A Novel Runoff Network Synchronous Prediction Model Based on Distribution Similarity Analyzing and Sparse Matrix Mapping

The dual-module High Temperature gas-cooled Reactor Pebble-bed Module (HTR-PM) demonstration reactor, led by the Institute of Nuclear and New Energy Technology of Tsinghua University, has achieved successful operation. To enhance economic efficiency, the HTR-PM600S incorporating six nuclear steam supply system (NSSS) modules has been established as the next-phase development objective. The complex coupling effects of multimodule systems necessitate systematic simulation to preinvestigate operational characteristics. The simulator, leveraging its full-scope simulation capabilities and real-time computational performance, has evolved from an operator training tool to a critical operational analysis platform. By integrating neutronics, primary/secondary circuit thermal-hydraulic models, and control systems, the HTR-PM simulator establishes a complete framework that plays an important role in nuclear power plant operational studies. When extending the HTR-PM simulator to HTR-PM600S modeling, direct replication of standard NSSS module configurations substantially increases simulation time cost. Particularly in the primary circuit thermal-hydraulic model, the linear expansion of the number of thermal component network matrices and the dimensions of the helium flow network matrix have driven single-step simulation time beyond the 100-ms time step, constituting the critical bottleneck for multimodule simulation. This study dissects the computational time cost of different simulation tasks and matrix characteristics of a helium flow network, proposing targeted optimization strategies: implementing multithreaded parallel computing for thermal component networks and adopting KLU sparse matrix solver for helium flow network. Through these optimizations, the primary circuit thermal-hydraulic computation time has been reduced from 160 to 45 ms, achieving real-time performance while reserving an over 50% temporal margin for subsequent complex model expansions.

To address the inherent complexity and nonlinearity of electroencephalogram (EEG) signals, this study proposes a refined classification framework, NeuroNetFusion, which strategically integrates and selects multi-context network-based features for improved performance. This framework enhances the performance of bio-signal classification models by integrating multi-directionally expressed cross-dependence information. Unlike prior EEG classification studies that mainly relied on single-domain or uni-directional features, our framework introduces a systematic multi-context integration strategy, which constitutes a primary contribution of this work. The process begins with preprocessing the EEG signals using a Savitzky-Golay (SG) filter to reduce noise. Next, the signals are decomposed into multiple frequency bands using the Discrete Wavelet Transform (DWT). The resulting data are then reconstructed from five bands into corresponding adjacency matrices. Following this, the signals are represented as two distinct types of networks: a causality network based on the Directed Transfer Function (DTF) and a correlation network using the Pearson correlation coefficient. To combine features from these two networks, we utilize the TF-IDF method to vectorize the non-zero element index sequences from the adjacency matrices. This procedure transforms sparse adjacency matrices into a quantitative representation, allowing us to assess the importance of each connection within the network. Additionally, a genetic algorithm is employed to select important TF-IDF features, optimizing them for classification tasks. Performance is evaluated by comparing several conventional machine learning models using the EEG source dataset (MTOUH), employing standard evaluation metrics. The proposed model achieved a final accuracy of 88.05%, representing an 8.73% absolute improvement over the baseline TCN model. This demonstrates the effectiveness of our method in identifying abnormalities in EEG signals. The key contribution lies in bridging causality- and correlation-based representations through TF-IDF-driven feature encoding, offering a novel pathway for interpretable and scalable EEG analysis. Our approach holds promise for future applications in cross-dependent bio-signal classification problems, paving the way for further research and development in this area.

