The block incomplete LU (BILU(0)) preconditioner is widely adopted for solving large-scale block-sparse linear systems arising from coupled partial differential equations (PDEs). However, strong inherent data dependencies and high memory bandwidth requirements of block matrix operations in the preconditioner pose significant challenges for efficient implementation on GPUs. Existing methods face the tradeoff between parallelism and convergence, and efforts to leverage the block properties remain limited. In this work, we introduce HiDAP-BILU for dependency-aware BILU(0) preconditioner on GPUs. HiDAP-BILU employs a block-centric hierarchical parallelism strategy to maximize concurrency. A block-to-block waiting scheme is proposed to preserve inter-block dependencies, while warp-level data distribution operations are utilized to maintain intra-block dependencies. Aligned with block-centric hierarchical parallelism, various architecture-aware optimizations are designed to minimize warp divergence and ensure coalesced memory access. Experiment results demonstrate that HiDAP-BILU achieves up to 4.77x and 3.10x speedups compared with state-of-the-art general methods for factorization and triangular solve, respectively. Additionally, it provides an average speedup of 3.11x in end-to-end BILU-preconditioned iterative solvers.

Air quality prediction is essential for public health and environmental management. Traditional approaches are categorized into physics-based and data-driven models. Physics-based models are accurate but computationally expensive, while data-driven models are efficient but may violate physical principles. Physics-informed approaches combine the strengths of both, but they often lack explicit decomposition of time series, and the Gated Recurrent Unit (GRU) encoder commonly used in these models often fails to capture important non-contiguous states. To address these limitations, we propose PI-TPDNet, a novel dual-branch framework with a fusion module. Specifically, the first branch decomposes the time series into trend, period and disturbance components for feature extraction, while the second branch leverages Neural Ordinary Differential Equations (Neural ODEs) and an enhanced GRU to learn physically consistent representations. These dual representations are then fused based on Graph Attention Networks (GAT) to boost overall prediction performance. Experiments on two real-world datasets demonstrate that PI-TPDNet achieves superior performance for 72-hour predictions, reducing prediction errors by up to 7.7% and providing a robust and efficient solution for practical air quality prediction.

As artificial intelligence models grow in complexity, optimizing neural network inference has become a critical challenge. Existing approaches often rely on manual, expert-driven tuning tailored to specific hardware, which lacks scalability across diverse architectures. In this paper, we propose an automated optimization framework with a hierarchical two-layer tuning mechanism. At the node level (intra-operator), we introduce an I/O lower-bound theory based on the Red-Blue Pebble game and the ( X 1 , X 2 )-Partition theorem to guide tiling and memory-mapping configurations. At the graph level (inter-operator), we employ a reinforcement learning (RL) strategy to adaptively identify optimal operator fusion boundaries across network topologies. By synergizing theoretical I/O constraints with graph-level adaptive fusion while accounting for search overhead, the framework systematically explores high-performance execution patterns. For the TileAttn operator, the ( X 1 , X 2 )-Partition theorem raises DRAM flow estimation accuracy from 77.5%-82.0% under X -Partition to 86.3%-94.8%. Our method reduces shared memory traffic by 10.79% on average, achieves the best performance in 65.3% of cross-platform cases and top-two in 86.1%, and its DQN-based fusion engine outperforms greedy strategies in 91.67% of scenarios. We further analyze the algorithm’s overhead and its amortization break-even points.

This paper introduces a physics-driven approach to improve fluid dynamics simulations of multi-body geometries with non-matching interfaces. Conventional methods often suffer from inaccuracies due to the lack of robust physical models. Our solution integrates Computational Fluid Dynamics (CFD) techniques and proposes a conservative interpolation algorithm that resolves interface mismatches without mesh modification. By using a dual-weighting scheme based on overlapping face areas, the algorithm ensures flux consistency across subdomains while maintaining high computational efficiency. Applicable to both structured and unstructured meshes, this simple yet robust method has been implemented in general-purpose CFD software and validated through complex cases. Specifically, in Couette flow between concentric cylinders, it shows a maximum 1.556% relative error in velocity distribution against analytical solutions, outperforming continuous mesh methods in accuracy. In reactor pressure vessel simulations, it achieves a pressure distribution error of 0.586% and a maximum flow distribution error of 1.645% compared to continuous mesh solutions. These results validate the method’s high accuracy and reliability in simulating diverse flow regimes, thus facilitating precise analyses for complex engineering problems.

