Orogenic magmatic Cu-Ni sulfide deposits represent an important source of copper, nickel, and other critical metals, and their formation is closely related to mantle-derived magma evolution and crust–mantle interaction. The Beishan Orogenic Belt, located in a key tectonic position within the Central Asian metallogenic domain, hosts numerous mafic–ultramafic intrusions formed during the post-collisional extensional stage, many of which are associated with varying degrees of Cu-Ni mineralization. However, uncertainties remain in distinguishing ore-bearing from barren intrusions and in evaluating their mineralization potential. In this study, whole-rock geochemical data of mafic-ultramafic intrusions from the Beishan Orogenic Belt were systematically compiled from published literature to construct a dataset comprising 360 samples. A total of 27 geochemical variables were selected as input features, including major elements (e.g., SiO₂, Al₂O₃, Fe₂O₃, MgO), rare earth elements (La-Lu), and trace elements (e.g., Rb, Nb, Sr, Y). Four machine learning modelswere employed to classify mineralized and non-mineralized intrusions and to evaluate their mineralization potential, such as multilayer perceptron (MLP), support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost). The results indicate that all models exhibit strong classification performance, with the MLP model achieving the highest accuracy of 97%, followed by XGBoost (94%), RF (91%), and SVM (90%). The modeling results demonstrate that whole-rock geochemical characteristics can effectively capture key metallogenic processes such as magmatic differentiation and crustal contamination, serving as reliable indicators for identifying ore-bearing intrusions. The discrimination method established based on the optimal model enables effective evaluation of the mineralization potential of mafic-ultramafic intrusions in the Beishan Orogenic Belt. This study not only provides a new technical approach for Cu-Ni sulfide exploration in the Beishan region but also offers methodological insights for assessing the mineralization potential of similar orogenic magmatic deposits.

Pyrite is the most common gold bearing mineral in gold deposits, and its geochemical composition can effectively indicate mineralization processes. This study systematically compiles trace element data of pyrite from epithermal, porphyry, and orogenic gold deposits worldwide. Nine commonly used geochemical features are selected, and three machine learning models including the K nearest neighbors algorithm, random forest, and support vector machine are employed for model training, with the aim of establishing an efficient and accurate method for identifying gold deposit types based on pyrite composition. The evaluation results of model classification performance show that the classification accuracies of the trained K nearest neighbors, random forest, and support vector machine models are 0.89, 0.96, and 0.95, respectively. In addition, pyrite compositional data from the East Kunlun orogenic gold deposits that are not involved in model training are used for independent validation to assess the applicability and accuracy of the models in practical scenarios. The results indicate that all three models correctly identify these pyrites as mainly derived from orogenic gold deposits, among which the support vector machine model achieves the best performance with an accuracy of 0.88. To address the black box issue in machine learning, an interpretable machine learning method is further introduced to quantify the contributions of different pyrite trace element features to the classification results. The SHAP analysis reveals that As, Co, and Cu are the three most influential trace element features for distinguishing porphyry and orogenic gold deposits, whereas Cu, As, and Ni exert the greatest influence on the identification of epithermal gold deposits. The results demonstrate that trace elements in pyrite can serve as effective indicators for discriminating genetic types of gold deposits. In the future, combined with machine learning approaches, pyrite composition has the potential to be further extended to the identification and classification of other types of mineral deposits, providing new technical approaches for studies on ore genesis and mineral exploration prediction.

Pyrite is the most common metallic sulfide in gold deposits, and its trace element composition effectively records the physicochemical properties of ore-forming fluids and mineralization environments, thus serving as an important indicator for discriminating genetic types of gold deposits. With the rapid development of information technology, geoscience research has entered a data-intensive era, and the accumulation of multi-source, multi-scale geochemical data provides new opportunities for quantitative classification of deposit types. However, existing studies are still limited by small sample sizes, insufficient coverage of deposit types, and relatively low classification accuracy. To address these issues, this study systematically compiles LA-ICP-MS trace element data of pyrite from seven types of gold deposits, including iron oxide–copper–gold (IOCG), volcanogenic massive sulfide (VMS), porphyry, low-sulfidation and high-sulfidation epithermal, Carlin-type, and orogenic deposits, and constructs a large geochemical dataset. After missing value control and centered log-ratio (CLR) transformation, Support Vector Machine (SVM) and Random Forest (RF) models are applied to classify deposit types, and their performance is systematically evaluated using cross-validation and multiple metrics. The results show that the SVM model based on the original CLR feature space achieves the best performance in terms of classification accuracy and stability, whereas the RF model, while maintaining strong predictive capability, reveals the key trace element combinations controlling classification through SHAP analysis. Overall, this study demonstrates that machine learning methods can effectively extract geologically meaningful mineralization signatures, and that the integration of high-performance classification models with interpretability analysis provides an effective approach for the identification and genetic study of gold deposit types in complex mineralization systems.