Short-term origin–destination (OD) prediction for urban rail transit is critical for efficient operation and dynamic management. Unlike conventional passenger flow forecasts that focus solely on aggregate indicators such as entries and exits, OD prediction reconstructs the complete station-to-station flow matrix, enabling fine-grained operational planning and providing a stronger basis for service design, transfer coordination, and capacity allocation from a network-wide demand perspective. However, the high dimensionality and sparsity of the OD matrix, along with complex spatiotemporal dependencies, make this task challenging. This paper introduces multi-dimensional holiday and weather tags, which are coupled with historical OD tensors into a unified multi-dimensional tensor input. The dual-graph spatial encoder captures the physical network topology and latent functional relationships between stations, while the hierarchical time-series Transformer models the short-term dynamics and periodic patterns. Furthermore, we integrate Tucker tensor decomposition in the decoder to efficiently reconstruct the OD prediction results and address data sparsity. Experiments on large-scale urban rail transit datasets show that Spatio-Temporal Graph Transformer Network (STGTN) significantly outperforms benchmark methods in prediction accuracy, with a 58% reduction in Mean Absolute Error (MAE) and an 11% increase in accuracy. Larger relative errors are mainly concentrated in peripheral or infrequently used OD pairs, whereas major travel corridors are predicted accurately. Ablation studies further confirm the contribution of each module and the external factors. These results demonstrate the effectiveness of tensor-based spatiotemporal modeling and highlight the practical value of OD prediction for short-term capacity deployment, transfer coordination, and congestion mitigation in urban rail transit systems.

ABSTRACT We introduce a dynamic factor correlation model whose core methodological innovation is a variation‐free parametrization of dynamic factor loadings, inspired by the generalized Fisher transformation. The model accommodates time‐varying correlations, heterogeneous heavy tails, and dependent idiosyncratic shocks. Applied to a Small Universe of 12 assets and a Large Universe of 323 stocks, the factor structure induces a sparse idiosyncratic correlation matrix with dependencies concentrated within subindustries, enabling scalability to high dimensions under a sparse block structure. Both factor loadings and correlations vary substantially. Allowing for heterogeneous heavy tails via convolution‐ distributions yields sizable improvements relative to Gaussian and multivariate‐ benchmarks.

This letter uses sparse matrix statistics to highlight a structural gap between North American grid models and the synthetic cases commonly used in research. North American grids exhibit area-based modularity, characterized by dense intra-area links and limited inter-area connectivity. By contrast, synthetic grids often over-mesh across areas. Analysis shows that these surplus tie lines increase fills in the Jacobian factorization, which reduces realism and computational efficiency. The observations confirm that inter-area connectivity influences sparse matrix behavior. This work emphasizes the importance of incorporating modular structure into synthetic grid design.

Multi-omics profiling-spanning proteomics, transcriptomics, and additional omics data types-is rapidly advancing, providing increasingly detailed maps of cellular identity and function. Yet, identifying rare cell populations while maintaining computational tractability remains a major challenge in large-scale multi-omics clustering. Here, we introduce the supercell paradigm, in which expression-coherent cells are grouped into intermediate units that preserve weak but biologically meaningful local structure across omics layers, thereby improving sensitivity to rare populations that are often masked at the conventional cluster level. Supercells are constructed using angle-aware similarity metrics and second-order co-occurrence neighbors, with impurity cells pruned by degree centrality. Building on this idea, we develop scHG, a high-order graph learning framework with an omics-weighted optimizer that adaptively balances contributions from gene expression, surface proteins, and chromatin accessibility while remaining scalable on large datasets through sparse matrix optimization and iterative graph refinement. Across six benchmark datasets (up to 30672 cells), scHG consistently outperforms state-of-the-art methods, improving mean ARI and NMI by 3.97% and 3.54%, respectively, while reducing runtime by 26.40%. Beyond overall clustering accuracy, scHG resolves fine-grained heterogeneity within conventionally defined T-cell populations and, importantly, uncovers rare populations-including dendritic-cell populations and NK-like B cells-that remain hidden under standard clustering pipelines. These results demonstrate that supercells provide not only an efficient intermediate representation for large-scale multi-omics integration, but also a practical mechanism for rare-cell detection.

