Applications of deep learning for mineral exploration and geological data analysis in mining

 Corresponding author: Mu’awiya Baba Aminu. Copyright © 2025 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution Liscense 4.0. Applications of deep learning for mineral exploration and geological data analysis in mining Raymond Kudzawu-D’Pherdd 1, 4 and Mu’awiya Baba Aminu 2, 3, * 1 Mechanical Engineering Department, Colorado School of Mines, Golden CO-USA. 2 School of Materials and Mineral Resources Engineering, University Sains Malaysia, Nibong Tebal, Malaysia. 3 Department of Geology, Federal University Lokoja, Kogi State, Nigeria. 4 Department of Sustainable Mineral Resources Development, School of Mines and Build Environment. Global Journal of Engineering and Technology Advances, 2025, 25(01), 269–282 Publication history: Received on 23 May 2025; revised on 05 July 2025; accepted on 07 July 2025 Article DOI: https://doi.org/10.30574/gjeta.2025.25.1.0209 Abstract The speed at which deep learning (DL) is developing has brought a new dawn in the field of mineral exploration and analysis of geological data, and it has provided extremely useful tools to meet the increasing complexity and size of the geoscience data. In this review the importance of DL applications in mineral exploration is discussed in general and within the fields of remote sensing image classification, geophysical anomaly detection, geochemical pattern recognition and drilling data interpretation in particular. Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Generative Adversarial Networks (GANs) and transformer models form DL architectures that have been found to be most effective in the automation and improvement of geological tasks previously limited by manual interpretation and a lack of scalability. Other challenges evidenced in the study are data heterogeneity, deficiency of labeled datasets, and interpretability of models. In addition, this paper addresses future research topics, with the integration of multimodal data, better explainability, transfer learning, and real-time decision-making systems. This review highlights how deep learning can change the face of geoscience today and how it can create a more novel, productive and controllable mineral exploration. Keywords: Deep Learning; Mineral Exploration; Geological Data Analysis; Convolutional Neural Networks; Remote Sensing; Geophysical Anomalies; Geochemical Modeling; Drilling Data 1. Introduction 1.1. Overview of Mineral Exploration and Geological Data Analysis Mineral exploration and study of geological data are key support columns of the mining business that make the basis of locating and developing subsurface mineral resources. In the past, exploration was done through field mapping, geophysical and geochemical survey, core drilling and manual analysis of the results (Reedman, 2012). Although these are known to be reliable methods, they tend to be labor bearing, costly and can have human bias/error associated with it. Recently, the phenomena of the development of highly sensitive sensors and satellite technologies have resulted in the emergence of voluminous high-dimensional data in the form of hyperspectral images, geochemical assays, digital elevation models (DEMs), and 3D seismic surveys (Carranza & Laborte, 2015). Our current ability to analyze data has been overrun by the size and sophistication of such large datasets, which have necessitated an increase in automated and scalable forms of analysis. In addition, the sector of the mining industry as a whole is under the pressure to find new deposits amid more demanding natural conditions (geological and Global Journal of Engineering and Technology Advances, 2025, 25(01), 269–282 270 environmental) (Liu et al., 2023). The mineral deposits are becoming deeper in depth and geologically complex, so mineral explorers need new tools to model, recognize patterns and predict. This stands in stark contrast to the rapidly escalating global demand for minerals. According to projections by the World Bank, the production of certain minerals will need to increase by as much as 500% by 2050 to support the transition to clean energy technologies (World Bank Group, 2018, 2020). Today, the analysis of geological data requires the incorporation of multiple types of data including spatial, temporal, geophysical and geochemical data to create the brain of models that will be able to support the reliance of fundamental choices (Cracknell & Reading, 2014). Because of this, the sphere of exploration is experiencing a digital transformation, and data science and computational intelligence are its leaders. Artificial intelligence (AI) or, specifically, machine learning (ML) has become a transformative force across disciplines in recent years - geoscience being no exception. Applications of