Pub Date : 2026-01-20DOI: 10.1016/j.acags.2026.100321
Weilin Chen, Jiyin Zhang, Chenhao Li, Xiaogang Ma
{"title":"Corrigendum to “Enhancing neuro-symbolic AI for mineral prediction via LLM-guided knowledge integration” [Appl. Comput. Geosci. 29 (2026) 100310]","authors":"Weilin Chen, Jiyin Zhang, Chenhao Li, Xiaogang Ma","doi":"10.1016/j.acags.2026.100321","DOIUrl":"10.1016/j.acags.2026.100321","url":null,"abstract":"","PeriodicalId":33804,"journal":{"name":"Applied Computing and Geosciences","volume":"29 ","pages":"Article 100321"},"PeriodicalIF":3.2,"publicationDate":"2026-01-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146037979","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-16DOI: 10.1016/j.acags.2026.100320
Hakim Saibi , Abdelhadi Hireche , Takeshi Tsuji , Mohammed Y. Ali , Ahmad B. Ahmad
This paper presents a comparative study of three deep learning (DL) architectures for one-dimensional magnetotelluric (MT) data inversion: long short-term memory (LSTM), gated recurrent unit (GRU), and Informer models. We developed a comprehensive framework for generating realistic synthetic MT data, training the models, and evaluating their performance through multiple quantitative metrics. Our synthetic dataset comprised 100,000 samples with 25 periods spanning 10−3 to 10 seconds, created using statistical parameters derived from real MT data. Each model was trained on apparent resistivity and phase responses to recover subsurface resistivity profiles. The results show that the recurrent neural network architectures (LSTM and GRU) slightly outperform the attention-based Informer model, with the LSTM achieving the best performance (MSE of 0.06455 , R of 0.45234). Despite their differing architectures, all the models successfully captured the major subsurface resistivity contrasts. When applied to real MT data from the UAE, the tested models showed promising results in terms of reconstructing subsurface structures. Overall, this study demonstrates the viability of DL approaches for MT inversion, with potential applications in efficient field-based subsurface imaging.
{"title":"Comparison of deep learning models for 1D magnetotelluric inversion","authors":"Hakim Saibi , Abdelhadi Hireche , Takeshi Tsuji , Mohammed Y. Ali , Ahmad B. Ahmad","doi":"10.1016/j.acags.2026.100320","DOIUrl":"10.1016/j.acags.2026.100320","url":null,"abstract":"<div><div>This paper presents a comparative study of three deep learning (DL) architectures for one-dimensional magnetotelluric (MT) data inversion: long short-term memory (LSTM), gated recurrent unit (GRU), and Informer models. We developed a comprehensive framework for generating realistic synthetic MT data, training the models, and evaluating their performance through multiple quantitative metrics. Our synthetic dataset comprised 100,000 samples with 25 periods spanning 10<sup>−3</sup> to 10<span><math><msup><mrow></mrow><mrow><mn>3</mn></mrow></msup></math></span> seconds, created using statistical parameters derived from real MT data. Each model was trained on apparent resistivity and phase responses to recover subsurface resistivity profiles. The results show that the recurrent neural network architectures (LSTM and GRU) slightly outperform the attention-based Informer model, with the LSTM achieving the best performance (MSE of 0.06455 <span><math><mrow><mi>Ω</mi><mspace></mspace><msup><mrow><mtext>m</mtext></mrow><mrow><mn>2</mn></mrow></msup></mrow></math></span>, R<span><math><msup><mrow></mrow><mrow><mn>2</mn></mrow></msup></math></span> of 0.45234). Despite their differing architectures, all the models successfully captured the major subsurface resistivity contrasts. When applied to real MT data from the UAE, the tested models showed promising results in terms of reconstructing subsurface structures. Overall, this study demonstrates the viability of DL approaches for MT inversion, with potential applications in efficient field-based subsurface imaging.</div></div>","PeriodicalId":33804,"journal":{"name":"Applied Computing and Geosciences","volume":"29 ","pages":"Article 100320"},"PeriodicalIF":3.2,"publicationDate":"2026-01-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145977262","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-09DOI: 10.1016/j.acags.2026.100318
Fan Xiao , Ao Tang , Dunhui Xiao , Shu Jiang , Qiuming Cheng
Tectonically induced fault systems create favorable environments for hydrothermal fluid migration and mineral precipitation, ultimately leading to the formation of certain deposits. Therefore, the numerical simulation of tectonic stress fields in mineralizing dynamic systems is crucial for understanding mineralization processes and for predictive modeling of ore prospectivity related to structure-controlled deposits. However, uncertainties in rock-physical parameters can undermine the reliability of simulation results. Although trial-and-error methods may seem to offer a solution to this issue, they typically necessitate extensive numerical simulations, leading to high computational costs. In this study, we developed a deep learning-based surrogate modeling method, using the Fankou lead-zinc deposit as a case study. This method enables rapid and accurate calculations of complex stress field distributions within mineralizing dynamic systems. We employed the Sobol method to analyze the sensitivity of the first, second and third principal stresses, to critial rock-physical parameters, including density, Poisson's ratio, and elastic modulus. This analysis allowed us to identify the critical parameters for the mineralization dynamics model. Subsequently, we utilized Latin hypercube sampling (LHS) to generate parameter sets within the key parameter space for numerical simulations, resulting in high-fidelity datasets of three distinct stress field distributions. Finally, we constructed an end-to-end surrogate model for each of the three stress fields based on deep neural networks, using the rock-physical parameters generated by LHS as input variables and the high-fidelity datasets obtained from numerical simulations as prediction variables. Evaluation metrics in the test dataset, including the correlation coefficient, root mean square error, mean relative error, and mean absolute error, indicate that the surrogate models perform well, effectively capturing the spatial distribution of the tectonic stress field within the mineralizing dynamic system. Our results also demonstrate that the average computational efficiency of the data-driven surrogate models is approximately 200 times greater than that of numerical simulation methods, providing an effective and rapid computational framework for parameter inversion and optimization related to the tectonic stress field in complex mineralizing dynamic systems.