The identification of cell types is a basic step of pipelines for Single-Cell RNA sequencing (scRNA-seq) data analysis. However, unsupervised clustering of cells from scRNA-seq data has multiple challenges: high dimensionality, sparseness of the expression matrix, and technical noise that generates false zero entries. In this study, we introduce new algorithms for clustering scRNA-seq data. The first algorithm builds a k-MST graph from distances obtained directly from the input data without dimensionality reduction. The computation follows an iterative procedure of k steps, calculating the edges of minimum spanning trees over different subgraphs obtained by removing edges selected in previous iterations. The Louvain algorithm is executed on the k-MST graph for cell clustering. We also explored an alternative based on neural networks, using an autoencoder to learn the parameters of a Gaussian mixture model. Benchmark experiments show that the algorithms have competitive accuracy, compared to previous solutions. Sequencing depth, number of cells and tissue types have important effects on the performance of the algorithms. Further experiments with scRNA-data taken from a patient with refractory epilepsy show that the autoencoder model achieved the best accuracy for this dataset, and the k-MST was competitive among graph-based approaches.

Statistical analysis of microbial count data derived from 16S rRNA or metagenomics sequencing poses unique challenges due to the sparse, compositional, and high-dimensional nature of the data. While QIIME 2 already provides many tools for data pre-processing and analysis, plugins for statistical regression, classification, and microbial network estimation tailored to compositional count data are relatively scarce. We present q2-classo and q2-gglasso, two novel QIIME 2 plugins that implement penalized regression, classification, and graphical modeling approaches for microbial compositional data. q2-classo enables the prediction of a continuous or binary outcome of interest using compositional microbiome data as predictors. Both sparse log-contrast regression and classification, as well as tree-aggregated log-contrast models are available. q2-gglasso enables the estimation of taxon-taxon association networks through sparse graphical model estimation, such as, e.g., the SPIEC-EASI framework, as well as adaptive and latent graphical models. The latent model can decompose taxon-taxon associations into a sparse direct interaction matrix and a latent (low-rank) matrix which enables robust principal component embedding of a data set. Within the QIIME 2 ecosystem we demonstrate their application on the Atacama soil microbiome dataset, illustrating robust model selection, classification, and microbial network estimation with covariates and latent factors. The software is freely available under the BSD-3-Clause License. Source code is available at https://github.com/bio-datascience/q2-gglasso and https://github.com/bio-datascience/q2-classo-latest, with installation through QIIME 2 and Docker. oleg.vlasovets@helmholtz-munich.de.

Abstract A number of modifications of the basic algorithms for constructing a multilevel structure to improve the performance of the algebraic multigrid method for both scalar and point-block systems are considered in this paper. We explore the basic operations of transposing and multiplying sparse matrices, as well as ways to select the maximum independent subset in the graph of strong connections, methods for constructing the prolongation operator, and approaches to aggressive coarsening that reduce the operation complexity of the method. It is shown that the construction of an extended prolongation operator can significantly increase the accuracy of the method, but at the cost of higher operator complexity and longer execution times. This disadvantage can be compensated either by filtering small weights from the prolongation operator, or by using aggressive coarsening. Several approaches to aggressive coarsening are considered. To confirm the conclusions, a number of numerical experiments were performed on a series of matrices from a publicly available collection for problems on progressively refined grids. The method applicability is evaluated on systems derived from adaptively generated grids. Some performance analisys of shared and hybrid memory is provided.

Abstract The sparse matrix–vector multiplication ( SpMV ) kernel is a key kernel in scientific and engineering applications, forming the core of many iterative solvers for linear systems and eigenvalue problems. Due to its low arithmetic intensity and irregular memory access patterns, SpMV remains memory-bound on modern architectures, making its efficient implementation particularly challenging. This paper presents vectorized SpMV routines for RISC-V processors with SIMD support, exploiting the RISC-V Vector Extension (RVV 1.0). We implement and evaluate three storage formats—CSR (Compressed Sparse Row), SELL- p (a vector-friendly variant of ELLPACK), and JDS (Jagged Diagonal Storage)—providing low-level implementations that leverage RVV intrinsics. Performance is assessed on two commercial RISC-V platforms (CanMV-K230 and BananaPi F3) with 128-bit and 256-bit vector registers, and on the EPAC research system featuring 16,384-bit vectors. Results show that the vectorized routines significantly outperform scalar baselines, achieving a variety of speed-ups depending on the format and architecture. These findings highlight the potential of open RISC-V architectures for high-performance sparse linear algebra and provide a foundation for future vector-aware sparse kernel optimizations.