AI in mineral exploration have proved useful in anomaly detection, orebody modeling, geological classification and prospectivity prediction of mineral occurrences (Abedini et al., 2023). Such systems have the ability to manipulate bulk data, find some underlying patterns, and make predictions without excessive human input. In contrast to conventional statistical approaches, ML has the ability to discover nonlinear relationships in complicated data and thus it is capable of analyzing data in the field of geology, where data tend to be heterogeneous and multi-dimensional (Sun et al., 2020). Nevertheless, although numerous machine learning tools have been practical, most techniques are associated with extensive preprocessing, manual features engineering, which can be biased or result in poor model performance. The limitation of the traditional approaches is solved in deep learning (DL), which is a subset of ML, by automating the extraction of high-level features of raw data through the use of neural network constructions (LeCun et al., 2015). It has been very useful in association with complex geological structures, seismic profiles, and satellites where it can interpret geology using non-linear relationships and hierarchical learning models. Examples of the most popular categories are convolutional neural networks (CNNs), which are suitable in image-based geological applications, including lithological mapping or alteration zone detection, and recurrent neural networks (RNNs), which are well suited to time-series data, yielding, e.g., drilling logs and geophysical signals (Zheng et al., 2023; Kong et al., 2022). The strength of DL is not purely the accuracy but also how it is scalable across different data and the geological environment. It has the strength of being resistant to noisy inputs, and flexibility in the integration of spatial and temporal sources of data that enables geoscientists to be more confident in their conclusions (Agboola et al., 2024). Furthermore, the transfer learning and attention mechanisms are promising innovations that are making it easier to decrease large, labeled data requirements which are notable in mineral exploration branches (Ahari, 2024). A recent systematic review of machine-learning applications across the mining value chain highlights a rapid increase in datadriven methods for exploration, extraction, and reclamation. The authors emphasize that the main limitations are not algorithmic but rather tied to inconsistent validation practices and the absence of standardized benchmark datasets. Their findings underscore the need for reproducibility frameworks as the field moves toward more automated geological decision-making (Jung & Choi, 2021). Recent workforce projections suggest that more than half of the current mining labor force in the United States will retire by 2029, creating a substantial risk of institutional knowledge loss. This trend underscores the need for digital and automated systems that can retain geological expertise independent of personnel turnover (Wood & van As, 2024). Despite the growing body of literature applying deep learning to mineral exploration, few studies have systematically evaluated how algorithmic advances, such as attention mechanisms, encoder–decoder networks, and transfer learning, translate into geological discovery outcomes. This review addresses that gap by synthesizing methodological progress, evaluating comparative performance, and identifying practical integration pathways across exploration workflows. 1.2. Purpose and Scope of the Review The goal of this review is to investigate and summarize the increasing amount of literature within the appropriates of deep learning application to mineral exploration and analysis of geological data. It provides the description of the basic DL architectures used in the sphere, the main examples of case studies and the discussion of the implementation of DL to different types of geological data, such as geophysical, geochemical, and remote sensing data. In addition, the review clarifies and highlights important limitations of reviews arising, as well as proposing possible research directions (unavailability of enough training data, model interpretability, and interface with geoscientific knowledge). Offering a systematic review of this rapidly developing phenomenon, the review represents a useful contribution to the treasury of knowledge needed by geoscientists, data scientists, and mining specialists in need to Global Journal of Engineering and Technology Advances, 2025, 25(01), 269–282 271 increase the effectiveness of their exploration and streamline their decision-making with the application of complex computational tools. 