Pub Date : 2026-01-07DOI: 10.1016/j.cageo.2025.106101
Lei Liu, Chao Song, Liangsheng He, Silin Wang, Xuan Feng, Cai Liu
This study presents a fast and flexible three-dimensional dual-parameter full waveform inversion (FWI) framework for ground penetrating radar (GPR), enabled by a hybrid compilation strategy that integrates custom CUDA kernel functions with the PyTorch automatic differentiation ecosystem. In the proposed workflow, computationally intensive operations are executed by highly optimized CUDA kernels, while PyTorch is employed only for lightweight tasks. This selective integration substantially reduces memory usage and avoids the runtime bottlenecks often encountered in GPR FWI, achieving an effective balance between efficiency and algorithmic adaptability. The framework supports simultaneous inversion of relative permittivity and electrical conductivity in large-scale 3D domains, providing a practical solution for multi-parameter GPR imaging. Its modular Python-based architecture further allows users to easily customize loss functions, regularization schemes, and optimization settings without modifying code, making the method suitable for rapid prototyping and methodological development. Numerical experiments on 2D and 3D models demonstrate excellent scalability and stable reconstruction performance, while a field-data example confirms that the method can reliably detect subsurface anomalies even under challenging zero-offset acquisition conditions. Overall, the proposed CUDA-PyTorch hybrid framework advances the state of the art in GPR FWI by combining high-performance GPU computing with the flexibility of modern deep-learning toolchains, offering a practical and extensible platform for future GPR FWI research.
{"title":"Fast ground penetrating radar dual-parameter full waveform inversion method accelerated by hybrid compilation of CUDA kernel function and PyTorch","authors":"Lei Liu, Chao Song, Liangsheng He, Silin Wang, Xuan Feng, Cai Liu","doi":"10.1016/j.cageo.2025.106101","DOIUrl":"10.1016/j.cageo.2025.106101","url":null,"abstract":"<div><div>This study presents a fast and flexible three-dimensional dual-parameter full waveform inversion (FWI) framework for ground penetrating radar (GPR), enabled by a hybrid compilation strategy that integrates custom CUDA kernel functions with the PyTorch automatic differentiation ecosystem. In the proposed workflow, computationally intensive operations are executed by highly optimized CUDA kernels, while PyTorch is employed only for lightweight tasks. This selective integration substantially reduces memory usage and avoids the runtime bottlenecks often encountered in GPR FWI, achieving an effective balance between efficiency and algorithmic adaptability. The framework supports simultaneous inversion of relative permittivity and electrical conductivity in large-scale 3D domains, providing a practical solution for multi-parameter GPR imaging. Its modular Python-based architecture further allows users to easily customize loss functions, regularization schemes, and optimization settings without modifying code, making the method suitable for rapid prototyping and methodological development. Numerical experiments on 2D and 3D models demonstrate excellent scalability and stable reconstruction performance, while a field-data example confirms that the method can reliably detect subsurface anomalies even under challenging zero-offset acquisition conditions. Overall, the proposed CUDA-PyTorch hybrid framework advances the state of the art in GPR FWI by combining high-performance GPU computing with the flexibility of modern deep-learning toolchains, offering a practical and extensible platform for future GPR FWI research.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"209 ","pages":"Article 106101"},"PeriodicalIF":4.4,"publicationDate":"2026-01-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146038649","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-06DOI: 10.1016/j.cageo.2026.106107
Siyi Qiu , Zhen Ye , Yusheng Xu , Jia Qian , Rong Huang , Tao Tao , Huan Xie , Xiaohua Tong
A good lunar Digital Elevation Model (DEM) is critical for successful lunar landings and exploration missions, providing essential terrain data for landing site selection and mission planning. However, existing lunar DEMs face significant limitations, such as insufficient resolution, incomplete data coverage, and errors arising from observational and processing methods. These limitations hinder their effectiveness in high-precision terrain analysis and practical applications. To address these challenges, this study proposes an adaptive regularized variational model with curvature smoothing constraints, aiming to robustly fuse multi-source DEMs with varying resolutions, noise levels, and data voids. The proposed model incorporates two key innovations: a weighted data fidelity term that dynamically adjusts based on slope anomaly detection and ensures accuracy consistency across multi-scale DEMs, and a curvature-constrained total variation regularization term that suppresses noise and artifacts while preserving terrain details. The variational problem is solved iteratively using the Alternating Direction Method of Multipliers, enabling the seamless integration of multi-source DEMs. Experimental validations on simulated and real lunar south pole datasets demonstrate the superior performance of the proposed method, achieving significant improvements in noise suppression, artifact removal, void filling, and terrain detail preservation. The results indicate that the proposed method is capable of generating high-quality, seamless DEMs with enhanced spatial consistency and accuracy.