{"title":"Simulating tectonic stress fields in mineralization dynamics using a deep neural networks-based surrogate model","authors":"Fan Xiao , Ao Tang , Dunhui Xiao , Shu Jiang , Qiuming Cheng","doi":"10.1016/j.acags.2026.100318","DOIUrl":"10.1016/j.acags.2026.100318","url":null,"abstract":"<div><div>Tectonically induced fault systems create favorable environments for hydrothermal fluid migration and mineral precipitation, ultimately leading to the formation of certain deposits. Therefore, the numerical simulation of tectonic stress fields in mineralizing dynamic systems is crucial for understanding mineralization processes and for predictive modeling of ore prospectivity related to structure-controlled deposits. However, uncertainties in rock-physical parameters can undermine the reliability of simulation results. Although trial-and-error methods may seem to offer a solution to this issue, they typically necessitate extensive numerical simulations, leading to high computational costs. In this study, we developed a deep learning-based surrogate modeling method, using the Fankou lead-zinc deposit as a case study. This method enables rapid and accurate calculations of complex stress field distributions within mineralizing dynamic systems. We employed the Sobol method to analyze the sensitivity of the first, second and third principal stresses, to critial rock-physical parameters, including density, Poisson's ratio, and elastic modulus. This analysis allowed us to identify the critical parameters for the mineralization dynamics model. Subsequently, we utilized Latin hypercube sampling (LHS) to generate parameter sets within the key parameter space for numerical simulations, resulting in high-fidelity datasets of three distinct stress field distributions. Finally, we constructed an end-to-end surrogate model for each of the three stress fields based on deep neural networks, using the rock-physical parameters generated by LHS as input variables and the high-fidelity datasets obtained from numerical simulations as prediction variables. Evaluation metrics in the test dataset, including the correlation coefficient, root mean square error, mean relative error, and mean absolute error, indicate that the surrogate models perform well, effectively capturing the spatial distribution of the tectonic stress field within the mineralizing dynamic system. Our results also demonstrate that the average computational efficiency of the data-driven surrogate models is approximately 200 times greater than that of numerical simulation methods, providing an effective and rapid computational framework for parameter inversion and optimization related to the tectonic stress field in complex mineralizing dynamic systems.</div></div>","PeriodicalId":33804,"journal":{"name":"Applied Computing and Geosciences","volume":"29 ","pages":"Article 100318"},"PeriodicalIF":3.2,"publicationDate":"2026-01-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145977263","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Detecting subtle anomalies in well log data can significantly enhance subsurface characterization and inform reservoir decision-making. Well logs provide detailed records of rock and fluid properties. However, identifying anomalous patterns in these data remains a persistent challenge due to the complexity of geological systems and limitations in traditional statistical models. Classical anomaly detection approaches, such as Gaussian mixture models (GMMs), tend to oversimplify the intricacies of well log responses and often misinterpret geological heterogeneities as tool noise. This misclassification leads to high false positive rates and unreliable anomaly identification. In this context, generative models, particularly generative adversarial networks (GANs), have shown promise in structured data domains; however, they remain underutilized in geoscience applications. This study aims to benchmark the performance of ensemble GANs (EGANs) against GMMs for anomaly detection in well log data. The proposed EGANs framework aggregates multiple independently trained GANs to enhance model stability and robustness. Anomalies are detected based on discriminator scoring, while performance is evaluated using precision, recall, and F1-score across four key logs: gamma ray, travel time, bulk density, and neutron porosity. The results demonstrate that EGANs consistently outperform GMMs across all logs, achieving higher precision (up to 0.70) and F1-scores (up to 0.79), with statistically significant improvements confirmed via paired t-tests. These findings highlight the ability of EGANs to model complex subsurface patterns and detect subtle deviations more effectively than conventional probabilistic methods. This study introduces the first application of EGANs to petrophysical anomaly detection, bridging deep learning with geoscience workflows. It offers a scalable framework for integrating data-driven anomaly detection into reservoir modeling, quality control, and near-real-time decision-making in drilling operations. Future work will focus on multivariate analysis, cross-basin validation, and real-time deployment, advancing toward a more intelligent, adaptive reservoir monitoring system.