2. Deep Learning Fundamentals for Geoscientific Applications 2.1. Deep Learning and Neural Networks: An Overview Deep learning (DL) is a part of machine learning (ML) that uses multiple-layer artificial neural networks (also referred to as deep neural networks) to learn high-level abstractions in data (Agboola & Hashemi, 2024). Due to the human brain structure and its operations, the models are comprised of interconnected nodes called neurons, the arrangement of which is input, hidden and output layers (Aggarwal, 2018). Every neuron provides inputs, performs a weighted sum, and, then, activates the signal using a non-linear effect, which is then transferred to the subsequent layer. DL models can learn complicated patterns and representations on raw data through this structure without the need of manual feature engineering (LeCun et al., 2015; Schmidhuber, 2015). Interpretable ensemble architectures such as deep forests have recently demonstrated performance comparable to deep neural networks while retaining transparent featureimportance outputs. In mineral prospectivity mapping, this approach achieved an AUC above 0.96, showing that highaccuracy prediction does not require fully opaque models. The results provide a counterexample to the assumption that explainability and performance are mutually exclusive (Dong & Zhang, 2024). Some of the most widespread deep learning architectures applied in geosciences are 2.2. Convolutional Neural Networks (CNNs) CNNs have been found to perform well in manipulation of spatial maps like satellite images, seismic sections and the geological maps. Convolutional filters enable them to automatically acquire spatial features in the form of faults, lineaments, mineral alteration zones, and stratigraphic boundaries (Zheng et al., 2022). CNNs have also found use in mineral exploration classifying lithologies and identifying geochemical anomalies as well as remote sensing data to define mineralized zones (Cracknell & Reading, 2014; Sun et al., 2020). Figure 1 ATT-CNN structure for mineral prospectivity modeling (Sun et al., 2024) Figure 1 illustrates the architecture of the Attention-based Convolutional Neural Network (ATT-CNN) proposed by Sun et al. (2024) for mineral prospectivity modeling. The model integrates conventional convolutional layers with an attention mechanism that adaptively weights spatial and spectral features, allowing the network to focus on geologically meaningful signals such as alteration halos, faults, and lineaments. In this configuration, attention layers act as feature selectors that suppress irrelevant noise from multisource geospatial data, thereby improving predictive accuracy. Including this figure highlights how modern CNN architectures can automatically extract and prioritize geological patterns, an advance that reduces manual feature engineering and strengthens reproducibility in mineral targeting workflows. Recurrent Neural Networks (RNNs): RNNs are sequential data tools and can learn about a time-varying input like borehole logs, well log data or geophysical time-series (Aminu et al., 2024). Yet, the problem of vanishing gradients is mitigated in the enhanced versions of RNNs called Long Short-Term Memory (LSTM) and Gated Recurrent Global Journal of Engineering and Technology Advances, 2025, 25(01), 269–282 272 Unit (GRU) models, which can capture long-range dependencies in sequences and, thus, can be used in geological modeling where the changes of the subsurface can be represented either by depth or by time (Graves et al., 2013; Sherstinsky, 2020). Transformer Models: Transformers were initially used in natural language processing, but they have recently been repurposed into spatial and sequential data in the field of geosciences because they can represent long-range dependencies with fewer parameters than RNNs; they can do this because of their self-attention mechanisms (Vaswani et al., 2017). Such models can be especially helpful when combining data sources that help to do more in-depth subsurface interpretations (i.e. remote sensing and geophysical logs) (Xie et al., 2024). Recurrent architectures such as LSTMs are increasingly used for sequential geochemical or temporal monitoring data, allowing the capture of oresystem evolution over time. 2.3. Types of Geological Data Used in Deep Learning Practical usage of deep learning in geological applications involving geoscientific data need a good comprehension of the different forms of geoscientific data. The type of these datasets is very diverse by format, scale and resolution: Geophysical Data: This data consists of Magnetic, gravity, radiometmic, and seismic data giving the structure and composition of subsurface. Such data is normally obtained during airborne or terrestrial surveys and are commonly employed to outline the faults, lunches, intrusive bodies, and potentially profitable mineral set ups (Zuo et al., 2019). CNNs and other deep learning models