{"title":"Adaptive variational fusion of multi-source lunar digital elevation models based on curvature regularization","authors":"Siyi Qiu , Zhen Ye , Yusheng Xu , Jia Qian , Rong Huang , Tao Tao , Huan Xie , Xiaohua Tong","doi":"10.1016/j.cageo.2026.106107","DOIUrl":"10.1016/j.cageo.2026.106107","url":null,"abstract":"<div><div>A good lunar Digital Elevation Model (DEM) is critical for successful lunar landings and exploration missions, providing essential terrain data for landing site selection and mission planning. However, existing lunar DEMs face significant limitations, such as insufficient resolution, incomplete data coverage, and errors arising from observational and processing methods. These limitations hinder their effectiveness in high-precision terrain analysis and practical applications. To address these challenges, this study proposes an adaptive regularized variational model with curvature smoothing constraints, aiming to robustly fuse multi-source DEMs with varying resolutions, noise levels, and data voids. The proposed model incorporates two key innovations: a weighted data fidelity term that dynamically adjusts based on slope anomaly detection and ensures accuracy consistency across multi-scale DEMs, and a curvature-constrained total variation regularization term that suppresses noise and artifacts while preserving terrain details. The variational problem is solved iteratively using the Alternating Direction Method of Multipliers, enabling the seamless integration of multi-source DEMs. Experimental validations on simulated and real lunar south pole datasets demonstrate the superior performance of the proposed method, achieving significant improvements in noise suppression, artifact removal, void filling, and terrain detail preservation. The results indicate that the proposed method is capable of generating high-quality, seamless DEMs with enhanced spatial consistency and accuracy.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"209 ","pages":"Article 106107"},"PeriodicalIF":4.4,"publicationDate":"2026-01-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980490","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-05DOI: 10.1016/j.cageo.2026.106106
Jaesung Park , Jina Jeong , Eun-Jung Holden
Accurate reconstruction of missing well-log data is essential for subsurface characterization and reservoir modeling but remains challenging under conditions of stratigraphic heterogeneity and multi-log incompleteness. This study introduces a deep learning framework that enhance missing log prediction by embedding lithological information as a contextual constraint. The proposed framework integrates a Conditional Variational Autoencoder (CVAE) with a Long Short-Term Memory (LSTM)-based lithology predictor, namely the Iterative Lithology-Constrained Hybrid CVAE–LSTM Network (ILCH-Net), in an iterative refinement process. The model was trained and validated on 45,809 samples from six wells in the Volve oil field, Norwegian North Sea, comprising five commonly acquired logs (GR, RHOB, NPHI, DTC, DTS) across three lithologies (claystone, sandstone, limestone). Quantitative evaluation demonstrates that ILCH-Net surpasses baseline approaches (Autoencoder, Iteratively refined autoencoder, LSTM), achieving lower root mean squared error (10.43 vs. 12.84 for LSTM) and improved distributional similarity (median Kolmogorov–Smirnov statistic of 0.15 with an interquartile range of 0.09 across the six test wells). Lithology-specific analysis further shows that reconstruction accuracy is highest for limestone and claystone, reflecting their distinct well-log responses, while sandstone exhibits greater variability due to depth-dependent compaction effects. These results confirm that lithological constraints not only enhance accuracy but also reduce inter-well variability, thereby yielding geologically consistent reconstructions. By embedding geological priors within a data-driven framework, ILCH-Net provides a robust and scalable solution for applications in reservoir characterization, digital rock modeling, and geomechanical analysis where incomplete or irregular logs are prevalent.