{"title":"Enhanced anomaly detection in well log data through the application of ensemble GANs","authors":"Abdulrahman Al-Fakih , Ardiansyah Koeshidayatullah , Tapan Mukerji , SanLinn I. Kaka","doi":"10.1016/j.acags.2025.100316","DOIUrl":"10.1016/j.acags.2025.100316","url":null,"abstract":"<div><div>Detecting subtle anomalies in well log data can significantly enhance subsurface characterization and inform reservoir decision-making. Well logs provide detailed records of rock and fluid properties. However, identifying anomalous patterns in these data remains a persistent challenge due to the complexity of geological systems and limitations in traditional statistical models. Classical anomaly detection approaches, such as Gaussian mixture models (GMMs), tend to oversimplify the intricacies of well log responses and often misinterpret geological heterogeneities as tool noise. This misclassification leads to high false positive rates and unreliable anomaly identification. In this context, generative models, particularly generative adversarial networks (GANs), have shown promise in structured data domains; however, they remain underutilized in geoscience applications. This study aims to benchmark the performance of ensemble GANs (EGANs) against GMMs for anomaly detection in well log data. The proposed EGANs framework aggregates multiple independently trained GANs to enhance model stability and robustness. Anomalies are detected based on discriminator scoring, while performance is evaluated using precision, recall, and F1-score across four key logs: gamma ray, travel time, bulk density, and neutron porosity. The results demonstrate that EGANs consistently outperform GMMs across all logs, achieving higher precision (up to 0.70) and F1-scores (up to 0.79), with statistically significant improvements confirmed via paired t-tests. These findings highlight the ability of EGANs to model complex subsurface patterns and detect subtle deviations more effectively than conventional probabilistic methods. This study introduces the first application of EGANs to petrophysical anomaly detection, bridging deep learning with geoscience workflows. It offers a scalable framework for integrating data-driven anomaly detection into reservoir modeling, quality control, and near-real-time decision-making in drilling operations. Future work will focus on multivariate analysis, cross-basin validation, and real-time deployment, advancing toward a more intelligent, adaptive reservoir monitoring system.</div></div>","PeriodicalId":33804,"journal":{"name":"Applied Computing and Geosciences","volume":"29 ","pages":"Article 100316"},"PeriodicalIF":3.2,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145977264","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-23DOI: 10.1016/j.acags.2025.100315
Shuai Hou , Danping Cao , Zhiyu Hou , Yali Zhu , Ziqiang Wang
Precise segmentation of core images acquired by X-ray computed tomography (X-CT) is fundamental for subsequent digital rock physics analysis. However, in cores with abundant high-density minerals, insufficient X-ray penetration often leads to distorted pore structures and severe artifacts, which compromise segmentation accuracy. To address this challenge, we propose a hybrid method that synergizes Radon transform with the DINOv2 vision foundation model. Our approach employs two stage strategy: first, a ‘hard correction’ via Radon transform to leverage the directional features of artifacts and suppress directional artifact patterns in the sinogram domain; second, a “soft learning” mechanism, where the fine-tuned DINOv2 model leverages its global semantic priors to discern and rectify residual artifact interference in the feature space. This hybrid approach not only suppresses artifact interference but also preserves the continuity and visual plausibility of pore structures. Experimental results demonstrate that, on synthetic artifact-contaminated samples, the model achieves incremental three-phase segmentation performance (mIoU = 0.755, F1 = 92.15 %). While Appling to real core images, with heavy mineral artifacts, the model demonstrates robust error-correcting capabilities, yielding segmentations that align more closely with the expected rock microstructure than the original ground truth annotations. This study provides an effective and validated framework for artifact resilient digital rock segmentation, offering substantial improvements for quantitative DRP studies.