have proved very promising in the automation of interpretation of these dataset as they have found out minor anomalies and patterns. Geochemical Data: Data is based on the chemical analysis of rocks, soils, stream sediments or water samples to determine the location of high concentrations of any element such as gold, copper or rare earth elements. Those dataset is important in identifying the anomaly, as well as map of mineral prospect (Zuo et al., 2019). Hidden connection between geochemical signatures and the underlying mineralization may be uncovered by the application of DL methods. Remote Sensing Data: Remote sensing data are information obtained through satellite-based platforms like Landsat, ASTER, Sentinel-2 and hyperspectral sensors that allow acquisition of the spatially continuous data regarding the surface of the earth. These are mineralogical composition, the vegetation indices, landforms, and hydrothermal alterations (Zheng et al., 2022). CNNs find wide application in the categorization and deciphering of remote sense images when it comes to mineral searching. Core Logs and Drilling Data: These so-called ground truth data contain core samples and lithological logs, gamma-ray logs and other borehole measurements information. They are essential in the interpretation in the vertical geologic variation of the subsurface. Relying on DL models, primarily LSTMs and transformers, automation of classifying and predicting lithological units based on well logs has been conducted (Fu et al., 2022; Xing et al., 2023). Topographic and Structural Data: Topographic and Structural data include digital elevation models (DEMs), structural lineaments, and fault maps to offer the context data to mineral exploration activities, and this source of information is best analyzed along with other data (Sun et al., 2024). When such disparate datasets are combined, DL models can create more precise and encompassing definitions of the geological context to enhance the productivity and the degree of success of mineral exploration programs (Syum et al., 2025). Graph neural networks (GNNs) are promising for relational geodata such as faults, lineaments, and drill-hole networks, providing a structure-aware alternative to pixel-based learning. 3. Applications of Deep Learning in Mineral Exploration Due to its capability to perform on high-dimensional, non-linear and heterogeneous data, deep learning (DL) has since grown to be of great value in mineral exploration. Ranging between satellite imagery to geophysical signals and borehole logs, DL models have shown a great promise in the recognition of patterns, anomaly detection, and predictive analyses. Here the application areas are discussed which are related to remote sensing image classification, detection of geophysical anomalies, geochemical pattern recognition, and interpretation of drilling data. 3.1. Lithological mapping using Remote Sensing Image Classification By contrast, the conventional lithological mapping is based on manual analysis of satellite/airborne imagery, which consumes much time due to the bias of human perception and complex spectral characteristics (Aminu et al., 2024). Global Journal of Engineering and Technology Advances, 2025, 25(01), 269–282 273 Deep learning architectures, such as Convolutional Neural Networks (CNNs), or U-Nets remove the laborious process of locating rock units and discover the spatial-contextual characteristics of multispectral, hyperspectral or radar observations. Farahbakhsh et al., (2025) also showed that CNNs using Landsat and ASTER data gave more accurate lithological map compared to the traditional classifiers and that they are applicable when generating lithological maps in areas that have complex geological conditions, such as Western Australia (Figure 2). Likewise, Transformer networks have recently become able to efficiently process sparse training labels on arid data, such as that used to map gossans in the Saudi Arabian Shield with 94.1 percent accuracy (Shirmard et al., 2022), using WorldView-3 imagery. These techniques save a lot in the cost of field surveys and enhance map reproduction. Figure 2 presents the deep-learning framework developed by Farahbakhsh et al. (2025) for alteration-zone mapping using multispectral and hyperspectral imagery. The workflow integrates a CNN-based encoder–decoder architecture that fuses spectral, spatial, and textural information to generate pixel-level alteration-probability maps. This approach exemplifies how deep networks translate remote-sensing data into predictive geological layers, bridging image classification and mineral prospectivity modeling. By visualizing data flow from raw satellite imagery through feature extraction to final mineralization mapping, the figure underscores the operational value of end-to-end learning