准确重建缺失的测井数据对于地下表征和储层建模至关重要，但在地层非均质性和多测井不完整的情况下仍然具有挑战性。本研究引入了一个深度学习框架，通过嵌入岩性信息作为上下文约束来增强缺失测井预测。该框架将条件变分自编码器（CVAE）与基于长短期记忆（LSTM）的岩性预测器（即迭代岩性约束混合CVAE - LSTM网络（ILCH-Net））集成在迭代改进过程中。该模型在挪威北海Volve油田6口井的45809个样本上进行了训练和验证，包括5种常用测井（GR、RHOB、NPHI、DTC、DTS），涵盖3种岩性（粘土岩、砂岩、石灰岩）。定量评价表明，ILCH-Net优于基线方法（Autoencoder、iterative refined Autoencoder、LSTM），实现了更低的均方根误差（10.43 vs. 12.84 LSTM），并改善了分布相似性（6口测试井的Kolmogorov-Smirnov统计量中位数为0.15，四分位数区间为0.09）。岩性分析进一步表明，石灰岩和粘土岩的重建精度最高，反映了它们独特的测井响应，而砂岩由于深度相关的压实作用而表现出更大的变异性。这些结果证实，岩性约束不仅提高了精度，还减少了井间的变异性，从而产生了地质一致性的重建。通过将地质先验嵌入到数据驱动的框架中，ILCH-Net为储层表征、数字岩石建模和地质力学分析等应用提供了强大且可扩展的解决方案，这些应用普遍存在测井不完整或不规则的情况。
{"title":"Deep-learning-based hybrid model with iterative lithology constraints for the enhanced prediction of missing well-logs","authors":"Jaesung Park , Jina Jeong , Eun-Jung Holden","doi":"10.1016/j.cageo.2026.106106","DOIUrl":"10.1016/j.cageo.2026.106106","url":null,"abstract":"<div><div>Accurate reconstruction of missing well-log data is essential for subsurface characterization and reservoir modeling but remains challenging under conditions of stratigraphic heterogeneity and multi-log incompleteness. This study introduces a deep learning framework that enhance missing log prediction by embedding lithological information as a contextual constraint. The proposed framework integrates a Conditional Variational Autoencoder (CVAE) with a Long Short-Term Memory (LSTM)-based lithology predictor, namely the Iterative Lithology-Constrained Hybrid CVAE–LSTM Network (ILCH-Net), in an iterative refinement process. The model was trained and validated on 45,809 samples from six wells in the Volve oil field, Norwegian North Sea, comprising five commonly acquired logs (GR, RHOB, NPHI, DTC, DTS) across three lithologies (claystone, sandstone, limestone). Quantitative evaluation demonstrates that ILCH-Net surpasses baseline approaches (Autoencoder, Iteratively refined autoencoder, LSTM), achieving lower root mean squared error (10.43 vs. 12.84 for LSTM) and improved distributional similarity (median Kolmogorov–Smirnov statistic of 0.15 with an interquartile range of 0.09 across the six test wells). Lithology-specific analysis further shows that reconstruction accuracy is highest for limestone and claystone, reflecting their distinct well-log responses, while sandstone exhibits greater variability due to depth-dependent compaction effects. These results confirm that lithological constraints not only enhance accuracy but also reduce inter-well variability, thereby yielding geologically consistent reconstructions. By embedding geological priors within a data-driven framework, ILCH-Net provides a robust and scalable solution for applications in reservoir characterization, digital rock modeling, and geomechanical analysis where incomplete or irregular logs are prevalent.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"209 ","pages":"Article 106106"},"PeriodicalIF":4.4,"publicationDate":"2026-01-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928790","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Scientific workflows are essential in modern geoscientific research, where complex, multi-stage computational pipelines are used to analyze heterogeneous environmental data. Ensuring the reproducibility and traceability of these workflows is critical but often challenging due to their intricacy and evolving structure. We introduce EWoPe (Embeddable Workflow Persistence), a lightweight and embeddable methodology and C++ library designed to persist and reconstruct scientific workflows over time. Unlike existing workflow systems and provenance tools, our method adds minimal overhead to workflow execution: it does not automate or optimize processes, but instead ensures persistence through lightweight and structured metadata. EWoPe models workflows as directed acyclic graphs (DAGs), in which each data node is linked to computational tasks through metadata. The latter captures input–output dependencies, algorithm parameters, execution commands, and intermediate results, supporting full traceability and reproducibility of computational histories. EWoPe offers dual usability: as a standalone command-line tool or as an embeddable component within larger applications. We show its flexibility and applicability through a case study involving a complex workflow leading to subsurface reaction-transport modeling, starting from boreholes data. The EWoPe library is publicly available and designed to be extensible, making it suitable for a wide range of scientific domains, including geochemistry, geophysics, environmental engineering, and any other fields where transparency and data integrity are critical.