Structured mesh generation serves as a crucial preprocessing step in numerical simulations and can be formulated as a mapping problem from geometry to structured mesh. Existing approaches typically establish an isolated mapping for each geometry. This geometry-specific paradigm fails to capture and leverage commonalities across geometries, inevitably requiring recomputation or costly retraining for new geometries. To overcome this limitation, we propose ICL-Mesh, a meta-learning framework based on in-context learning (ICL) for structured mesh generation. It treats learning one mapping as one task and trains a single neural network to extract commonalities across tasks and learn from in-context examples within each task, enabling rapid generalization to unseen tasks without parameter updates. Experimental results demonstrate that ICL-Mesh effectively generalizes to diverse geometries with only a few context examples, and even without examples. It also exhibits robustness to in-context example order sensitivity and can be extended to various mesh generation scenarios, including mesh refinement and coarsening.

Deep learning (DL) has emerged as a powerful paradigm for solving ill-posed Electromagnetic Inverse Scattering Problems (ISPs). However, standard deterministic models often produce over-smoothed reconstructions without assessing prediction reliability. While generative diffusion models offer highfidelity priors, they are prone to physical “hallucinations” that violate Maxwell's equations. To bridge this gap, this paper proposes PhyDiff, a physics-guided diffusion framework for trustworthy electromagnetic imaging. Unlike pure data-driven approaches, PhyDiff integrates the electromagnetic forward operator into the reverse sampling process via a Manifold Constrained Gradient Descent (MCGD) mechanism. This strategy steers the generative trajectory toward the solution subspace consistent with measurement data, eliminating physical inconsistencies. Comprehensive experiments on the Austria Profile benchmark demonstrate that PhyDiff achieves State-of-the-Art (SOTA) performance (SSIM $\approx \mathbf{0. 9 2 6})$, significantly outperforming traditional methods, U-Net, and existing generative baselines. The framework's robustness is further validated via cross-solver tests using FDTDgenerated data. Crucially, PhyDiff inherently enables pixel-wise uncertainty quantification, providing physically interpretable confidence maps that enhance the trustworthiness of AI-driven decision-making in safety-critical scenarios.

Grounded Multimodal Named Entity Recognition (GMNER) extends traditional Multimodal Named Entity Recognition (MNER) by aiming to extract named entities, their types, and the bounding boxes of their corresponding visual objects from text-image pairs, thereby supporting the construction of Multimodal Knowledge Graphs (MKGs). The performance of GMNER heavily relies on manual annotations of textual entities and their associated visual regions, which are both time-consuming and costly to obtain. To address these limitations, especially in low-resource settings, existing studies have introduced data augmentation techniques to enrich annotated text-image pairs. However, current approaches still face two major limitations. First, text-based generation methods often produce incorrect entity types due to inadequate constraints on entity typing. Second, image-based generation methods fail to maintain entity–object consistency, frequently generating images that do not preserve the visual regions corresponding to the mentioned entities. To address these issues, we propose a generative multimodal data augmentation framework with entity–object consistency generation (EOCG). The framework automatically produces high-quality text-image pairs that preserve entity–object alignment, along with their corresponding entity annotations and bounding boxes. It comprises two stages: In the multimodal text generation stage, to generate entities with type constraints, we introduce an entity-aware text generation method that retains key entities to enhance control over entity types during generation. In the multimodal image generation stage, to improve the alignment between text entities and target objects, we generate images consistent with textual semantics while preserving the visual regions corresponding to entities. This design allows the bounding boxes in the original image to be effectively reused as annotations for entity grounding, as we can use the original annotation information. Experimental results demonstrate that EOCG significantly boosts GMNER performance on multiple benchmark datasets under low-resource settings, thereby validating the effectiveness of the proposed framework in both synthesis quality and entity recognition.

