Aromatic secondary organic aerosol (SOA) is a key component of fine particulate matter in urban areas of China，contributing significantly to air pollution. In this study，a semi⁃explicit aromatic oxidation scheme (SXP) was implemented in the atmospheric chemistry module of the Earth System Model to better represent the oxidation processes of anthropogenic aromatic volatile organic compounds (VOCs). The SXP scheme explicitly simulates the competition between RO2 autoxidation and termination reactions and is used to evaluate the contribution of autoxidation to aromatic SOA formation. Compared with the default model (BASE) that uses the volatility basis set (VBS)，the simulation results of the SXP scheme reduces the average bias in surface SOA concentrations by approximately 5%. The differences between the simulation results of SXP and BASE show distinct regional and seasonal patterns，with SXP simulating higher SOA levels in southern China and during the summer months. Mechanistic analysis indicates that the differences arise primarily from the influence of NO x on aromatic SOA formation. Under low⁃concentration NO conditions (<15 ppbv)，SXP simulates 20%~50% higher aromatic SOA compared to BASE. In contrast，under high⁃concentration NO conditions (>25 ppbv)，SXP simulates approximately 10% lower SOA levels than BASE. This is due to the inclusion of the competition between RO2 autoxidation and NO termination reactions，where high NO concentrations suppress autoxidation and reduce SOA formation efficiency. Finally，the study examines changes in aromatic SOA during 2013-2019，a period of significant emission reductions in China. The results provide insights into the trends of anthropogenic SOA under NO x reduction scenarios，helping to improve understanding of the impact of NO on aromatic SOA formation and its regional distribution. This work underscores the importance of accounting for the competition between autoxidation and NO termination reactions in accurately predicting SOA formation in response to emission controls.
The Local Climate Zone (LCZ) is a high⁃resolution urban land use classification. Incorporating LCZ maps as input data into numerical models can provide detailed urban morphological parameters，thereby enhancing the precision of simulations in urban areas. This study focuses on a heat wave event in Shanghai from July 19 to 28，2017，and employs the Weather Research and Forecasting (WRF) model coupled with the LCZ scheme for numerical simulation. Three experimental cases CTL，NOAH，and NOURB were designed to investigate the contributions of land use change and anthropogenic heat to the urban heat island in Shanghai during the heat wave，as well as the impact of urbanization on short⁃term heavy precipitation. The results indicate that when both urban land use changes and anthropogenic heat emissions are considered，the model successfully reproduces the canopy urban heat island at night and the urban heat island circulation daytime in Shanghai. At night，both land use change and anthropogenic heat positively contribute to the canopy urban heat island intensity，with contributions of 0.89 ℃ and 1.02 ℃，respectively. During the day，anthropogenic heat contributes positively by 0.27 ℃ to the canopy urban heat island (UHI)，while land use change exerts a negative contribution of 0.28 ℃ due to the shading effect of buildings. The simulation of the short⁃term heavy precipitation process on July 22，2017，in Shanghai found that the presence of cities reduces the convective available potential energy (CAPE) in the lower atmosphere by reducing water vapor and altering temperature vertical distribution，thereby suppressing short⁃term heavy precipitation in urban areas. The findings of this study contribute to a better understanding of the urban heat island during summer heat waves and provide theoretical insights for addressing potential heat stress from future heat wave events and improving the forecasting capability for small⁃scale precipitation in urban areas during summer.