Accurate identification of ore-forming intrusions is critical for porphyry copper exploration. Previous studies have extensively used trace-element compositions of magmatic accessory minerals, such as zircon and apatite, to evaluate the metallogenic potential of magmas. However, the trace-element compositions of accessory minerals are controlled by multiple factors, which may limit their effectiveness in recording ore-forming magma signatures. Biotite is a common rock-forming mineral in intermediate to felsic igneous rocks, and its major-element compositions are mainly controlled by the parental magma, making it a potential indicator mineral for magmatic metallogenic potential. Based on a systematic compilation of biotite data from porphyry copper ore-forming intrusions and barren intermediate to felsic intrusions worldwide, this study selected nine major elements of biotite to construct three machine-learning models, namely support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGB), for identifying porphyry copper ore-forming intrusions. The results show that all three models perform well, with accuracies of 0.9561, 0.9614, and 0.9587, respectively, indicating that this approach can effectively identify ore-forming intrusions related to porphyry copper mineralization. SHAP analysis further reveals that FeO^T, Na₂O, and MnO are the three most important features controlling model outputs, and that biotite from ore-forming intrusions is generally characterized by low FeO^T, low MnO, and high Na₂O. These results demonstrate that the major-element compositions of biotite can effectively indicate the metallogenic potential of intrusions, providing a new approach for identifying ore-forming intrusions in porphyry copper systems and for guiding exploration of similar deposit types.

The orthoamphibole rocks exposed in the Hengshan-Wutai area are characterized by a distinctive mineral assemblages (anthophyllite + garnet) and unique geochemical signatures (enriched in Fe, Mg and Al, depleted in Ca and alkalis). However, their protolith and petrogenesis remain a subject of long-standing debate due to the perturbed effects of subsequent metamorphism and tectonic deformation. In this study, we conducted a systematic investigation of these rocks by integrating detailed field surveys and petrographic observations with machine learning algorithms and geochemical fingerprints. For the machine learning techniques, we constructed a geochemical database comprising Archean basalts and pelitic rocks and designed a Random Forest classification model by training classifers with the geochemical compositional database. Combined with Isocon mass balance analysis, which should be possible to formulate the relationships between immobile and potentially mobile elements, the trained classifers enable to classify the orthoamphibole rocks exhibiting a strong genetic affinity with the Archean basalts. Moreover, their rare earth element (REE) patterns and elevated Co and Sc concentrations are consistent with those of Archean basalts and distinct from typical pelitic rocks. Isocon mass balance analysis indicates that the protoliths underwent a net mass loss of approximately 26%, accompanied by intense leaching of mobile components such as Ca and Na. Notably, the decoupling of Zr–Hf, the anomalous enrichment of Th, and the scattered distribution of Zn suggest that the precursors underwent mechanical sorting during sedimentation. Additionally, zircon U–Pb dating yielded a primary detrital age peak at ～2520 Ma, suggesting a uniform provenance derived mainly from local ～2.52 Ga magmatic rocks. Metamorphic zircon ages cluster around 1.92 Ga, recording a Paleoproterozoic collision event in the North China Craton. Consequently, we propose that the protoliths of the Hengshan anthophyllite rocks was a mafic-rich sedimentary rock derived from Archean basaltic sources. These rocks formed through chemical weathering of the basaltic source, followed by depositional processes and subsequent amphibolite-facies metamorphism at ~1.92 Ga.