Mesh generation plays a crucial role in numerical simulation. Recently, neural network-based methods (intelligent methods) for structured mesh generation have gained significant attention due to their ability to greatly enhance meshing efficiency while maintaining mesh quality. However, existing intelligent methods face two main challenges. The first challenge is the lack of generalization ability, where even minor geometric variations can lead to invalid meshes, necessitating network retraining. The second is their heavy dependence on labeled data or prior knowledge, where the quality of the data has a substantial impact on the training results. In this paper, we propose PI-MeshONet, a generalizable and self-supervised approach for structured mesh generation. This method transforms the mesh generation task into an operator learning task, enabling generalization to different geometric boundaries with only a single training session. Additionally, the entire training process is governed by physics-informed loss functions, thereby eliminating the need for labeled data. Experimental results show that PI-MeshONet outperforms existing methods and demonstrates high meshing efficiency, particularly in scenarios involving frequent changes in geometries. This demonstrates an application-driven extension of operator learning for structured mesh generation, providing guidance for future research.

This paper presents an adaptive density-clustering rule generation method (termed ADRG). The key contributions are twofold. First, we apply a novel density-clustering algorithm that integrates natural neighborhood searching with granular-ball computing to perform password clustering. By adaptively constructing granular balls to characterize local data structures, the algorithm identifies high-density password clusters that serve as the foundation for deriving mangling rules. Second, we introduce a hierarchical prioritization mechanism specifically tailored for rule transformations, which applies a “whole-to-part, composite-to-atomic” decomposition strategy to impose a logical and scalable ordering on rule functions. Collectively, these components form an efficient, robust, and interpretable framework for automated password rule generation.To evaluate the effectiveness of ADRG, we conduct comprehensive experiments on 11 real-world breached password datasets. The results demonstrate the following: (i) Under a fixed guessing budget, ADRG achieves a significantly higher password recovery rate than both Hashcat’s built-in rule set and existing automated mangling-rule generation techniques; (ii) The generated mangling rules reveal previously undocumented password modification patterns, exhibiting strong complementarity with current rule sets. When integrated with other rule collections, ADRG yields absolute recovery rate improvements ranging from 1.0% to 7.64% across the evaluated datasets.

The advancement of GPU technology has revolutionized Computational Fluid Dynamics (CFD) by enabling efficient generation of high-precision flow field data. However, a critical challenge lies in reconciling the fast convergence of data-dependent implicit schemes with GPU parallelism, as existing methods struggle to strike a balance between computational efficiency and convergence accuracy. This study systematically evaluates four implicit approaches implemented in the industrial solver FlowStar, focusing on efficiency-convergence trade-offs stemming from serial dependencies that limit GPU utilization. To address these limitations, a balanced coloring and hyperplane method is proposed to optimize hardware resource allocation, while a unified evaluation framework quantifies performance across five unstructured grid configurations. Experimental results demonstrate that the balanced coloring and hyperplane implementations achieve comprehensive speedups of 2.33× and 1.15× against the conventional level-set parallelization. This work provides actionable insights for designing parallel CFD algorithms that harmonize parallelism with numerical stability.

