Critical challenges still exist in the practical modeling of precise three-dimensional (3D) geological models with borehole data in geological engineering owing to its sparse distribution. Traditional data-driven methods often neglect macroscopic geo-environmental information, such as Digital Elevation Models (DEM) and distance to faults. The neural network models have demonstrated impressive modeling capabilities, but they are difficult to represent complex and non-linear geological features since its limitation in linear weights with fixed structure. To address these issues, a Kolmogorov-Arnold Network (KAN)-based dual-branch model for 3D geological modeling was proposed. The spatial position encoding branch is designed to capture spatial dependencies and lithological distribution laws by mapping 3D coordinates into a high-dimensional feature space. In parallel, the geo-environmental factor branch leverages the flexible fitting capability of KAN to parse complex response relationships between geo-environmental factors and subsurface lithology, effectively encoding geological prior knowledge into constraint features. A mining area in Ordos, Inner Mongolia, China was selected as case study. The results were systematically compared against five traditional machine learning algorithms. Results demonstrated the proposed method significantly outperforms benchmark models in evaluation metrics and effectively characterizes strata's structural morphology and stratigraphic relationships. Ablation studies further demonstrate that both the KAN structure and the geo-environmental data processing branch significantly enhance the reliability of the model. Therefore, our method offers a novel solution for the accurate modeling of complex geological bodies from sparse borehole data.
Multi-source precipitation products are increasingly used in hydrological modelling, climate diagnostics, and environmental risk assessment, but selecting an appropriate product remains difficult because candidate datasets differ in format, resolution, uncertainty structure, and performance across regions and diagnostic objectives. This study presents PREMISE v2.1, a configuration-driven decision-support framework that integrates rule-based data acquisition, format conversion, metadata harmonization, basin-oriented preprocessing, extreme and drought-related index generation, multi-metric evaluation, and MCDM into a reproducible workflow. A case study over China and three representative basins demonstrates the practical value of the framework. The main findings are as follows: (1) CHIRPSv3-rnl achieved the best national-scale performance, with a bias of −0.21 mm day−1, RMSE of 0.62 mm day−1, CC of 0.94, KGE of 0.89, and an entropy-weighted TOPSIS score of 1.00; (2) product suitability was strongly basin-dependent, with CMORPH_BLD and CPC ranking highest in the Haihe River Basin, CN05-1 and CPC in the Pearl River Basin, and MSWEP-V315 in the Southwest Basin, with top TOPSIS scores ranging from 0.95 to 1.00; and (3) diagnostic-specific evaluation showed that CHIRPSv3-rnl achieved the highest national extreme-index score of 0.92, while CHM_PREV2 and CN05-1 showed strong drought-oriented SPI consistency in the Haihe and Pearl River basins, with SPI correlations of 0.96–0.99 and RMSE values as low as 0.16–0.27. These results confirm that no single precipitation product is uniformly optimal across all domains or objectives. PREMISE provides an open, extensible, and auditable framework for transparent precipitation-product screening and application-specific environmental data selection.