The analog⁃to⁃digital converter (ADC) serves as a key interface bridging the physical world and digital systems，playing a critical role in modern electrical systems such as edge sensing and intelligent computing. However，ADCs based on CMOS technology are constrained by manufacturing processes，which typically allow them to be constructed only using basic components such as capacitors，resistors，and transistors. The absence of non⁃volatile，tunable devices in the analog domain means that traditional ADCs often incur significant power consumption and chip area overhead when calibrating non⁃ideal characteristics or attempting to improve sampling efficiency across diverse scenarios. This paper introduces memristors into ADC design，leveraging their non⁃volatile and continuously tunable resistance to achieve in⁃situ calibration of circuit non⁃idealities and arbitrary programmability of quantization characteristics without any additional hardware overhead. Simulation results based on the 180 nm process demonstrate that the memristor⁃based ADC can perform in⁃situ circuit calibration solely through memristor programming，reducing the maximum differential nonlinearity by 71.6% and the maximum integral nonlinearity by 48.0%. Moreover，it maintains excellent calibration performance consistently across various process corners，temperatures，and supply voltages. Furthermore，by programming the memristors，arbitrary quantization curve shapes—such as linear，exponential，logarithmic，and "S"⁃shaped—can be reconfigured on the same ADC. This enables the memristor⁃based ADC to adapt its quantization curve to the statistical characteristics of different real⁃world signals (e.g.，images，audio)，thereby maximizing sampling efficiency. Simulation results confirm that the memristor⁃based ADC achieves higher resolution with lower power consumption. The memristor ADC proposed in this paper highlights the significant potential of emerging memristor devices in advancing integrated circuit development for intelligent edge systems.
Efficiently leveraging imperfect expert demonstration data is one of the key challenges in the field of imitation learning. Imperfect demonstrations，which entail the loss of fine⁃grained local details，can lead to error accumulation. To address scenarios with incomplete demonstrations，we propose a robust imitation learning method based on generative trajectory modeling. Our approach utilizes a Decision Transformer to generate continuous state⁃action sequences conditioned on incomplete demonstrations，thereby completing the trajectories. The quality of these generated trajectories is controlled by setting a target return. The completed trajectories are then fed into a diffusion model for further trajectory generation. Additionally，we introduce a novel scoring mechanism to evaluate whether the trajectories generated by the diffusion model conform to the environmental dynamics constraints present in the expert demonstrations. This model can generate high⁃reward trajectories that adhere to the dynamics constraints while effectively completing imperfect trajectories，thereby significantly reducing error accumulation and improving planning accuracy. Experiments show that our method outperforms existing offline reinforcement learning approaches. It successfully addresses the planning⁃to⁃reality mismatch problem in high⁃dimensional continuous spaces and demonstrates superior robustness and generalization capabilities.
Multi⁃label learning has found extensive application in practical scenarios such as text classification. However, in real⁃world settings, labels are often partially missing due to high annotation costs, human oversight, or complex data collection processes. Such incomplete supervision significantly impairs a model's ability to capture the underlying label structures. Most existing methods rely on a global low⁃rank assumption to model the label matrix and recover latent label structures. However, this assumption is often inadequate for capturing the complex and diverse relationships among labels. To address this limitation, we propose a Local Low⁃Rank Decomposition method for Multi⁃Label Learning with Missing Labels (LLRD⁃MLML). Our approach constructs label⁃correlated subsets and applies low⁃rank modeling to these subsets to uncover complex structural relationships among labels. Furthermore, we introduce a local⁃global joint learning mechanism that integrates the local models with a unified prediction model. To preserve the structural consistency among samples, a graph regularization constraint is incorporated. The resulting optimization problem is solved using an alternating optimization strategy. Experimental results on multiple public multi⁃label datasets demonstrate that the proposed method achieves superior performance and stability across various label missing ratios, thereby validating the effectiveness of local low⁃rank structure modeling in scenarios with missing labels.