Neural architecture search and hyperparameter optimization are critical for improving deep neural network performance. However, existing approaches typically treat them as isolated problems or optimize them sequentially, ignoring the intrinsic coupling between architectures and hyperparameters. This necessitates a joint hyperparameter and architecture search method. Nevertheless, such joint optimization is highly challenging due to the vast coupled search space and the prohibitive cost of candidate evaluation. To address these challenges, we propose AutoEJS, an efficient and effective three-stage hybrid iterative optimization framework for joint search. First, we utilize a representative proxy dataset for rapid warm-up to identify promising regions efficiently. Second, to overcome the tendency of conventional surrogate-assisted search heuristics to become trapped in local optima, we integrate large language models as semantic explorers to inject diverse, high-potential candidates into the iterative process. Finally, a selective re-evaluation mechanism robustly transfers knowledge to the full dataset, ensuring precise fine-tuning with a minimal budget. Extensive experiments on benchmark datasets demonstrate that AutoEJS achieves significantly higher efficiency than advanced iterative optimization heuristics. Moreover, it consistently boosts the accuracy of established network architectures across diverse datasets, yielding improvements of 2.5

Accurately modeling the spatio-temporal dynamics of blast wave propagation remains a longstanding challenge due to its highly nonlinear behavior, sharp gradients, and burdensome computational cost. While machine learning-based surrogate models offer fast inference as a promising alternative, they suffer from degraded accuracy, particularly evaluated on complex urban layouts or out-of-distribution scenarios. Moreover, autoregressive prediction strategies in such models are prone to error accumulation over long forecasting horizons, limiting their robustness for extended-time simulations. To address these limitations, we propose RGD-Blast, a robust and generalizable deep surrogate model for high-fidelity, long-term blast wave forecasting. RGD-Blast incorporates a multi-scale module to capture both global flow patterns and local boundary interactions, effectively mitigating error accumulation during autoregressive prediction. We introduce a dynamic-static feature coupling mechanism that fuses time-varying pressure fields with static source and layout features, thereby enhancing out-of-distribution generalization. Experiments demonstrate that RGD-Blast achieves a two-order-of-magnitude speedup over traditional numerical methods while maintaining comparable accuracy. In generalization tests on unseen building layouts, the model achieves an average RMSE below 0.01 and an R2 exceeding 0.89 over 280 consecutive time steps. Additional evaluations under varying blast source locations and explosive charge weights further validate its generalization, substantially advancing the state of the art in long-term blast wave modeling.

Hybrid meshes are widely adopted in industry due to their generation efficiency and superior capability to balance computational accuracy and cost. However, initially generated hybrid meshes often compromise element regularity to guarantee topological validity and boundary fitting, making mesh smoothing an indispensable post-processing procedure to rectify local distortions. Since heterogeneous elements in hybrid meshes involve divergent optimization objectives, most existing smoothing methods remain restricted to isomorphic meshes such as all-triangular or all-quadrilateral meshes. Furthermore, the majority of current approaches struggle to effectively eliminate negative-area elements, which severely limits their practical applicability in complex scenarios.To address these challenges, we proposes EAGLE, an Element-Aware Graph Learning Engine, for unsupervised hybrid mesh smoothing. The framework integrates three key components: (1) a lightweight graph neural network to extract features of heterogeneous mesh elements, (2) a geometric constraint module to maintain structural consistency during the smoothing process, and (3) a plug-and-play loss function to support end-to-end unsupervised optimization. Extensive experiments conducted on surface meshes with several geometries verify that the proposed method achieves superior performance over existing baselines across various evaluation metrics. Notably, the EAGLE shows impressive robustness through its zero-shot generalization when applied to isomorphic meshes.

The analysis of penetration mechanics is critical for the offensive targeting and defensive design of underground facilities. Although computational methods are fundamental to penetration analysis, they are often constrained by a trade-off between accuracy and computational efficiency. Emerging artificial intelligence (AI) methods, with inherent strengths in modeling complex high-dimensional relationships from available data, provide promising alternatives for building intelligent surrogate models. This study proposes a fusion-enhanced radial basis function network (FE-RBFN) for penetration prediction, solving forward and inverse problems with multi-fidelity data. FE-RBFN employs three interconnected subnetworks to extract features and capture nonlinear correlations at varying fidelity levels. To overcome the challenge of data scarcity, FE-RBFN embeds a data fusion strategy to fully leverage multi-fidelity data from multiple sources. The experimental results demonstrate that our network yields rapid and precise predictions, outperforming traditional machine learning methods. Notably, in multi-fidelity scenarios, FE-RBFN exhibits robust prediction accuracy despite the limited availability of high-fidelity data.