{"title":"EWoPe: A light Embeddable WOrkflow PErsistence tool for geoscientific pipeline reproducibility","authors":"Marianna Miola , Daniela Cabiddu , Simone Pittaluga , Micaela Raviola , Marino Vetuschi Zuccolini","doi":"10.1016/j.cageo.2025.106099","DOIUrl":"10.1016/j.cageo.2025.106099","url":null,"abstract":"<div><div>Scientific workflows are essential in modern geoscientific research, where complex, multi-stage computational pipelines are used to analyze heterogeneous environmental data. Ensuring the reproducibility and traceability of these workflows is critical but often challenging due to their intricacy and evolving structure. We introduce <span>EWoPe</span> (Embeddable Workflow Persistence), a lightweight and embeddable methodology and C++ library designed to persist and reconstruct scientific workflows over time. Unlike existing workflow systems and provenance tools, our method adds minimal overhead to workflow execution: it does not automate or optimize processes, but instead ensures persistence through lightweight and structured metadata. <span>EWoPe</span> models workflows as directed acyclic graphs (DAGs), in which each data node is linked to computational tasks through metadata. The latter captures input–output dependencies, algorithm parameters, execution commands, and intermediate results, supporting full traceability and reproducibility of computational histories. <span>EWoPe</span> offers dual usability: as a standalone command-line tool or as an embeddable component within larger applications. We show its flexibility and applicability through a case study involving a complex workflow leading to subsurface reaction-transport modeling, starting from boreholes data. The <span>EWoPe</span> library is publicly available and designed to be extensible, making it suitable for a wide range of scientific domains, including geochemistry, geophysics, environmental engineering, and any other fields where transparency and data integrity are critical.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"209 ","pages":"Article 106099"},"PeriodicalIF":4.4,"publicationDate":"2026-01-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145908904","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-03DOI: 10.1016/j.cageo.2025.106103
Daniel Siervo, Yangkang Chen
Most earthquake detection workflows are based on an optimized short-term-average/long-term-average (STA/LTA) ratio, especially in regions with relatively sparse station geometry and poor velocity models. As the magnitude threshold is lowered to enable more complete earthquake analysis, more false alarms are occurring daily in earthquake monitoring. Here, we propose a high-fidelity approach, called hybrid shallow and deep learning (HSDL), to automatically classify potential earthquakes detected by an optimized STA/LTA workflow as true positives or false positives. To facilitate classification, we leverage an advanced deep learning phase picker, the earthquake compact convolutional transformer (EQCCT), which provides several classification features. These features include the counts of P&S picks, the average, minimum, maximum, and standard deviation of P&S probabilities, and the S/P pick count ratios. On a moderate dataset containing 200 real earthquakes and 200 fake earthquake waveforms, we achieve 100% accuracy across all metrics for both the random forest and XGBoost methods. On a larger dataset of 1500 events, we still achieve a precision of 1.0, a recall above 0.99, and an F1 score above 0.99 for both the random forest and XGBoost methods, with XGBoost achieving slightly higher accuracy. We also analyzed the feature importance and found that the maximum S-pick probability and the S/P pick count ratio play the most critical roles in classification. The proposed method provides a highly effective and efficient approach for fine-tuning the automatic earthquake catalog using the optimized STA/LTA method, leveraging existing tools such as the deep-learning-based phase picker and XGBoost.
{"title":"HSDL: A novel and practical method to refine automatic earthquake catalog using hybrid shallow and deep learning","authors":"Daniel Siervo, Yangkang Chen","doi":"10.1016/j.cageo.2025.106103","DOIUrl":"10.1016/j.cageo.2025.106103","url":null,"abstract":"<div><div>Most earthquake detection workflows are based on an optimized short-term-average/long-term-average (STA/LTA) ratio, especially in regions with relatively sparse station geometry and poor velocity models. As the magnitude threshold is lowered to enable more complete earthquake analysis, more false alarms are occurring daily in earthquake monitoring. Here, we propose a high-fidelity approach, called hybrid shallow and deep learning (HSDL), to automatically classify potential earthquakes detected by an optimized STA/LTA workflow as true positives or false positives. To facilitate classification, we leverage an advanced deep learning phase picker, the earthquake compact convolutional transformer (EQCCT), which provides several classification features. These features include the counts of P&S picks, the average, minimum, maximum, and standard deviation of P&S probabilities, and the S/P pick count ratios. On a moderate dataset containing 200 real earthquakes and 200 fake earthquake waveforms, we achieve 100% accuracy across all metrics for both the random forest and XGBoost methods. On a larger dataset of 1500 events, we still achieve a precision of 1.0, a recall above 0.99, and an F1 score above 0.99 for both the random forest and XGBoost methods, with XGBoost achieving slightly higher accuracy. We also analyzed the feature importance and found that the maximum S-pick probability and the S/P pick count ratio play the most critical roles in classification. The proposed method provides a highly effective and efficient approach for fine-tuning the automatic earthquake catalog using the optimized STA/LTA method, leveraging existing tools such as the deep-learning-based phase picker and XGBoost.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"209 ","pages":"Article 106103"},"PeriodicalIF":4.4,"publicationDate":"2026-01-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928789","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-29DOI: 10.1016/j.cageo.2025.106100
Bohan Chen , Yan Zhang , Liangliang Yao , Haichao Wang , Qifeng Chen , Miaomiao Wang
The finite-difference time-domain (FDTD) method incurs significant computational overhead in high-precision, long sampling, and large-scale elastic wave forward numerical simulations, due to its inherent drawbacks, such as the need for repeated wave field iteration and the large number of grid points. To address this problem, this paper proposes a 1D convolution operator elastic wave forward acceleration method based on FDTD theory, the staggered grids, and the first-order velocity-stress equation while maintaining the original wave field iteration and grid point count. The method takes the finite difference coefficients as the 1D convolution kernel weights, transforms the finite difference partial derivative computation into the form of convolution operation, and makes full use of the high arithmetic intensity and parallel characteristics of convolution computation in GPUs to achieve efficient solving of spatial first-order derivatives. The matrix transposition optimization strategy is introduced to reorganise the storage layout of column direction data to improve the efficiency of reading column direction data and maximize the performance of the convolution operation. At the same time, a parallel matrix multiplication mechanism is designed to further improve the performance of convolutional computation. The proposed method achieves comparable numerical simulation accuracy to the FDTD method of the same order. It show a time efficiency improvement of 62.82 % in high-precision imaging, 58.48 % in long sampling, and 44.03 % in large-scale models.