x射线计算机断层扫描（X-CT）获得的岩心图像的精确分割是后续数字岩石物理分析的基础。然而，在富含高密度矿物的岩心中，x射线穿透不足往往会导致孔隙结构畸变和严重的伪影，从而影响分割精度。为了解决这一挑战，我们提出了一种将Radon变换与DINOv2视觉基础模型协同的混合方法。我们的方法采用两阶段策略：首先，通过Radon变换进行“硬校正”，以利用伪影的方向性特征并抑制sinogram域中的方向性伪影模式；第二，“软学习”机制，其中微调的DINOv2模型利用其全局语义先验来识别和纠正特征空间中的残余人工干扰。这种混合方法不仅抑制了伪影干扰，而且保持了孔隙结构的连续性和视觉上的合理性。实验结果表明，该模型在人工制品污染样品上取得了增量式三相分割效果（mIoU = 0.755, F1 = 92.15%）。当应用于真实岩心图像时，该模型显示出强大的纠错能力，产生的分割比原始的地面真实注释更接近预期的岩石微观结构。该研究为伪影弹性数字岩石分割提供了一个有效且经过验证的框架，为定量DRP研究提供了实质性的改进。
{"title":"Synergizing Radon transform and DINOv2 for artifact-resilient digital rock segmentation","authors":"Shuai Hou , Danping Cao , Zhiyu Hou , Yali Zhu , Ziqiang Wang","doi":"10.1016/j.acags.2025.100315","DOIUrl":"10.1016/j.acags.2025.100315","url":null,"abstract":"<div><div>Precise segmentation of core images acquired by X-ray computed tomography (X-CT) is fundamental for subsequent digital rock physics analysis. However, in cores with abundant high-density minerals, insufficient X-ray penetration often leads to distorted pore structures and severe artifacts, which compromise segmentation accuracy. To address this challenge, we propose a hybrid method that synergizes Radon transform with the DINOv2 vision foundation model. Our approach employs two stage strategy: first, a ‘hard correction’ via Radon transform to leverage the directional features of artifacts and suppress directional artifact patterns in the sinogram domain; second, a “soft learning” mechanism, where the fine-tuned DINOv2 model leverages its global semantic priors to discern and rectify residual artifact interference in the feature space. This hybrid approach not only suppresses artifact interference but also preserves the continuity and visual plausibility of pore structures. Experimental results demonstrate that, on synthetic artifact-contaminated samples, the model achieves incremental three-phase segmentation performance (mIoU = 0.755, F1 = 92.15 %). While Appling to real core images, with heavy mineral artifacts, the model demonstrates robust error-correcting capabilities, yielding segmentations that align more closely with the expected rock microstructure than the original ground truth annotations. This study provides an effective and validated framework for artifact resilient digital rock segmentation, offering substantial improvements for quantitative DRP studies.</div></div>","PeriodicalId":33804,"journal":{"name":"Applied Computing and Geosciences","volume":"29 ","pages":"Article 100315"},"PeriodicalIF":3.2,"publicationDate":"2025-12-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145977261","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Recent studies have revealed the remarkable capabilities of multimodal large models (MLLMs) in general vision-and-language tasks, generating increasing interest in their application to geoscientific domains. Rock image recognition and lithological description constitute fundamental skills for geologists and represent one of the earliest application scenarios in geosciences where artificial intelligence technologies have been actively explored. As rock image recognition and description are inherently multimodal tasks, involving both visual and textual data, models are required to understand the interactions between modalities and how these interactions affect information transfer. However, the simultaneous comprehension of rock images and their corresponding lithological descriptions remains a significant challenge for current machine learning frameworks. The impact of multimodal co-training on downstream geoscientific performance is still unconfirmed. This study presents PetroMind, a domain-adapted multimodal petrographic foundation model built upon Qwen2.5-VL, together with SA-Rock, a novel LLM-based metric designed to assess the semantic accuracy of rock image generation descriptions to evaluate the capabilities of MLLMs in petrographic multimodal tasks. In the rock image classification task, PetroMind achieves performance comparable to ViT-Base-Patch16-224, with accuracy and Macro-F1 scores exceeding or approaching 97 %. This demonstrates PetroMind's strong capability in long-tailed image learning and highlights its effectiveness in substantially improving classification accuracy for few-shot rock image categories. In the lithological description generation task, PetroMind attains BLEU-4 and SA-Rock scores of 0.644 and 7.864, respectively, indicating good few-shot learning performance. The SA-Rock metric shows that the model produces highly accurate descriptions of rock structure, texture, and colour, while leaving considerable scope for improvement in the description of mineral composition. Ablation experiments further indicate that task-specific LoRA adapters are more effective for high-resource tasks such as image classification, whereas a single shared LoRA adapter demonstrates superior multi-task interaction in low-resource captioning scenarios. This study demonstrates the potential of MLLM-based architectures in jointly understanding rock images and their associated geological descriptions.