Image anomaly detection (IAD) usually requires a separated train set to build an inductive model, which then infers on the test set. However, the cost of collecting and labeling training images has inspired zero-shot IAD (ZS-IAD), which directly processes the test set without the train set. Most ZS-IAD methods resort to pre-trained foundation models (e.g., CLIP), which rely on external prompts and lack adaptation to the target IAD scene. By contrast, context-guided ZS-IAD methods have recently attracted a growing interest: They not only avoid using external prompts by exploiting scene-specific context clues within unlabeled images, but also achieve superior performance to prior ZS-IAD counterparts. Unfortunately, existing context-guided ZS-IAD methods suffer from two vital flaws: The absence of train set forces them to set key hyperparameters blindly, which leads to unreliable performance. Besides, they do not actively handle mixed anomalies that disturb the learning process. To this end, we propose to automate context-guided ZS-IAD by a novel Self-Constrained Anomaly Detector (SCAD), which makes the following contributions: (1) We propose a novel self-constrained mechanism that can automatically determine proper values for key hyperparameters. (2) We design a new online self-constrained sampler that terminates the time-consuming sampling process by a proper stopping point, which can significantly reduce the computational cost. (3) We develop self-constrained normality refinement strategies that can actively constrain anomalies' impact and automatically rectify the stopping threshold. To the best of our knowledge, this is also the first work that addresses hyperparameter selection in the IAD realm. Experiments show that SCAD not only yields comparable performance to classic IAD solutions, but also matches ZS-IAD solutions enhanced by hindsight knowledge (i.e., hyperparameters validated on the test set).

Identifying interfacial binders for energetic materials remains constrained by the inherent trade-off between computational speed and predictive fidelity. Experimental screening is labor-intensive, while empirical force-field molecular dynamics (MD) offers efficiency at the cost of accuracy. Conversely, high-fidelity approaches such as free-energy perturbation (FEP) are computationally prohibitive for large chemical spaces. Here, we develop conventional molecular dynamics-free energy perturbation(cMD-FEP), an automated Uni-Mol-based machine learning framework that seamlessly bridges conventional MD and FEP to enable high-fidelity, high-throughput screening of interfacial binders. cMD-FEP employs staged, independent task-specific fine-tuning-first on large-scale MD trajectories and subsequently on FEP data-yielding an end-to-end predictive pipeline that directly maps molecular structures to interfacial interaction and free energies. Trained on datasets of similar to 6.3 x 104 simulation-derived entries, cMD-FEP achieves FEP-level accuracy with orders-of-magnitude acceleration, screening similar to 106 octogen (HMX)-binder pairs within 10 min and predicting reliable free energies for similar to 3 x 104 candidates. Clustering of top-ranking binders identifies several noncanonical chemotypes with superior adhesion strength. Experimental validation on nitrogen-containing binders confirms the predicted free-energy trends, demonstrating strong agreement with cMD-FEP results. This framework exhibits remarkable generalizability and robustness, offering a potential route for data-driven discovery of organic interfacial functional materials beyond energetic systems.