To address insufficient multi-source data fusion, strong subjectivity, low quantification, and difficulty in uncertainty quantification in traditional mineral exploration target optimization methods, this study constructs a probabilistic decision-making model integrating Hybrid Improved Fuzzy Analytic Hierarchy Process (HI-FAHP), Monte Carlo simulation, and machine learning. The model aims to achieve systematic fusion of multi-source information and probabilistic expression of results, improving the scientific reliability of exploration risk assessment in the early exploration stage. Taking the Zushimiao polymetallic mining area in Luanchuan County, Henan Province, as a case study, a HI-FAHP system is established by combining improved AHP, principal component analysis, and fuzzy evaluation. Weights are objectively determined for geological indicators (ore-bearing strata, fault structures, intrusive rocks, geophysical anomalies) and geochemical indicators (four groups: Ag-Pb-Sb, Au-Mo-As-Bi, Cu-V-Zn, W). Monte Carlo simulation is introduced to randomly perturb membership degrees (perturbation coefficient 0.08, triangular distribution) and grade boundary values (uniform distribution), generating probability distributions and ranking probabilities for comprehensive membership degrees of each target area. Three machine learning models—Decision Tree, Random Forest, and XGBoost—are trained on 5,000 randomly generated score datasets based on HI-FAHP weights. XGBoost is selected via error analysis to invert geological and geochemical weights and calculate the mineralization potential index. Five prospecting target areas (Ⅰ–Ⅴ) are delineated. Monte Carlo simulation yields comprehensive membership rankings: Target Ⅰ (79.55)>Target Ⅳ (76.70)>Target Ⅱ (74.50)>Target Ⅲ (71.30)>Target Ⅴ (67.20). XGBoost shows the best fit (R²=0.90), with inverted weights of 0.25 for geological and 0.75 for geochemical features, and consistent mineralization potential index ranking. Engineering verification confirms that Target I exposes galena ore bodies, Target IV delineates copper ore bodies, while the other targets show weak mineralization, highly consistent with predictions. The HI-FAHP-Monte Carlo-machine learning integrated model achieves objective multi-source data fusion and uncertainty quantification. XGBoost effectively inverts physically meaningful feature weights. Verified by engineering results, the model is reliable and suitable for multi-source data integration and exploration risk assessment in the early exploration stage.

With the advancement of large language models and knowledge graph technologies, the construction of geoscience knowledge systems based on semantic modeling and logical reasoning has gradually become an important technical approach for intelligent mineral exploration. To address the long-standing issue of single-level ontological semantics in geoscience knowledge representation, this paper proposes a dual-axis ontology design centered on the "concept axis-function axis." This design integrates key geological elements such as stratigraphy, lithology, structure, and magmatic rocks, combined with exploration indicators including ore-bearing lithology, ore-controlling structures, and geochemistry, to construct a concept-function dual-axis ontology structure. Leveraging large language models, entity recognition and relation extraction from geological texts are achieved, and the Neo4j platform is utilized for graph-structured storage and reasoning applications of mineral deposit knowledge. Taking the Qukulekedong Gold-Antimony Ore Concentration Area as a case study, the dual-axis ontology is embedded into prompt engineering to enable knowledge extraction in the study area using large language models. Through semantic association and structural analysis of the knowledge graph, logical reasoning and pattern induction for exploration models are conducted. The results indicate that the dual-axis ontology model, by collaboratively representing geoscience entities through the dual semantic dimensions of the concept axis and function axis, demonstrates significant advantages in semantic retrieval, relationship organization, and knowledge reasoning. This model overcomes the limitations of traditional geological knowledge, which primarily relies on static attribute descriptions, and expands from structural representation to functional semantics. It provides a solid theoretical foundation and methodological support for constructing geoscience knowledge systems with intelligent representation and logical reasoning capabilities.