pipelines in mineral exploration. Collectively, these studies illustrate a methodological shift from pixel-based classification toward spatial-contextual learning, where attention mechanisms and encoder, decoder structures enhance feature interpretability and geological realism. Table 1 Common Types of Geological Data and Their Relevance to Deep Learning Applications (After Moon et al., 2006; Bergen et al., 2019) Type of Geological Data Description and Typical Characteristics Deep Learning Relevance Geophysical Data Magnetic: Variations in Earth's magnetic field due to magnetic minerals; grid/profile data. CNNs for anomaly detection; RNNs for profile analysis. Gravity: Variations in Earth's gravitational field due to density contrasts; grid data. CNNs for density structure interpretation. Seismic: Acoustic waves propagating through the Earth to image subsurface structures; 2D/3D volumes. CNNs (U-Net) for fault detection, facies classification. Electromagnetic (EM): Electrical conductivity variations; profile/grid data. CNNs for conductive body identification, inversion. Geochemical Data Concentrations of elements in soil, rock, water, vegetation samples; point data, multi-element. NNs, Autoencoders for pattern recognition, anomaly identification. Remote Sensing Data Satellite Imagery: Multispectral, hyperspectral, radar images; spatial/spectral resolution. CNNs for lithological mapping, alteration detection. Drone/UAV Imagery: High-resolution orthophotos, 3D models from photogrammetry. CNNs for detailed surface mapping, outcrop analysis. Drilling Data Core Logs: Lithology, RNNs for lithofacies mineralogy, alteration, structural features, RQD; sequential textual/image data. prediction; CNNs for image-based logging. Assay Data: Element concentrations from drill core; point data along boreholes. NNs for grade estimation, interpolation. Geotechnical Logs: Rock strength, RQD, fracture frequency; numerical/textual logs. RNNs for rock mass characterization. Mapping Data Geological Maps: Lithological units, structural features, contacts; vector/raster data. CNNs for map interpretation, feature extraction. Global Journal of Engineering and Technology Advances, 2025, 25(01), 269–282 274 3.2. Geophysical Anomaly Detection Large data (magnetic, gravity, seismic) arise in geophysical surveys, and the mineralization effect may be obscured by noise or in regional trends. DL models identify minute irregularities associated with ore bodies as a consequence of finding latent data distributions. Sun et al. (2024) confirm that autoencoders use magnetic data to clean data to detect IOCG (iron oxide copper-gold) targets. Generative Adversarial Networks (GANs) have been used to generate realistic anomalies which are then used to boost small training data sets; Sun et al. (2020) apply GANs to locate kimberlite pipes in the glaciated terrains of Canada, boosting anomaly detection success by 30 percent compared to conventional filtering. In case of 3D seismic data, 3D CNNs (Zheng, 2022) can identify volumetric features that are related to volcanic massive sulphides (VMS) and accurately target these features below the surface. 3.3. Ore Targeting by Geochemical Pattern Recognition The nonlinear interactions between elements are high-dimensional and do not lend to easy statistical modeling through traditional statistics. DL analyzes multi-element signatures that are indicators of mineralization. He et al. (2022) used Random Forests combined with Deep Neural Networks (DNNs) to study the geochemistry of soil in the Jiaodong gold province of China with an AUC score of 0.91 during prospectivity mapping. LSTM networks imply the sequential modeling of geochemical alterations in drill cores, and Xing et al., (2023) predict Cu-Ni-PGE zones in the Canadian Shield (F1-score: 0.89). Graph Neural Networks (GNNs) (Zuo et al., 2019) have recently been used to model regional geochemical surveys as spatial graphs and have been able to identify porphyry copper districts in Chile with 93 percent accuracy, by learning the correlations between elements and also their spatial interactions. 3.4. Drilling Data Interpretation and Predictive Modeling Drilling generates heterogeneous data (core images, downhole assays, geophysics) where manual integration is errorprone. DL unifies these streams for predictive modeling. ResNet-50 classifiers automate lithology identification from core scans, Agrawal& Govil (2023) reduced logging time by 85% in a porphyry copper project. Hybrid 1D CNN-LSTM architectures fuse downhole geophysics and assays to predict gold grades (Santos et al., 2021), cutting assay turnaround time while maintaining low error (RMSE: 0.23 ppm). Most innovatively, Fu et al. (2022) combined Transformers with Gaussian Processes to generate 3D orebody models from sparse multi-hole assays, reducing grade uncertainty by 90% and optimizing drill-hole planning. Machine-learning workflows that integrate drillhole and geophysical data now achieve structural and lithological interpretations that closely align with expert geologists. These systems substantially reduce the amount of manual logging and allow predictive models to be updated in near-real-time as new drilling results are acquired. This illustrates a practical pathway toward automated subsurface modeling in active exploration settings (Wedge et al., 2019). 