Exploiting matrix symmetry to halve memory footprint offers a substantial opportunity for accelerating memory-bound computations like Sparse Matrix-Vector Multiplication (SpMV). However, symmetric SpMV incurs data conflicts when concurrently writing the output vector. Previous approaches fail to address this issue efficiently, i.e., either are non-scalable or yield poor performance for large high-bandwidth irregular matrices. This paper extends DCS-SpMV , a D ivide-and- C onquer (DC) based shared-memory implementation of S ymmetric SpMV. The key idea of DCS-SpMV is to recursively divide and reorder the matrix-induced conflict graph into independent subgraphs for parallel execution, and construct separate subgraphs to avoid data conflicts. The DC approach naturally transforms the input matrix into a low-conflict part and a high-conflict part, which motivates us to design a conflict-aware hybrid solution DCH-SpMV that executes these two parts using DCS-SpMV and the standard SpMV, respectively. We also develop a machine learning model for DCH-SpMV to predict the optimal number of DC recursions on a given matrix and architecture. In this work, we further optimize the hybrid DC implementation by reducing data conflicts before the DC preprocessing. First, we present a conflict-pruning strategy to decouple certain highly dense columns or rows from the conflict graph of a symmetric matrix. Second, we implement a heuristic to adaptively select the lower or upper triangular part of a symmetric matrix, leading to fewer data conflicts. Our optimizations not only facilitate the DC preprocessing, but also improve the performance of DCH-SpMV. We evaluate our work on both x86 and ARM multi-core CPUs using 298 symmetric sparse matrices from the SuiteSparse Matrix Collection. Our new optimizations improve the performance of previous version [42] by up to 4.89 ×, demonstrating significant speedup over the state-of-the-art approaches including the vendor-tuned Intel oneMKL library.

Sparse matrix–matrix multiplication (SpMM) is a fundamental operation in scientific computing with broad applications across numerous domains. Tiling is a key optimization technique for improving data locality and is widely adopted in high-performance computing. However, the irregular data access patterns inherent to SpMM make it challenging to exploit tiling effectively for data reuse. In this paper, we propose MaSpMM , a memory-aware SpMM framework that integrates cache-aware tiling with a segment-oriented data layout. MaSpMM stores matrices as continuous segments to enhance data locality within each tile. Moreover, since many sparse matrices in real-world applications exhibit symmetry, we further develop MaSpMM-Sym, an extension that recursively partitions symmetric matrices to eliminate write conflicts and further improve locality. To adapt to diverse scenarios, we finally introduce MaSpMM-Adap, which adaptively selects the most suitable approach for each input matrix. Comprehensive evaluations on both x86 and ARM CPUs demonstrate that MaSpMM-Adap achieves average speedups of up to 1.86 × over Intel oneMKL, 1.84 × over ASpT, and 1.75 × over J-Stream.

The growing number of on-orbit spacecraft and the increasing volume of telemetry data have made intelligent anomaly detection in multi-channel telemetry essential for mission operations. Current spacecraft anomaly detection methods primarily rely on statistical models or time-series deep learning approaches, which often fail to explicitly model spatiotemporal dependencies across multiple telemetry channels. This shortcoming limits their ability to capture the dynamically evolving and intricately coupled relationships between variables. To overcome this limitation, a Progressive Spatiotemporal Graph (PSTG) model is proposed for anomaly detection in multi-channel spacecraft telemetry. PSTG employs a multi-scale patch embedding module to extract hierarchical semantic features from multi-channel time series, effectively reducing the dimensionality of the spatiotemporal graph. It constructs a sparse adjacency matrix using a multi-head attention mechanism that integrates intra-channel temporal dynamics, inter-channel spatial correlations, and cross-channel spatiotemporal interactions. An improved multi-head graph attention network then captures pairwise dependencies among nodes within the adjacency matrix. As a result, PSTG encodes rich spatiotemporal representations derived from intricate variable interactions, enabling accurate, real-time prediction of multi-channel telemetry. Furthermore, a dynamic thresholding mechanism is incorporated into PSTG to perform online anomaly detection based on prediction residuals. Extensive experiments on real-world spacecraft telemetry data collected over 84 months show that PSTG outperforms eleven state-of-the-art benchmark methods in almost all cases across multiple evaluation metrics. Finally, visualizations of the learned adjacency and attention matrices are presented to interpret the spatiotemporal modeling process, providing operators with actionable insights into the detected anomalies and facilitating root cause analysis.