Accurate prediction of deep tight sandstone reservoir property parameters is essential for reservoir evaluation and sweet spot prediction. However, existing models struggle to fully characterize the complex nonlinear relationships between parameters and cannot achieve simultaneous prediction of multiple parameters. Taking the Yangxia Formation tight sandstone reservoir in the Kuqa Depression of the Tarim Basin's Jurassic System as a case study, this paper conducted high-pressure mercury intrusion and nuclear magnetic resonance experiments on cores to extract reservoir micro-parameters. Average pore throat radius, displacement pressure, fractal dimension, and sorting coefficient are selected as feature parameters. A multi-task learning (MTL) framework was introduced based on a single-task XGBoost model, with an attention mechanism employed to enhance feature selection capabilities. This study proposes a joint prediction model combining Extreme Gradient Boosting (XGBoost) and MTL, suitable for small-sample datasets, to simultaneously predict permeability and porosity. Multiple comparative experiments were conducted to evaluate model performance from multiple perspectives, including prediction accuracy and error metrics. The research findings indicate: ① Both high-pressure mercury intrusion porosimetry and NMR fractal dimension curves exhibit a "three-segment" characteristic, indicating that deep tight sandstone reservoirs possess typical multifractal features; ② The XGBoost-MTL model achieved high correlation coefficients of 0.91 and 0.95 for predicted and actual values in pore-permeability prediction, respectively. Its root mean square error was reduced by approximately 70% compared to the single-task XGBoost model, demonstrating excellent predictive performance; ③ The pore throat scale in the study area is dominated by medium-sized pores. Using the attention score method, it was found that the fractal dimension has the most significant impact on reservoir permeability, while the average pore throat radius has the most significant impact on porosity. This study presents a reliable data-driven approach for predicting the physical properties of deep tight sandstone reservoirs. The findings have been successfully applied in reservoir evaluation of the Yangxia Formation in the Kuqa Depression, with prediction errors controlled within 8.3%. This approach holds significant implications for the efficient development of deep tight sandstone gas reservoirs.

As oilfield development enters the middle and late stages, formation energy deficit becomes increasingly severe and injection-production conflicts become more prominent. Conventional empirical injection allocation and numerical simulation methods often face challenges such as high computational cost, difficult history matching, and insufficient prediction accuracy when applied to complex reservoirs with strong heterogeneity and rapid dynamic changes. To address these problems, an injection-production optimization method integrating production dynamic data, physical constraints, and adaptive machine learning is proposed. In this method, LSTM and XGBoost are used to construct a parallel prediction framework, and the prediction model is selected through multi-index competition. Kalman filtering and nonlinear diffusion filtering are adopted to reconstruct injection-production response signals, and a physics-constrained ANN combined with gradient sensitivity analysis is used to invert inter-well connectivity coefficients. On this basis, a dual-objective injection allocation optimization model considering both oil production enhancement and water-cut control is established and solved using the differential evolution algorithm. A case study from B Oil Production Plant shows that the physics-constrained LSTM achieves an R² of 0.9521 and reduces the MSE to 0.6513, outperforming the standard LSTM and XGBoost in both prediction accuracy and physical consistency. The inter-well connectivity inversion results identify dominant high-water-cut flow channels and agree well with geological understanding. After optimization, block oil production increases from 175.00 m³/d to 176.10 m³/d, with an incremental oil production of 1.10 m³/d. The results indicate that coupling data-driven models with seepage-physics constraints can reduce reliance on repeated history matching, improve the evaluation efficiency of injection-production schemes, and provide technical support for refined injection allocation and stable oil production with water control in high-water-cut oilfields.

Existing rock mass stability analysis methods are mostly based on simplified structural models, making it difficult to accur ately reflect the complex spatial network characteristics of structural planes. To address the lack of intelligent construction methods for three-dimensional structural plane networks and stability analysis using UAV-acquired data, this paper is based on UAV oblique photogrammetry to obtain multi-view images of rock masses. A deep learning model is constructed by fusing geometric and texture features to achieve automatic identification of rock mass structural planes and intelligent extraction of attitude parameters, thereby constructing a three-dimensional structural plane network model. This network is introduced into rock mechanics analysis and combined with block theory to automatically identify potentially unstable blocks and quantitatively calculate their stability. Finally, method validation is carried out from four dimensions: structural plane identification accuracy, block identification integrity, network topology rationality, and stability analysis consistency, while 3D visualization is used to reveal rock mass failure modes and stability characteristics. Experimental results show that the proposed method achieves an accuracy of 94.2%, a recall rate of 92.5%, and an intersection over union (IoU) of 87.6% for structural plane recognition. Block identification results indicate that when the volume increases from 10 m³ to 1000 m³, the safety factor decreases from 1.3 to 0.6, revealing the scaling effect law. By fusing geometric and texture features, this method optimizes the structural plane identification model, enhances the topological integrity of three-dimensional network construction, and elevates the dimension of rock mass stability analysis from individual structural planes to the network level, significantly improving the accuracy and engineering applicability of stability assessment.