In recent years，large speech models have demonstrated strong capabilities in cross⁃modal representation and cross⁃lingual generation，offering a new paradigm for end⁃to⁃end speech translation (E2E ST). Existing E2E ST methods have gradually leveraged these advantages to alleviate alignment and optimization challenges inherent in traditional training. However，directly generating translations from speech remains difficult，as the model is required to simultaneously perform acoustic modeling，semantic understanding，and cross⁃lingual generation within a single objective. This challenge becomes more severe when parallel data are limited，often leading to unstable training and degraded translation quality. To address these issues，this paper proposes a staged training and policy optimization method for end⁃to⁃end speech translation. Under a unified autoregressive generation framework，the proposed method organizes keywords，ASR transcriptions，and target translations into a structured output sequence and progressively enhances the model's speech understanding and translation generation capabilities through three stages. First，keyword prediction and speech transcription are jointly modeled to construct a dual⁃granularity source representation comprising a keyword⁃level semantic skeleton and a complete source transcription，thereby establishing stable acoustic–semantic correspondences. Second，the full translation objective is introduced，and mixed auxiliary labels are constructed by combining human annotations with predictions from the previous stage，improving the model’s adaptability to intermediate⁃representation noise during inference. Finally，reinforcement learning based on Group Relative Policy Optimization (GRPO) is adopted，with reward modeling applied exclusively to the final translation subsequence，to further refine the translation generation strategy. Experiments on three MuST⁃C language pairs using Qwen2⁃Audio，Qwen2.5⁃Omni，and Qwen3⁃Omni demonstrate that the proposed method achieves significant improvements in both BLEU and COMET scores，confirming that the integration of staged training，mixed auxiliary labels，and translation⁃level policy optimization effectively enhances both the stability and cross⁃lingual generation quality of end⁃to⁃end speech translation.
Density Peaks Clustering (DPC) is a density⁃based clustering algorithm capable of automatically identifying clusters of arbitrary shapes without requiring the number of clusters to be pre⁃specified. However，when processing datasets containing clusters with significant density variations，DPC tends to erroneously select multiple cluster centers within dense clusters while overlooking the true centers in sparse clusters. Furthermore，DPC's single⁃step chained allocation strategy is prone to a “domino effect”，where the misallocation of a single data point can trigger a cascade of erroneous assignments for subsequent points. To address these issues，this paper proposes a novel Density Peaks Clustering algorithm based on Mutual Nearest Neighbors and Dempster⁃Shafer Theory，termed MDS⁃DPC. First，by integrating a sample's distance⁃based similarity with its neighborhood cohesion，we redefine the calculation of local density to effectively mitigate inter⁃cluster density disparities. Second，to more accurately characterize the relative positional relationships between samples，we refine the relative distance metric by leveraging both local and global distribution characteristics，specifically through the integration of local density peaks and a mutual nearest neighbor graph. Finally，we introduce Dempster⁃Shafer theory and employ a multi⁃stage assignment strategy coupled with a hierarchical fine⁃tuning mechanism for cross⁃cluster connected samples to enhance the overall accuracy of sample allocation. We evaluated the proposed MDS⁃DPC algorithm against six state⁃of⁃the⁃art clustering algorithms on nine synthetic and twelve real⁃world datasets. The experimental results demonstrate that MDS⁃DPC achieves superior clustering performance.
Multi⁃view clustering aims to exploit the consensus and complementarity across different views，yet existing methods face two major bottlenecks: first，global topology modeling relies on predefined similarity measures and fixed neighborhoods，making it difficult to adapt to complex data distributions; second，sample⁃level view quality varies significantly，and static weighting strategies fail to characterize fine⁃grained reliability changes，particularly lacking robustness in missing data scenarios. To this end，we propose a multi⁃view clustering framework based on Dynamic Anchors and quality⁃aware Mixture⁃of⁃Experts (DAMC⁃MoE). First，a learnable dynamic anchor mechanism replaces traditional predefined similarity measures，achieving end⁃to⁃end deep coupling between topological structure modeling and feature representation learning. Building on this，a quality⁃aware mixture⁃of⁃experts module is introduced，which generates quality tokens from sample⁃level completeness and signal⁃to⁃noise ratio to guide the gating mechanism for adaptive routing，realizing a paradigm shift from conventional view⁃level weighting to sample⁃level fine⁃grained perceptual fusion. Finally，a three⁃level contrastive learning mechanism is constructed to jointly reinforce semantic alignment from inter⁃view，intra⁃view，and local⁃global perspectives. In comprehensive comparative experiments on 5 benchmark datasets against 11 state⁃of⁃the⁃art algorithms，DAMC⁃MoE demonstrates superior clustering performance. Friedman test results further indicate that DAMC⁃MoE achieves significantly higher average rankings across three clustering evaluation metrics compared to all baseline methods.