This paper addresses the challenges in 3D geological modeling, specifically the difficulties in integrating multi-source heterogeneous geological data and constraints, which often lead to inefficient and inflexible surface reconstruction. We propose VolSur, a novel and efficient geological volume surface reconstruction framework. It provides a high-performance, flexible, and scalable solution for integrating diverse constraints, particularly suitable for 3D geological modeling workflows requiring rapid iteration and dynamic updates. The core innovation lies in the design of a unified, data-structure-agnostic Boolean volumetric intermediate representation layer, which fundamentally differentiates our method from existing voxelized Boolean combination approaches. Built on a high-performance sparse voxel (N-tree) structure, this layer transforms the complex surface reconstruction task into parallelizable Boolean bitwise operations performed at the internal node bitmask level, enabling an efficient integrated pipeline for constraint incorporation, volume fusion and surface generation. Based on an efficient, high-performance sparse voxel (N-tree) data structure, the framework seamlessly integrates various native geological volume models, including regular block, octree block, and variable block models. It uniformly incorporates both explicit volumetric constraints and implicit volumetric constraints. By performing efficient Boolean combination operations on this unified representation, the framework rapidly reconstructs an adaptive, quad-dominant surface mesh that preserves the stair-stepped boundaries characteristic of native volume models. Furthermore, the framework supports incremental updating by specifying constraints on regions with attribute changes, avoiding redundant global computations and significantly improving reconstruction efficiency. Compared to traditional volumetric surface reconstruction methods, the proposed framework integrates both explicit and implicit constraints. It achieves efficient surface reconstruction for large-scale voxel models on commodity hardware, with its adaptive reconstruction algorithm capable of reducing the number of quadrilateral faces. Experiments on multiple geological datasets validate the framework's computational efficiency and reliability in handling large-scale, multi-scale, multi-constraint geological volume data, offering a practical solution for dynamic surface reconstruction in geological modeling. The core algorithms of the framework are currently integrated into commercial geological application software.
In Geological Carbon Sequestration (GCS), reactive transport modeling of CO2-water-rock interactions is computationally prohibitive due to stringent requirements for small time steps and high-resolution grids in traditional numerical solvers. This study develops a novel coupled prediction model (CPM) integrating computational fluid dynamics (CFD) with deep learning to overcome this bottleneck. The method employs a Convolutional Long Short-Term Memory (ConvLSTM) neural network trained on high-fidelity data points generated by OpenFOAM in a real carbonate rock representative elementary volume, capturing a wide range of dissolution patterns. The trained model achieves exceptional short-term prediction accuracy, with normalized root mean square errors of merely 0.62% for concentration and 1.24% for porosity fields, visually indistinguishable from CFD results. To ensure long-term stability, a hybrid strategy of "four steps neural prediction followed by one step CFD correction" is proposed, effectively constraining error accumulation over 40000 s simulations. While accurately reproducing major morphological evolution trends, minor smoothing occurs at complex fluid-rock interfaces. Quantitative analysis demonstrates consistent predictive performance across a range of injection rates within the investigated carbonate REV. Crucially, CPM achieves a 9.8× speedup over conventional CFD under identical computational resources, maintaining physical consistency through periodic physics-based corrections. This work provides a framework for efficient reactive transport prediction, with direct implications for geological carbon sequestration monitoring and reservoir engineering.
Phased array-based look-ahead of drill bits technology serves as a common method in logging-while-drilling (LWD). It utilizes an array of acoustic sources to excite synthetic acoustic wavefields for detecting geological anomalies. Therefore, understanding the spatial–temporal evolution of the multi-source synthesized acoustic wavefield in geological structures is crucial. Previous numerical strategies mainly depend on finite element methods with refined meshes that are computationally expensive. Recently, physics-informed neural networks (PINNs) have attracted considerable attention as a novel paradigm for solving partial differential equations (PDEs). However, PINNs face over-fitting and point-wise weight imbalance challenges when learning multi-source complex wavefields, potentially limiting phased array-based acoustic wavefield reconstruction. This work presents a soft residual attention-based physics-informed neural network (SRA-PINN) to address these problems and reconstruct the phased array-based synthetic wavefield using sparse data. Firstly, the acoustic wave equation is embedded into the loss function as a penalty term to supervise the training process of the deep neural network to characterize the evolutionary process of the synthetic acoustic wavefield. Furthermore, the proposed soft residual attention mechanism assigns point-wise adaptive weights based on the point-wise PDEs residual, thereby enhancing reconstruction quality. Finally, numerical results demonstrate that the proposed SRA-PINN-based method is not only more accurate than the standard PINN but also requires less training data than purely data-driven methods. It efficiently reconstructs multi-source overlapping acoustic wavefields with limited data and without meshes.