Large-scale mantle convection simulations repeatedly solve sparse velocity-pressure systems, and the multigrid velocity solver often dominates the total runtime. This paper presents an MT-3000-oriented heterogeneous multigrid solver for CitcomCU on the new-generation Tianhe supercomputer. The solver keeps the original governing equations, finite-element discretization, Uzawa iteration, and geometric multigrid framework, while redesigning the dominant kernels and data movement for the explicitly managed memory hierarchy of MT-3000. The central design is an 8-color block Gauss-Seidel smoother co-designed with ELL-like node-based storage, topology-derived neighbor indexing, AM-aware row-wise streaming, ping-pong double buffering, host-accelerator collaboration, and communication-computation overlap. These techniques expose block-level parallelism, improve locality, and reduce avoidable DMA traffic without increasing the number of multigrid cycles in the tested cases. For the Busse1993 benchmark, the optimized solver remains consistent with the original CitcomCU implementation: after 4000 time steps, the full-domain relative L2 errors are 0.47% for temperature and 2.27% for the three-component velocity vector. Compared with the unmodified production CPU-based CitcomCU baseline, the complete heterogeneous execution mode achieves system-level speedups of up to 4.91x in average multigrid-cycle time and 4.71x in average complete-Uzawa-solve time per simulation step. Weak scaling maintains 59.54% efficiency on 12,800 processes, demonstrating the effectiveness of MT-3000-specific algorithmic and data-movement redesign for matrix-based geodynamic multigrid solvers.

Mesh generation remains a key technology in many areas where numerical simulations are required. As numerical algorithms become more efficient and computers become more powerful, the percentage of time devoted to mesh generation becomes higher. In this paper, we present an improved structured mesh generation method. The method formulates the meshing problem as a global optimization problem related to a physics-informed neural network. The mesh is obtained by intelligently solving the physical boundary-constrained partial differential equations. To improve the prediction accuracy of the neural network, we also introduce a novel auxiliary line strategy and an efficient network model during meshing. The strategy first employs a priori auxiliary lines to provide ground truth data and then uses these data to construct a loss term to better constrain the convergence of the subsequent training. The experimental results indicate that the proposed method is effective and robust. It can accurately approximate the mapping (transformation) from the computational domain to the physical domain and enable fast high-quality structured mesh generation.

Sparse Matrix-Matrix Multiplication (SpMM) is a fundamental kernel across scientific computing and machine learning. While prior work accelerates SpMM using Tensor Cores, no existing sparse kernel exploits the asynchronous features of modern GPU architectures, such as NVIDIA's Tensor Memory Accelerator (TMA) and warp specialization. This work systematically studies how these features impact SpMM performance and introduces two co-designed kernels. For structured sparsity, we optimize a warp-specialized producer-consumer pipeline overlapping TMA data transfer with WGMMA computation using Block Compressed Sparse Row (BCSR) format. For irregular sparsity, we design a Window Compressed Sparse Row (WCSR) kernel that loads the sparse operand via TMA and splits large row-windows across thread blocks for load balancing. Our WCSR kernel outperforms all prior SpMM kernels on SuiteSparse matrices (1.47x over AccSpMM, 6.24x over cuSPARSE). Our BCSR kernel achieves a combined 2.66x end-to-end speedup on Qwen2.5-7B prefill at 90

High-frequency electromagnetic simulation, crucial for defense and communications applications, requires parallel computing to tackle the escalating computational costs. As an alternative to traditional numerical methods, Physics-Informed Neural Networks (PINNs) struggle with high-frequency problems due to spectral bias, while their variant Feature-Enhanced-Physics-Informed Radial Basis Network (FE-PIRBN) can better capture high-frequency solutions but suffer from quadratic computational complexity O(NM) with respect to the number of radial basis centers. To address this limitation, we propose KRBFNet, a high-performance computing (HPC) oriented framework that implements and scales the established paradigm of localized RBF approximations within a physics-informed network. By employing k-nearest neighbor (KNN) search to dynamically select relevant RBF centers, this approach reduces the dominant computational cost of RBF evaluations from O(NM) to O(Nk) (while requiring an initial O(NM) distance computation), maintaining accuracy while significantly improving efficiency. Experimental results demonstrate that our method achieves comparable accuracy to FE-PIRBN while delivering a 2.55x speedup on a single GPU and a 7.02× speedup on 8 GPUs (relative to a baseline with 3 GPUs.). This work bridges classical meshless PDE solvers (e.g., RBF-FD) with modern deep learning, providing a scalable, GPU-accelerated framework for parametric electromagnetic simulations.