最近的研究揭示了多模态大模型（mllm）在一般视觉和语言任务中的卓越能力，引起了人们对其在地球科学领域应用的越来越多的兴趣。岩石图像识别和岩性描述是地质学家的基本技能，也是人工智能技术在地球科学中最早应用的领域之一。由于岩石图像识别和描述本质上是多模态任务，涉及视觉和文本数据，因此需要模型来理解模态之间的相互作用以及这些相互作用如何影响信息传递。然而，同时理解岩石图像及其相应的岩性描述仍然是当前机器学习框架面临的一个重大挑战。多模式协同训练对下游地球科学性能的影响尚未得到证实。本研究提出了PetroMind，一个基于Qwen2.5-VL的自适应多模态岩石学基础模型，以及SA-Rock，一个新的基于llm的度量，旨在评估岩石图像生成描述的语义准确性，以评估mllm在岩石学多模态任务中的能力。在岩石图像分类任务中，PetroMind实现了与viti - base - patch16 -224相当的性能，准确率和Macro-F1得分超过或接近97%。这证明了PetroMind在长尾图像学习方面的强大能力，并突出了其在大幅提高少量岩石图像类别分类精度方面的有效性。在岩性描述生成任务中，PetroMind的BLEU-4和SA-Rock得分分别为0.644和7.864，显示了良好的少拍学习性能。SA-Rock度量表明，该模型对岩石结构、纹理和颜色的描述非常准确，同时在矿物成分的描述方面留下了相当大的改进空间。消融实验进一步表明，特定任务的LoRA适配器在高资源任务（如图像分类）中更有效，而单个共享的LoRA适配器在低资源字幕场景中表现出更好的多任务交互。这项研究证明了基于mllm的体系结构在共同理解岩石图像及其相关地质描述方面的潜力。
{"title":"PetroMind: A multimodal petrographic model for rock image classification and lithological description generation","authors":"Zhongliang Chen , Chaojie Zheng , Mingming Zhang , Zhaoqi Hu , Jianchao Duan","doi":"10.1016/j.acags.2025.100314","DOIUrl":"10.1016/j.acags.2025.100314","url":null,"abstract":"<div><div>Recent studies have revealed the remarkable capabilities of multimodal large models (MLLMs) in general vision-and-language tasks, generating increasing interest in their application to geoscientific domains. Rock image recognition and lithological description constitute fundamental skills for geologists and represent one of the earliest application scenarios in geosciences where artificial intelligence technologies have been actively explored. As rock image recognition and description are inherently multimodal tasks, involving both visual and textual data, models are required to understand the interactions between modalities and how these interactions affect information transfer. However, the simultaneous comprehension of rock images and their corresponding lithological descriptions remains a significant challenge for current machine learning frameworks. The impact of multimodal co-training on downstream geoscientific performance is still unconfirmed. This study presents PetroMind, a domain-adapted multimodal petrographic foundation model built upon Qwen2.5-VL, together with SA-Rock, a novel LLM-based metric designed to assess the semantic accuracy of rock image generation descriptions to evaluate the capabilities of MLLMs in petrographic multimodal tasks. In the rock image classification task, PetroMind achieves performance comparable to ViT-Base-Patch16-224, with accuracy and Macro-F1 scores exceeding or approaching 97 %. This demonstrates PetroMind's strong capability in long-tailed image learning and highlights its effectiveness in substantially improving classification accuracy for few-shot rock image categories. In the lithological description generation task, PetroMind attains BLEU-4 and SA-Rock scores of 0.644 and 7.864, respectively, indicating good few-shot learning performance. The SA-Rock metric shows that the model produces highly accurate descriptions of rock structure, texture, and colour, while leaving considerable scope for improvement in the description of mineral composition. Ablation experiments further indicate that task-specific LoRA adapters are more effective for high-resource tasks such as image classification, whereas a single shared LoRA adapter demonstrates superior multi-task interaction in low-resource captioning scenarios. This study demonstrates the potential of MLLM-based architectures in jointly understanding rock images and their associated geological descriptions.</div></div>","PeriodicalId":33804,"journal":{"name":"Applied Computing and Geosciences","volume":"29 ","pages":"Article 100314"},"PeriodicalIF":3.2,"publicationDate":"2025-12-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145977530","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-03DOI: 10.1016/j.acags.2025.100310
Weilin Chen, Jiyin Zhang, Chenhao Li, Xiaogang Ma
Integrating domain knowledge into machine learning (ML) models is critical for achieving reliable and interpretable predictions in complex scientific fields such as geoscience. In several recent studies centered on the so-called Neuro-Symbolic AI (NSAI) frameworks, symbolic geological knowledge was successfully combined with traditional ML algorithms to improve the prediction of mineral deposit types. The fast development of Large Language Model (LLM) brings new opportunities to further enhance the NSAI applications. In this study, to construct the symbolic component of NSAI, we used an LLM to automatically extract, structure, and transform descriptive knowledge from authoritative geoscience textbooks into a machine-readable format. The result captures geochemical signatures, lithological settings, and alteration features associated with various mineral systems. The structured knowledge was integrated into a decision tree classifier by embedding each sample with a vectorized representation of its corresponding deposit type. Compared to conventional ML models trained solely on geochemical data, our NSAI model achieved significantly higher accuracy on the test sets, indicating improved generalization. Moreover, the NSAI model demonstrated consistent performance across a broader set of