphology. The development of zircon-quartz pseudomorphs serves as a key mineral index for zirconium (Zr) and REE mineralization in alkaline granite-hosted rare metal-REE deposits. In-depth research on the characteristics and genesis of zircon-quartz pseudomorphs is crucial for understanding the physicochemical conditions and metallogenic mechanisms of rare metal mineralization in alkaline granite-hosted deposits. However, the limitations of traditional identification methods in pseudomorph boundary annotation and mineral phase delineation, such as manual micro-petrographic observation, have greatly restricted the analysis of zircon-quartz pseudomorphs and the development of quantitative mineral analysis. To address these issues, we apply the YOLOv8n-seg Deep Learning (DL) model for instance segmentation to accurately identify and quantify the microscopic features of zircon-quartz pseudomorphs. The results show that the YOLOv8 instance segmentation algorithm achieves more than 90% accuracy in the automatic recognition of zircon-quartz pseudomorphs and effectively extracts the morphological features of the pseudomorph. Additionally, feature maps are introduced throughout the instance segmentation process, highlighting and quantifying the critical characteristic regions for pseudomorph identification through feature contribution. These regions, including the textural structure, mineral phase distribution, and crystal boundaries, which are closely related to fluid migration and alteration, exhibit distinct identification features. Our DL-based instance segmentation model provides an accurate and efficient method for pseudomorph identification and quantitation, demonstrating tremendous potential for understanding metallogenic mechanisms for rare metals mineralization in alkaline granite-hosted deposits.