{"title":"An acceleration method for elastic wave forward modeling based on 1D convolution operator","authors":"Bohan Chen , Yan Zhang , Liangliang Yao , Haichao Wang , Qifeng Chen , Miaomiao Wang","doi":"10.1016/j.cageo.2025.106100","DOIUrl":"10.1016/j.cageo.2025.106100","url":null,"abstract":"<div><div>The finite-difference time-domain (FDTD) method incurs significant computational overhead in high-precision, long sampling, and large-scale elastic wave forward numerical simulations, due to its inherent drawbacks, such as the need for repeated wave field iteration and the large number of grid points. To address this problem, this paper proposes a 1D convolution operator elastic wave forward acceleration method based on FDTD theory, the staggered grids, and the first-order velocity-stress equation while maintaining the original wave field iteration and grid point count. The method takes the finite difference coefficients as the 1D convolution kernel weights, transforms the finite difference partial derivative computation into the form of convolution operation, and makes full use of the high arithmetic intensity and parallel characteristics of convolution computation in GPUs to achieve efficient solving of spatial first-order derivatives. The matrix transposition optimization strategy is introduced to reorganise the storage layout of column direction data to improve the efficiency of reading column direction data and maximize the performance of the convolution operation. At the same time, a parallel matrix multiplication mechanism is designed to further improve the performance of convolutional computation. The proposed method achieves comparable numerical simulation accuracy to the FDTD method of the same order. It show a time efficiency improvement of 62.82 % in high-precision imaging, 58.48 % in long sampling, and 44.03 % in large-scale models.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"208 ","pages":"Article 106100"},"PeriodicalIF":4.4,"publicationDate":"2025-12-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145885238","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-21DOI: 10.1016/j.cageo.2025.106098
Yehoon Kim , Ho-rim Kim , Heewon Jung
The computational challenge of predicting reactive transport in heterogeneous porous media originates from resolving complex pore-scale geometries and sharp concentration gradients near solid-fluid interfaces. This study introduces PRT-DeepONet (Pore-scale Reaction Transport Deep Operator Network), a geometry-aware neural operator that predicts local concentration fields in porous media with linear and nonlinear reactions. The architecture comprises two branch networks—a CNN branch for encoding the spatial patterns of binary porous media and an FNN branch for extracting parametric controls in partial differential equations—and a trunk network augmented with a geodesic distance function. This geodesic encoding addresses the inherent limitation of convolutional neural networks in maintaining geometric fidelity at solid-fluid interfaces by providing explicit transport pathway information absent in conventional architectures. PRT-DeepONet was trained on lattice Boltzmann simulations spanning various Péclet and Damköhler numbers, encompassing diffusion-to advection-dominated transport and slow to fast reaction kinetics. For steady-state predictions, PRT-DeepONet achieves an average RMSE below 0.04 while preserving complex grain geometries that baseline models fail to capture, with computational speedups of 3–5 orders of magnitude—reducing simulation times from minutes to milliseconds. The architecture successfully extends to transient problems, accurately predicting temporal concentration evolution for both reversible sorption and Monod kinetics. PRT-DeepONet demonstrates robust interpolation for unseen parametric conditions and time points, with performance improving with denser sampling of training data. These capabilities position PRT-DeepONet as an efficient tool for subsurface applications requiring rapid evaluation of reactive transport, including groundwater contamination assessment, CO2 sequestration modeling, and nuclear waste disposal safety analysis.