In this paper, we propose a novel element-wise subset selection method for the alternating least squares (ALS) algorithm, focusing on low-rank matrix factorization involving matrices with missing values, as commonly encountered in recommender systems. While ALS is widely used for providing personalized recommendations based on user-item interaction data, its high computational cost, stemming from repeated regression operations, poses significant challenges for large-scale datasets. To enhance the efficiency of ALS, we propose a core-elements subsampling method that selects a representative subset of data and leverages sparse matrix operations to approximate ALS estimations efficiently. We establish theoretical guarantees for the approximation and convergence of the proposed approach, showing that it achieves similar accuracy with significantly reduced computational time compared to full-data ALS. Extensive simulations and real-world applications demonstrate the effectiveness of our method in various scenarios, emphasizing its potential in large-scale recommendation systems.

Objective.Ultrasound (US) is commonly used to guide minimally invasive procedures; however, its effectiveness is often limited by poor needle-tip visibility. This work presents a 3D US imaging and needle-tip tracking system based on US multilateration, toward real-time, safe and efficient guidance.Approach.The proposed system integrates a fiber-optic hydrophone within a medical needle to receive US transmissions from a sparse 2D matrix array probe used for 3D imaging. Chirp excitation was employed to improve the signal-to-noise ratio and enable robust time-of-arrival estimation. A multilateration algorithm was employed for needle-tip localization relative to the US probe, requiring fewer US elements and achieving substantially faster performance than existing methods.Main results.The system achieved spatial tracking accuracy of1.27±0.65 mm using only 13 active elements, representing an order-of-magnitude increase in tracking speed from 3.77 to 45.45 Hz compared to a conventional delay-and-sum-based tracking algorithm with 256 elements. The translational potential of the system was further demonstrated withex vivochicken tissue, and a clinically realistic femoral nerve block phantom.Significance.These results demonstrate the feasibility of a multilateration-based system that combines 3D anatomical visualization with needle tracking. With further improvements towards real-time operation and validation under clinically relevant conditions, this approach could improve the precision and safety of US-guided interventions.

Transcriptome profiling is a cornerstone of functional genomics, enabling the detailed characterization of gene expression in health and disease. Bulk RNA sequencing (bulk RNAseq) remains the most widely used approach in clinical and large-cohort studies due to its cost-effectiveness, robustness, and comprehensive transcriptome coverage. However, bulk RNAseq inherently averages gene expression signals across heterogeneous cell populations, thereby masking cellular diversity and obscuring rare cell types. In contrast, single-cell RNA sequencing (scRNAseq) enables a high-resolution analysis of cellular heterogeneity, allowing the identification of distinct cell types, transitional states, and developmental trajectories. Nevertheless, scRNAseq is associated with higher cost, limited scalability, increased technical noise, sparse expression matrices, and protocol-dependent biases introduced during tissue dissociation or nuclear isolation. In this review, we summarize the conceptual and methodological foundations of integrating bulk RNAseq and scRNAseq data, emphasizing their complementary strengths and limitations. We discuss how scRNAseq-derived cell-type atlases can serve as reference matrices for computational reconstruction (deconvolution) of bulk RNAseq profiles and examine key sources of technical and biological variability. Furthermore, we outline major integration strategies, including reference-based deconvolution, pseudobulk aggregation, and Bayesian joint modeling to provide an overview of widely used analytical tools and essential components of scRNAseq data processing workflows.