Large⁃scale graph data，such as knowledge graphs and social networks，are ubiquitous. The enormous size of these graphs poses significant challenges for analytical tasks like frequent pattern mining (FPM). Graph reduction techniques have emerged as a key enabler for large⁃scale graph analysis，as they can drastically reduce graph size while preserving critical information. However，existing graph reduction methods primarily focus on retaining specific attribute information or minimizing global information loss，often neglecting the preservation of local high⁃order structures. This limitation leads to suboptimal support for FPM tasks. To address this issue，we propose a Structure⁃Preserving Graph Reduction Framework (SPGRF) that retains a higher quantity of high⁃quality frequent patterns while reducing data scale. First，we introduce a local⁃to⁃global edge importance evaluation method that guides the construction of an initial reduction skeleton by precisely identifying key nodes and core edges. Second，to compensate for structural degradation caused by reduction，we design a neighborhood⁃based skeleton enhancement mechanism to improve the structural diversity and completeness of the reduced graph. Extensive experiments on real⁃world graphs demonstrate the superiority of our framework across multiple metrics. The generated reduced graphs effectively preserve the original "backbone" structure: when the graph is reduced to 10% of its original size
RNA interference (RNAi) is a promising therapeutic strategy. However,existing siRNA efficacy predictors lack uncertainty quantification. Here,we present an OligoFormer⁃based model to quantify both model and data uncertainty using the Huesken and Mixset datasets. Our analysis reveals that data uncertainty dominates overall uncertainty. Using it as a filter，we derived low⁃uncertainty subsets,denoted as Huesken' and Mixset'. In cross⁃validation，the Huesken' achieved AUC，PRC，F1⁃score,and PCC scores of 0.950，0.941，0.936,and 0.791，respectively,while Mixset' scored 0.901，0.930，0.773，and 0.733，surpassing the performance on the original data. In cross⁃dataset tests，models trained on Huesken' outperformed those on the original data，with AUC，PRC，and PCC improving by 1.29%，0.93%，and 0.72%,respectively. Uncertainty estimation thus enhances both prediction confidence and model generalizability via data filtering. Crucially，for deep learning–based siRNA prediction，improving data quality is more impactful than increasing data quantity.
Conventional stereo upmixing methods typically operate on audio content within the time⁃frequency domain. However，in complex musical mixtures，the overlapping of multiple sources in this domain often leads to biased estimation of correlation and source location parameters. To mitigate this issue，this paper proposes a two⁃path stereo music upmixing method incorporating source separation，built upon the existing frequency⁃domain upmixing framework. By decomposing the mixed music signal into multiple constituent sources and processing them individually，the proposed method effectively reduces interference caused by time⁃frequency overlap during parameter estimation. Subjective listening tests demonstrate that，compared with several representative upmixing methods and an ablation variant without the source separation module，the proposed approach achieves superior performance in terms of overall perceived audio quality.
Sound source localization is crucial in industrial detection，medical diagnosis and defense，where accurate signal acquisition in noisy environments is essential. Recently，non⁃Hermitian physics has shown remarkable progress in improving sensing sensitivity owing to counterintuitive phenomena such as eigenvalue splitting near the exceptional point. However，these designs cannot effectively suppress noise，which limits high⁃fidelity signal acquisition and causes localization errors. We propose a sound source localization method based on non⁃Hermitian sensing，capable of suppressing noise disturbances while capturing signals. The underlying mechanism lies in a dynamical system with history⁃dependent Hamiltonian transitions，which confines signals to discrete states against noise interference. On this basis，one⁃dimensional acoustic array studies confirm that non⁃Hermitian sensing achieves accurate time⁃delay detection at different noise levels. When extended to two dimensions，simulations show that the proposed method reduces average distance and angular errors by approximately 82.11% and 68.37%，respectively，compared with conventional sensors. Our study provides a new approach for accurate sound source localization in noisy environments and highlights the potential of non⁃Hermitian physics in anti⁃interference target detection.