deposit types, including those with extremely limited training samples. In particular, the NSAI framework improved predictive stability and accuracy even for minority classes with only 3 to 5 samples, where traditional ML models tend to overfit or fail. This robustness underscores the value of incorporating expert-level geological knowledge into data-driven pipelines. In our result assessment, the SHAP (SHapley Additive exPlanations) analysis further revealed that symbolic knowledge vectors contributed substantially to the model's decision-making process, confirming their importance in enhancing interpretability and predictive power. Our work demonstrates that LLM-guided knowledge extraction offers an effective and scalable way to integrate structured domain knowledge into mineral prediction tasks. We hope the work can also provide insights for other geoscientific applications of NSAI.
{"title":"Enhancing neuro-symbolic AI for mineral prediction via LLM-guided knowledge integration","authors":"Weilin Chen, Jiyin Zhang, Chenhao Li, Xiaogang Ma","doi":"10.1016/j.acags.2025.100310","DOIUrl":"10.1016/j.acags.2025.100310","url":null,"abstract":"<div><div>Integrating domain knowledge into machine learning (ML) models is critical for achieving reliable and interpretable predictions in complex scientific fields such as geoscience. In several recent studies centered on the so-called Neuro-Symbolic AI (NSAI) frameworks, symbolic geological knowledge was successfully combined with traditional ML algorithms to improve the prediction of mineral deposit types. The fast development of Large Language Model (LLM) brings new opportunities to further enhance the NSAI applications. In this study, to construct the symbolic component of NSAI, we used an LLM to automatically extract, structure, and transform descriptive knowledge from authoritative geoscience textbooks into a machine-readable format. The result captures geochemical signatures, lithological settings, and alteration features associated with various mineral systems. The structured knowledge was integrated into a decision tree classifier by embedding each sample with a vectorized representation of its corresponding deposit type. Compared to conventional ML models trained solely on geochemical data, our NSAI model achieved significantly higher accuracy on the test sets, indicating improved generalization. Moreover, the NSAI model demonstrated consistent performance across a broader set of deposit types, including those with extremely limited training samples. In particular, the NSAI framework improved predictive stability and accuracy even for minority classes with only 3 to 5 samples, where traditional ML models tend to overfit or fail. This robustness underscores the value of incorporating expert-level geological knowledge into data-driven pipelines. In our result assessment, the SHAP (SHapley Additive exPlanations) analysis further revealed that symbolic knowledge vectors contributed substantially to the model's decision-making process, confirming their importance in enhancing interpretability and predictive power. Our work demonstrates that LLM-guided knowledge extraction offers an effective and scalable way to integrate structured domain knowledge into mineral prediction tasks. We hope the work can also provide insights for other geoscientific applications of NSAI.</div></div>","PeriodicalId":33804,"journal":{"name":"Applied Computing and Geosciences","volume":"29 ","pages":"Article 100310"},"PeriodicalIF":3.2,"publicationDate":"2025-12-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145705705","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-01DOI: 10.1016/j.acags.2025.100307
Jian Li, Jinchao Xing, Lujun Wei, Chuankun Li, Xiaoyan Li
Accurate separation of P- and S-waves is crucial for multi-component seismic imaging in shallow subsurface studies. We proposed a transfer learning framework based on the U-FCN architecture that combines multiscale feature fusion from U-Net with a fully convolutional neural network. Synthetic datasets were generated using finite-difference simulation and Helmholtz decomposition to train the network in a data-driven manner. For field data adaptation, we employed a transfer learning strategy involving: (1) freezing early feature extraction layers, (2) fine-tuning the final some layers, and (3) incorporating dilated convolutions to enhance feature extraction. Numerical simulations and field experiments demonstrate that the proposed approach achieves accurate P- and S-wave separation. On the field seismic data acquired from the Loess Plateau, the separated P-waves achieved R=0.952, SSIM=0.906, and PSNR=29.197 dB, while the S-waves reached R=0.938, SSIM=0.885, and PSNR=28.846 dB, with an inference time of only 1.3 s. Compared with conventional methods, our approach achieves cleaner P- and S-wave separation with significantly higher computational efficiency and without relying on prior model parameters. The results confirm the robustness and scalability of the method for real-world seismic applications, effectively bridging the gap between synthetic-based training and field data interpretation.