{"title":"PRT-DeepONet: Geometry-aware neural operator for efficient prediction of pore-scale concentration fields","authors":"Yehoon Kim , Ho-rim Kim , Heewon Jung","doi":"10.1016/j.cageo.2025.106098","DOIUrl":"10.1016/j.cageo.2025.106098","url":null,"abstract":"<div><div>The computational challenge of predicting reactive transport in heterogeneous porous media originates from resolving complex pore-scale geometries and sharp concentration gradients near solid-fluid interfaces. This study introduces PRT-DeepONet (Pore-scale Reaction Transport Deep Operator Network), a geometry-aware neural operator that predicts local concentration fields in porous media with linear and nonlinear reactions. The architecture comprises two branch networks—a CNN branch for encoding the spatial patterns of binary porous media and an FNN branch for extracting parametric controls in partial differential equations—and a trunk network augmented with a geodesic distance function. This geodesic encoding addresses the inherent limitation of convolutional neural networks in maintaining geometric fidelity at solid-fluid interfaces by providing explicit transport pathway information absent in conventional architectures. PRT-DeepONet was trained on lattice Boltzmann simulations spanning various Péclet and Damköhler numbers, encompassing diffusion-to advection-dominated transport and slow to fast reaction kinetics. For steady-state predictions, PRT-DeepONet achieves an average RMSE below 0.04 while preserving complex grain geometries that baseline models fail to capture, with computational speedups of 3–5 orders of magnitude—reducing simulation times from minutes to milliseconds. The architecture successfully extends to transient problems, accurately predicting temporal concentration evolution for both reversible sorption and Monod kinetics. PRT-DeepONet demonstrates robust interpolation for unseen parametric conditions and time points, with performance improving with denser sampling of training data. These capabilities position PRT-DeepONet as an efficient tool for subsurface applications requiring rapid evaluation of reactive transport, including groundwater contamination assessment, CO<sub>2</sub> sequestration modeling, and nuclear waste disposal safety analysis.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"208 ","pages":"Article 106098"},"PeriodicalIF":4.4,"publicationDate":"2025-12-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145885240","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-20DOI: 10.1016/j.cageo.2025.106097
Hongliang Li , Zhewen Xu , Zelong Fang , Nong Zhang , Changzheng Liu , Renge Zhou , Xiaohui Wei
The Madden–Julian Oscillation (MJO) represents the predominant driver of sub-seasonal variability within tropical regions. In deep learning of weather forecasting, achieving reliable accuracy in MJO prediction remains challenging, making sub-seasonal forecasts generally probabilistic. In this paper, we reveal the challenges that impede MJO forecasts, including distributional drift when the circulation passes from the Indian Ocean across Australia, hindered by obstacles like lands or shoals. In addition, we find it non-trivial to extract the sophisticated spatio-temporal relationships in climate data. To address these issues, we propose MJOFormer, an adaptive land-ocean spatio-temporal transformer, with (1) a land-ocean sampler to address distributional drift by adaptively partitioning across terrain; (2) a dynamic attention mechanism to compensate for the absence of spatial features by adaptively tackling the spatio-temporal correlation; (3) a cuboid method to improve efficiency by parallel training. Comprehensive experiments exhibit that MJOFormer possesses competitive, outperforming existing methods with better accuracy, stability, and efficiency.