4. Applications of Deep Learning in Geological Data Analysis Deep learning has emerged as a powerful tool in geological data analysis, offering innovative solutions to complex challenges in subsurface interpretation, rock classification, structural detection, and data processing (Ahari, 2024). By leveraging advanced neural networks, geoscientists can now extract meaningful insights from vast and diverse geological datasets with unprecedented accuracy and efficiency (Zhao et al., 2024). These applications are transforming traditional workflows, reducing human bias, and accelerating decision-making processes across the mining, oil, and gas industries (Liu et al., 2023). Prospectivity modeling using deep autoencoders combined with large geochemical and geophysical datasets has demonstrated meaningful improvements over manually weighted index overlays. In one case study, more than forty variables were compressed into a low-dimensional representation that still preserved the spatial signatures of skarn mineralization. This provides evidence that unsupervised deep learning can extract exploration features that are not visually obvious to interpreters (Xiong et al., 2018). Deep-learning models trained on seafloor imagery have been used to classify terrain, detect hydrothermal mound structures, and infer mineralization potential at resolutions impractical for manual interpretation. These applications broaden the scope of ML in exploration to environments that cannot be physically mapped at scale, such as deep-sea deposits (Juliani & Juliani, 2021). 4.1. 3D Subsurface Modeling and Visualization The construction of accurate 3D subsurface models is critical for resource exploration and reservoir characterization (Zhou et al., 2020). Traditional methods rely heavily on manual interpretation of seismic data, well logs, and geophysical surveys, which can be time-consuming and subjective (Zhang et al., 2019). Deep learning techniques, particularly convolutional neural networks (CNNs) and generative adversarial networks (GANs), have demonstrated remarkable success in automating this process (Wu et al., 2021). CNNs excel at identifying geological features from seismic reflections (Li et al., 2022), while GANs can generate realistic synthetic models to augment limited datasets (Sun Global Journal of Engineering and Technology Advances, 2025, 25(01), 269–282 275 &AlRegib, 2020). Transformer-based architecture further enhances model performance by capturing long-range dependencies in seismic volumes (Vaswani et al., 2023). Figure 2 Deep learning framework for mapping alteration zones utilising remote sensing data. (After Farahbakhsh et al., 2025) 4.2. Rock Type Classification Using Core and Thin-Section Images Accurate rock classification is fundamental to geological analysis, yet manual identification from drill cores and thin sections remains labor-intensive (Alzubaidi et al., 2021). Deep learning models, particularly deep convolutional neural networks (DCNNs), have proven highly effective in automating this task (Karimpouli& Tahmasebi, 2019). These models can distinguish between various rock types, such as sandstone, shale, and granite, with high precision by analyzing textural and mineralogical patterns in images (Jiang et al., 2020). Recent advancements in self-supervised learning have further improved classification accuracy by reducing reliance on large labeled datasets (Chen & Zhang, 2022). Additionally, multimodal learning frameworks that integrate core images with geochemical and geophysical data provide a more comprehensive understanding of rock properties (Wang et al., 2023). 4.3. Fault and Fracture Detection from Seismic Data Faults and fractures play a crucial role in hydrocarbon migration, groundwater flow, and reservoir stability (Zeng et al., 2021). Traditional seismic interpretation methods often struggle with subtle or complex fault geometries (Wu & Hale, 2020). Deep learning approaches, such as U-Net architectures with attention mechanisms, have shown exceptional performance in detecting and delineating faults within 3D seismic datasets (Zhao et al., 2022). These models leverage hierarchical feature extraction to highlight discontinuities in seismic reflections, even in noisy or low-resolution data (Guo et al., 2023). Graph neural networks (GNNs) offer another promising avenue by representing seismic data as spatial graphs (Yang et al., 2022). The application of semi-supervised learning techniques further enhances fault detection by minimizing the need for extensive manual annotations (Liu & Zhang, 2023). 