{"title":"Separation of P- and S-waves on shallow subsurface using transfer learning","authors":"Jian Li, Jinchao Xing, Lujun Wei, Chuankun Li, Xiaoyan Li","doi":"10.1016/j.acags.2025.100307","DOIUrl":"10.1016/j.acags.2025.100307","url":null,"abstract":"<div><div>Accurate separation of P- and S-waves is crucial for multi-component seismic imaging in shallow subsurface studies. We proposed a transfer learning framework based on the U-FCN architecture that combines multiscale feature fusion from U-Net with a fully convolutional neural network. Synthetic datasets were generated using finite-difference simulation and Helmholtz decomposition to train the network in a data-driven manner. For field data adaptation, we employed a transfer learning strategy involving: (1) freezing early feature extraction layers, (2) fine-tuning the final some layers, and (3) incorporating dilated convolutions to enhance feature extraction. Numerical simulations and field experiments demonstrate that the proposed approach achieves accurate P- and S-wave separation. On the field seismic data acquired from the Loess Plateau, the separated P-waves achieved R<span><math><msup><mrow></mrow><mrow><mn>2</mn></mrow></msup></math></span>=0.952, SSIM=0.906, and PSNR=29.197 dB, while the S-waves reached R<span><math><msup><mrow></mrow><mrow><mn>2</mn></mrow></msup></math></span>=0.938, SSIM=0.885, and PSNR=28.846 dB, with an inference time of only 1.3 s. Compared with conventional methods, our approach achieves cleaner P- and S-wave separation with significantly higher computational efficiency and without relying on prior model parameters. The results confirm the robustness and scalability of the method for real-world seismic applications, effectively bridging the gap between synthetic-based training and field data interpretation.</div></div>","PeriodicalId":33804,"journal":{"name":"Applied Computing and Geosciences","volume":"28 ","pages":"Article 100307"},"PeriodicalIF":3.2,"publicationDate":"2025-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145624028","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-01DOI: 10.1016/j.acags.2025.100309
Orhun Aydin
Groups of structural measurements based on orientation similarity are indicative of deformation mechanisms and are important measures to infer the deformation history of planetary surfaces. Despite methods for defining groups based only on orientation or spatial proximity, a general framework for defining orientation-based groups under multiple constraints is lacking. A second challenge pertains to the computational challenge of clustering large structural data due to the large volume and velocity of the data collected as a part of field-based, earth-observing, and planetary missions. In this paper, we propose a general clustering framework for defining groups in angular data based on orientation similarity that can be constrained with relational constraints, such as spatial proximity or prior knowledge of geologic units. We represent the similarity of geologic measurements with a similarity graph where similarity links (graph edges) are defined via the clustering evidence accumulated by re-clustering of data with varying parameters. We showcase the use of a spectral gap measure to define the optimal number of clusters for the evidence graph. We apply the proposed method to define groups of compaction bands using field data collected from the Valley of Fire, NV. We compare our results to a state-of-the-art Bingham mixture model. Results indicate the realism of the proposed method in terms of mapping distinct structural groups under different spatial proximity constraints.