{"title":"MJOFormer: An adaptive land-ocean spatio-temporal transformer for Madden–Julian Oscillation forecasting","authors":"Hongliang Li , Zhewen Xu , Zelong Fang , Nong Zhang , Changzheng Liu , Renge Zhou , Xiaohui Wei","doi":"10.1016/j.cageo.2025.106097","DOIUrl":"10.1016/j.cageo.2025.106097","url":null,"abstract":"<div><div>The Madden–Julian Oscillation (MJO) represents the predominant driver of sub-seasonal variability within tropical regions. In deep learning of weather forecasting, achieving reliable accuracy in MJO prediction remains challenging, making sub-seasonal forecasts generally probabilistic. In this paper, we reveal the challenges that impede MJO forecasts, including distributional drift when the circulation passes from the Indian Ocean across Australia, hindered by obstacles like lands or shoals. In addition, we find it non-trivial to extract the sophisticated spatio-temporal relationships in climate data. To address these issues, we propose MJOFormer, an adaptive land-ocean spatio-temporal transformer, with (1) a land-ocean sampler to address distributional drift by adaptively partitioning across terrain; (2) a dynamic attention mechanism to compensate for the absence of spatial features by adaptively tackling the spatio-temporal correlation; (3) a cuboid method to improve efficiency by parallel training. Comprehensive experiments exhibit that MJOFormer possesses competitive, outperforming existing methods with better accuracy, stability, and efficiency.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"208 ","pages":"Article 106097"},"PeriodicalIF":4.4,"publicationDate":"2025-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145885239","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-13DOI: 10.1016/j.cageo.2025.106095
Zifa Wang , Jinfeng Dai , Dengke Zhao , Xiangying Wang , Jianming Wang , Zhaoyan Li , Yongcheng Feng , Zhaodong Wang
Ground motion prediction is central to earthquake engineering and disaster assessment, but traditional ground motion prediction models (GMPMs) struggle to capture the complex nature of seismic wave propagation. GMPMs based on a single machine learning algorithm also exhibit unsatisfactory performance when handling high-dimensional nonlinear relationships and large datasets. This study proposes a novel ground motion prediction model, MoE-XGB, which combines the Mixture of Experts (MoE) architecture with the XGBoost algorithm. Through a gating network, it dynamically assigns weights to adaptively handle the heterogeneity of earthquake data. The model innovatively integrates latitude and longitude features of stations and seismic sources, primarily acting as proxies for relative positions between stations and epicenters. The model was trained and validated using a strong-motion database of shallow crustal earthquakes in Japan and tested for cross-regional generalization with a New Zealand earthquake dataset. Results show that the MoE-XGB model, trained on a 1997–2019 strong-motion dataset, improves the mean squared error (MSE) by 39.2 %, reduces the standard deviation by 22.0 %, and increases the correlation coefficient by 4.8 % compared to the XGBoost-SC model, which is a regression model based on XGBoost and specifically designed for predicting seismic motions in the shallow crust region (SC). The inclusion of latitude and longitude features, primarily acting as proxies for relative positions between stations and epicenters, significantly enhances prediction accuracy. Cross-regional testing in New Zealand confirms the model's robust generalization to earthquake events in other regions. By efficiently integrating spatial features and a dynamic expert mechanism, the MoE-XGB model provides a high-precision, highly generalizable solution for ground motion prediction.
{"title":"A mixture of experts model for shallow crustal earthquake ground motion prediction in Japan","authors":"Zifa Wang , Jinfeng Dai , Dengke Zhao , Xiangying Wang , Jianming Wang , Zhaoyan Li , Yongcheng Feng , Zhaodong Wang","doi":"10.1016/j.cageo.2025.106095","DOIUrl":"10.1016/j.cageo.2025.106095","url":null,"abstract":"<div><div>Ground motion prediction is central to earthquake engineering and disaster assessment, but traditional ground motion prediction models (GMPMs) struggle to capture the complex nature of seismic wave propagation. GMPMs based on a single machine learning algorithm also exhibit unsatisfactory performance when handling high-dimensional nonlinear relationships and large datasets. This study proposes a novel ground motion prediction model, MoE-XGB, which combines the Mixture of Experts (MoE) architecture with the XGBoost algorithm. Through a gating network, it dynamically assigns weights to adaptively handle the heterogeneity of earthquake data. The model innovatively integrates latitude and longitude features of stations and seismic sources, primarily acting as proxies for relative positions between stations and epicenters. The model was trained and validated using a strong-motion database of shallow crustal earthquakes in Japan and tested for cross-regional generalization with a New Zealand earthquake dataset. Results show that the MoE-XGB model, trained on a 1997<strong>–</strong>2019 strong-motion dataset, improves the mean squared error (MSE) by 39.2 %, reduces the standard deviation by 22.0 %, and increases the correlation coefficient by 4.8 % compared to the XGBoost-SC model, which is a regression model based on XGBoost and specifically designed for predicting seismic motions in the shallow crust region (SC). The inclusion of latitude and longitude features, primarily acting as proxies for relative positions between stations and epicenters, significantly enhances prediction accuracy. Cross-regional testing in New Zealand confirms the model's robust generalization to earthquake events in other regions. By efficiently integrating spatial features and a dynamic expert mechanism, the MoE-XGB model provides a high-precision, highly generalizable solution for ground motion prediction.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"208 ","pages":"Article 106095"},"PeriodicalIF":4.4,"publicationDate":"2025-12-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145791802","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-13DOI: 10.1016/j.cageo.2025.106096
Samuel C. Boone , Ling Chung , Noel Faux , Usha Nattala , Thomas Church , Chenghao Jiang , Malcolm McMillan , Sean Jones , David Liu , Han Jiang , Kris Ehinger , Tom Drummond , Barry Kohn , Andrew Gleadow
Apatite fission-track (FT) thermochronology is widely used for constraining the thermal evolution of crustal rocks. However, manual FT identification is time-intensive and subjective. Although recent AI-based approaches have shown promise, performance often declines for complex, natural samples due to limited and overly idealised training data.