Deep learning enables efficient analysis of complex data and finds applications across diverse domains such as healthcare, cybersecurity, and image and text recognition. However, implementing such applications in resource-constrained environments presents challenges due to limited memory, low processing power, and difficulty in handling computationally intensive tasks. This paper explores compression and decompression strategies to reduce memory overhead and enhance the computational efficiency of Deep Neural Network (DNN) inference. We propose a hardware-efficient clustering architecture that compresses the intermediate layer outputs of the DNN. Compressing the DNN layer outputs introduces sparsity by reducing the number of non-zero activations, which in turn lowers the computational complexity of subsequent layers. Further in this paper, we investigate the adaptability of Delta-Compressed Sparse Row (dCSR), a sparse matrix storage format originally developed for embedded systems, on FPGA and GPU platforms. We present a custom decompression unit for reconstructing sparse matrices represented in dCSR format. The proposed decompression architecture is implemented on an FPGA, and its performance is evaluated in terms of resource utilization, execution speed, and power consumption. We also implement a dCSR-based sparse DNN on a GPU that exhibits improvement in terms of sparse parameter storage and execution speed. The dCSR-based sparse DNNs achieve an average storage reduction of up to 4% and a maximum speedup of 1.92× compared to the baseline implementations on GPU.

I describe a method for accurately refining the geometrical parameters of segmented X-ray area detectors on the basis of serial crystallography data, using 'Millepede' - an algorithm created for a very similar problem in high-energy physics. The Millepede method for serial crystallography builds on the approach of Brewster et al. [Acta Cryst. (2018), D74, 877-894], in which the detector parameters are refined simultaneously with the parameters for each individual crystal. This accounts for the mutual dependency between the parameters and thereby avoids the bias and slow convergence problems that have afflicted older approaches in which the deviations between observed and calculated Bragg peak positions were taken directly as the updates for the detector panel positions. The Millepede method uses the special structure of the least-squares normal equations to reduce them to a much smaller form that can be solved very quickly, even compared with the sparse matrix methods used previously. This makes it practical to refine the detector geometry frequently and thereby maintain accurate calibration without specialized alignment campaigns. Tilts of detector panels out of the plane can be reliably refined, as can the overall distance of the detector in the beam direction. With a simulated test case, the new method produced panel shifts within 7% of the correct values with only one iteration, and produced almost exactly correct shifts after a second iteration. A simulated out-of-plane panel rotation was correctly determined to within 0.001°. Applied to experimental data from an X-ray free-electron laser, the method increased the indexable fraction of frames from 30% to 91% in a single iteration, and to 96% after two further iterations. Computing the geometry updates on the basis of 2060 crystals took only 0.819 s on desktop computing hardware, including the time taken to read the required data from disk. The scaling was found to be very close to linear for up to 100 980 sets of crystal parameters, which took only 78.2 s to process under the same conditions. The method has been applied as part of a real-time feedback system at a synchrotron radiation beamline, in which an out-of-plane detector tilt of 0.04° was detected and corrected. Possible further applications are also described here.

ABSTRACT Accurate spatial prediction on the sphere is fundamental for global environmental applications such as climate monitoring and oceanographic analysis. Existing approaches, however, often struggle to balance computational efficiency, predictive accuracy, and the ability to accommodate heterogeneous spatial structures. We propose a multi‐resolution spatial modeling framework that integrates thin‐plate spline (TPS) basis functions with Gaussian process modeling. The framework begins with a fixed‐effects representation based on a hierarchy of nearly orthogonal TPS basis functions ordered by smoothness, thereby providing a multi‐resolution decomposition of spatial variation. This allows large‐scale patterns to be captured efficiently while preserving interpretability. To represent localized dependencies, we extend the model with a random effect governed by a tapered Matérn covariance, which models fine‐scale structure while ensuring scalability through sparse matrix operations. Model complexity is adaptively controlled using the conditional Akaike information criterion (cAIC), which simultaneously selects the number of basis functions and determines the contribution of the Gaussian process component. Numerical experiments and a global sea surface temperature application show how our method balances predictive accuracy with computational feasibility, establishing its role as a powerful solution for large‐scale spatial modeling on the sphere.