4.4. Automation in Logging, Sampling, and Data Interpretation Automation of the logging of wells, sampling of the core, and the interpretation of the geological data is a big step towards the optimization of the processes (Tian et al., 2021). Deep learning models make these processes more efficient as they collect valuable data out of various sources (Smith et al., 2022). Using natural language processing (NLP) Global Journal of Engineering and Technology Advances, 2025, 25(01), 269–282 276 approaches, it is possible to extract the lithologies descriptions based on unstructured texts in an automatic way (Brown et al., 2023). The algorithms of reinforcement learning are retrained to optimize real-time drilling decisions based on information obtained by sensors in the hole (Wang & Chen, 2023). Yet another use of automated machine learning (AutoML) frameworks is to further improve log interpretation through dynamic choice of most appropriate algorithm to a particular geological setting (Zhang & Li, 2024). Studies comparing neural networks, SVMs, and evolutionary hybrids with conventional geostatistical estimators show that ML models can handle highly skewed grade distributions and sparse drill datasets more effectively. These newer approaches also allow multi-output prediction, enabling simultaneous estimation of grade and geological uncertainty. This trend indicates that ML is moving beyond experimental use into resource classification workflows (Battalgazy et al., 2023). Automated extraction of text, figures, and tables from historical exploration reports has now reached the point where knowledge graphs can be built directly from unstructured PDF archives. This enables semantic search across decades of reports without manual curation, reducing information-retrieval bottlenecks in early-stage targeting. The work illustrates how ML is transforming not just geological modeling, but the data-access layer itself (Qiu et al., 2024). 5. Progression of deep learning in Mineral Exploration. Over the past several years, mineral exploration has seen considerable transformative advances through the adoption of deep learning and machine learning techniques. These innovations have steadily addressed peculiar core challenges which include handling complex, multi-source geological data, automating extraction of geological knowledge, and improving prediction accuracy for mineral prospectivity. Table 2 Comparison of Traditional and Deep Learning Approaches in Key Geological Applications (After Wang et al., 2020; Di et al., 2018; Pires et al., 2020). Application Area Traditional Approach Deep Learning Approach Advantages of DL Lithological Mapping Manual interpretation of aerial photos, limited spectral analysis. CNN-based classification of multispectral/hyperspectral imagery. Higher accuracy, speed, automation, finer detail. Geophysical Anomaly Detection Visual inspection, statistical filtering, CNNs for automatic feature extraction, denoising, and anomaly highlighting. Detects subtle anomalies, reduces human bias, faster processing. rule-based algorithms. Geochemical Pattern Univariate statistics, NNs/Autoencoder s for complex Uncovers non-linear Recognition principal component analysis, cluster analysis. multi-element pattern recognition, predictive mapping. relationships, better targeting in complex datasets. Drill Core Logging Manual visual inspection, descriptive logging, subjective. CNNs for image-based classification (lithology, alteration, structures); RNNs for sequential log interpretation. Consistent, objective, faster, reduced human error. 3D Geological Modeling Manual implicit modeling, interpolations based on limited data. Generative models (GANs, VAEs) for automated implicit modeling from sparse data. Accelerated model generation, realistic geological realizations, uncertainty quantification. Fault/Fracture Detection (Seismic) Manual picking/interpretation, attribute analysis (e.g., semblance). U-Net based semantic segmentation of seismic volumes. Automated, more accurate, identifies complex fault networks. Global Journal of Engineering and Technology Advances, 2025, 25(01), 269–282 277 The foundational milestone in this journey came in 2018, when Xiong et al. introduced an unsupervised deep autoencoder network combined with big data analytics to map skarn-type iron deposits in China’s Fujian metallogenic zone. They confronted the challenge of detecting buried mineral deposits in dense forested regions where direct geophysical information is weak or incomplete. Their deep learning model ingeniously