We introduce two open-access convolutional neural networks (CNNs) for automatic detection of surface-intersecting FTs in apatite and mica. The first, HALtracks 2D, uses paired reflected- and transmitted-light surface images, while HALtracks 3D incorporates an additional 3D stack of transmitted light images. HALtracks 2D exhibits mean accuracies (94.2–91.6 %) that are as good or better than both HALtracks 3D and all previous FT algorithms across a broad range of apatite fission track densities (up to 8.54 × 106 tracks/cm2) on expert-curated reference data. This improvement is due to a comprehensive training dataset comprising a wider range of track densities and etch-pit morphologies.
Unexpectedly, HALtracks3D performed worse (91.5–80.1 %), likely because reflected-light information—critical for recognising track openings—became underrepresented among multiple transmitted-light inputs during CNN training. At very high track densities (>8.54 × 106 tracks/cm2) pushing the analytical boundaries of optical fission-track counting, Coincidence Mapping (Gleadow et al., 2009) remains more accurate than HALtracks 2D. Thermochronologists might therefore consider utilising a combination of automated fission-track algorithms depending on FT density.
Future work could expand the open-access training dataset to include a broader range of apatite FT specimens, and increased metadata for targeted CNN training on spurious features such as surface imperfections and dislocations, which are misidentified as fission tracks by existing algorithms. The open-access testing dataset presented here provides a benchmark for evaluating future FT algorithms.
Nevertheless, HALtracks 2D's enhanced accuracy brings apatite FT analysis significantly closer to full automation, with the potential to mitigate observer bias, reduce inter-laboratory variability, and broaden the accessibility of the technique to the wider geoscience community.
{"title":"Raising the bar: Deep learning on comprehensive database sets new benchmark for automated fission-track detection","authors":"Samuel C. Boone , Ling Chung , Noel Faux , Usha Nattala , Thomas Church , Chenghao Jiang , Malcolm McMillan , Sean Jones , David Liu , Han Jiang , Kris Ehinger , Tom Drummond , Barry Kohn , Andrew Gleadow","doi":"10.1016/j.cageo.2025.106096","DOIUrl":"10.1016/j.cageo.2025.106096","url":null,"abstract":"<div><div>Apatite fission-track (FT) thermochronology is widely used for constraining the thermal evolution of crustal rocks. However, manual FT identification is time-intensive and subjective. Although recent AI-based approaches have shown promise, performance often declines for complex, natural samples due to limited and overly idealised training data.</div><div>We introduce two open-access convolutional neural networks (CNNs) for automatic detection of surface-intersecting FTs in apatite and mica. The first, <em>HALtracks 2D</em>, uses paired reflected- and transmitted-light surface images, while <em>HALtracks 3D</em> incorporates an additional 3D stack of transmitted light images. <em>HALtracks 2D</em> exhibits mean accuracies (94.2–91.6 %) that are as good or better than both <em>HALtracks 3D</em> and all previous FT algorithms across a broad range of apatite fission track densities (up to 8.54 × 10<sup>6</sup> tracks/cm<sup>2</sup>) on expert-curated reference data. This improvement is due to a comprehensive training dataset comprising a wider range of track densities and etch-pit morphologies.</div><div>Unexpectedly, <em>HALtracks3D</em> performed worse (91.5–80.1 %), likely because reflected-light information—critical for recognising track openings—became underrepresented among multiple transmitted-light inputs during CNN training. At very high track densities (>8.54 × 10<sup>6</sup> tracks/cm<sup>2</sup>) pushing the analytical boundaries of optical fission-track counting, <em>Coincidence Mapping</em> (Gleadow et al., 2009) remains more accurate than <em>HALtracks 2D</em>. Thermochronologists might therefore consider utilising a combination of automated fission-track algorithms depending on FT density.</div><div>Future work could expand the open-access training dataset to include a broader range of apatite FT specimens, and increased metadata for targeted CNN training on spurious features such as surface imperfections and dislocations, which are misidentified as fission tracks by existing algorithms. The open-access testing dataset presented here provides a benchmark for evaluating future FT algorithms.</div><div>Nevertheless, HALtracks 2D's enhanced accuracy brings apatite FT analysis significantly closer to full automation, with the potential to mitigate observer bias, reduce inter-laboratory variability, and broaden the accessibility of the technique to the wider geoscience community.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"208 ","pages":"Article 106096"},"PeriodicalIF":4.4,"publicationDate":"2025-12-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145791271","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","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