{"title":"Graph-based evidence accumulation for clustering 3D orientation measurements in planetary surface mapping under relational constraints","authors":"Orhun Aydin","doi":"10.1016/j.acags.2025.100309","DOIUrl":"10.1016/j.acags.2025.100309","url":null,"abstract":"<div><div>Groups of structural measurements based on orientation similarity are indicative of deformation mechanisms and are important measures to infer the deformation history of planetary surfaces. Despite methods for defining groups based only on orientation or spatial proximity, a general framework for defining orientation-based groups under multiple constraints is lacking. A second challenge pertains to the computational challenge of clustering large structural data due to the large volume and velocity of the data collected as a part of field-based, earth-observing, and planetary missions. In this paper, we propose a general clustering framework for defining groups in angular data based on orientation similarity that can be constrained with relational constraints, such as spatial proximity or prior knowledge of geologic units. We represent the similarity of geologic measurements with a similarity graph where similarity links (graph edges) are defined via the clustering evidence accumulated by re-clustering of data with varying parameters. We showcase the use of a spectral gap measure to define the optimal number of clusters for the evidence graph. We apply the proposed method to define groups of compaction bands using field data collected from the Valley of Fire, NV. We compare our results to a state-of-the-art Bingham mixture model. Results indicate the realism of the proposed method in terms of mapping distinct structural groups under different spatial proximity constraints.</div></div>","PeriodicalId":33804,"journal":{"name":"Applied Computing and Geosciences","volume":"28 ","pages":"Article 100309"},"PeriodicalIF":3.2,"publicationDate":"2025-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145690945","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-01DOI: 10.1016/j.acags.2025.100308
H.N. Mahendra , V. Pushpalatha , S. Mallikarjunaswamy , S. Rama Subramoniam , Arjun Sunil Rao , N.C. Sanjay Shekar
This study presents an analysis of land use and land cover changes in Mysuru Taluk, Karnataka, India, over a two-decade period (2004–2014 and 2014–2024), using Vision Transformers (ViTs) to enhance classification accuracy. Using Linear Imaging Self-Scanning Sensor (LISS-III) remote sensing data, our approach combines the powerful feature extraction capabilities of ViTs to address the complexities inherent in multi-temporal satellite data. Traditional methods for LULC mapping face challenges due to variability in land features and temporal changes, which impact classification accuracy. By employing ViTs, we aim to overcome these limitations through their self-attention approach, which can capture long-range dependencies in the data, thus offering a more refined classification process. Our study results show improved overall classification accuracy across the assessed years, achieving 95.07 % in 2004, 95.79 % in 2014, and reaching 96.74 % in 2024. These progressive results highlight the efficiency of ViTs in accurately classifying and detecting subtle land cover changes over time. Further, change detection analysis results show that the built-up area increased by 17.25 %, and agricultural land decreased by 16.24 % over two decades. The findings will assist policymakers and urban planners develop strategies to manage urbanization effectively while minimizing environmental impacts.
{"title":"Analyzing land use land cover changes in Mysuru taluk, Karnataka state, India using vision transformers","authors":"H.N. Mahendra , V. Pushpalatha , S. Mallikarjunaswamy , S. Rama Subramoniam , Arjun Sunil Rao , N.C. Sanjay Shekar","doi":"10.1016/j.acags.2025.100308","DOIUrl":"10.1016/j.acags.2025.100308","url":null,"abstract":"<div><div>This study presents an analysis of land use and land cover changes in Mysuru Taluk, Karnataka, India, over a two-decade period (2004–2014 and 2014–2024), using Vision Transformers (ViTs) to enhance classification accuracy. Using Linear Imaging Self-Scanning Sensor (LISS-III) remote sensing data, our approach combines the powerful feature extraction capabilities of ViTs to address the complexities inherent in multi-temporal satellite data. Traditional methods for LULC mapping face challenges due to variability in land features and temporal changes, which impact classification accuracy. By employing ViTs, we aim to overcome these limitations through their self-attention approach, which can capture long-range dependencies in the data, thus offering a more refined classification process. Our study results show improved overall classification accuracy across the assessed years, achieving 95.07 % in 2004, 95.79 % in 2014, and reaching 96.74 % in 2024. These progressive results highlight the efficiency of ViTs in accurately classifying and detecting subtle land cover changes over time. Further, change detection analysis results show that the built-up area increased by 17.25 %, and agricultural land decreased by 16.24 % over two decades. The findings will assist policymakers and urban planners develop strategies to manage urbanization effectively while minimizing environmental impacts.</div></div>","PeriodicalId":33804,"journal":{"name":"Applied Computing and Geosciences","volume":"28 ","pages":"Article 100308"},"PeriodicalIF":3.2,"publicationDate":"2025-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145690943","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号