fused 42 geological, geochemical, and geophysical variables, enabling anomaly detection via reconstruction errors. The model identified high prospectivity zones with strong spatial correspondence to known mineral deposits, establishing a new paradigm for integrating multi-modal data with DL. However, spatial resolution constraints and certain modeling assumptions were acknowledged as areas for refinement. Building on this foundation, Shi Li et al. (2019) harnessed the power of convolutional neural networks (CNN), specifically AlexNet, to interpret 2D geological “images” derived from chemical element maps and geological factors. Their approach tackled the intricate spatial prediction of manganese deposits in China’s Songtao–Huayuan area. By leveraging CNN’s spatial pattern recognition strengths, their model achieved perfect training accuracy and robust validation, successfully delineating prospective ore zones. This marked a pivotal step, demonstrating how computer vision architectures could be repurposed for geological data analysis. While impactful, the study’s focus on manganese and data scale limitations highlighted opportunities for broader application and enhanced 3D integration. Dumakor-Dupey and Arya (2021) provides a crucial consolidation of machine learning applications in mineral resource estimation, reviewing how ML methods like Artificial Neural Networks (ANN), Support Vector Machines (SVM), and Random Forests (RF) excel in capturing complex, nonlinear geological patterns that challenge classical geostatistics. They emphasized ML’s role in automating traditionally manual and linear estimation processes, spotlighting its growing maturity. Yet, the authors also urged the community to tackle interpretability and data quality challenges to fully leverage these tools. Haiming Liu et al. (2022) continued with this practical trend and demonstrated the efficacy of Random Forests in mapping gold mineral prospectivity in Ontario, Canada. Their integration of 32 geophysical and geochemical variables achieved high accuracy in identifying mineralized zones, especially those aligned with the Larder Lake-Cadillac deformation zone - a key structural control on gold deposits. Their work exemplified how combining diverse data sources with ML enhances exploration targeting, while also highlighting the sensitivity of models to data quality and the value of ensemble learning strategies. Cai Liu et al. (2023) progress to more sophisticated architectures, they introduce a self-attention back-propagation neural network (SA-BPNN) tailored for porphyry-epithermal mineral prospectivity in Tibet. By integrating selfattention mechanisms with expert geological input, they overcame challenges of uncertain data and sample imbalance, outperforming classical ML models like SVM and U-Net. Their framework not only improved prediction accuracy but also advanced target delineation, reflecting a methodological leap in blending advanced DL architectures with domain expertise. The study’s focus on a specific deposit type suggests fertile ground for extending this approach. The work of Hasan et al. (2023) underscore the vast potential of AI and ML with a broad to accelerate mineral exploration, while addressing data quality, model transparency, and ethical considerations as key hurdles simultaneously. Their call for improved algorithms, data stewardship, and ethical frameworks signifies the complex socio-technical landscape in which these technologies operate. Azhari et al. (2023) complement this with acritical review of deep learning implementations across mining, revealing a predominant focus on extraction (72% of studies) with CNNs dominating, primarily for image analysis. However, exploration and reclamation remain underrepresented, as welas being hampered by limited data availability and proprietary restrictions. Their analysis advocates for enhanced sensor diversity, public datasets, and interpretability to broaden DL’s mining impact. In 2024, Jintao Tao et al. innovated with a deep learning-based named entity recognition (NER) model specifically designed for Chinese literature on granitic pegmatite lithium deposits. Their MENER model, incorporating CNN, BiLSTM, multi-head attention, and CRF, dramatically improved the automated extraction of detailed geological entities, addressing the bottleneck of manual literature review. Though currently languageand deposit-type specific, this work points toward scalable, domain-tailored knowledge discovery. Concurrently, Magdalena Radulescu et al. (2024) advanced sustainable mining by integrating a quantized deep learning EfficientDet model with financial technology (FinTech) tools. Their approach delivered near-perfect mineral identification accuracy alongside real-time environmental monitoring and investment transparency, heralding a new