Wei Lü

As brain–machine interfaces (BMIs) and neural recording technologies evolve, there is an increasing demand for edge computing systems capable of processing large amounts of neural data in real‐time to alleviate the data transmission challenges and improve BMI performance. In this work, we propose a memristor‐based reservoir computing (RC) system that leverages the short‐term memory dynamics of memristive devices to process recorded neural signals. We validate the proposed system on a behavioral state classification task using neural spike recordings from a mouse during free movement. The system achieves robust classification performance over a 2‐week period and demonstrates resilience to device‐to‐device variations and limited training data. The proposed system further enabled ablation analysis to identify the dominate neurons responsible for particular actions. These results demonstrate the effectiveness of memristor‐based RC systems as promising solutions for energy‐efficient, real‐time neural signal processing in future BMI systems.

Electronic Health Records (EHRs) provide a rich source of high-dimensional biomedical signals for clinical decision support. However, these signals, typically represented as patient-medication interaction matrices, present two fundamental challenges for signal processing: 1) extreme sparsity, and 2) a missing data mechanism that is often Missing Not At Random (MNAR), where the pattern of missingness is correlated with the unobserved true signal values. Furthermore, valuable relational information encapsulated in patient similarity networks is often ignored. To address these challenges, we propose PSMR-MNAR, a novel probabilistic graphical model for biomedical signal completion. Our approach explicitly models the MNAR generative process and integrates inter-patient relationships as a graph-based prior to regularize the problem. By defining a patient-specific "reference level," the model learns to adaptively leverage information from clinically similar patients. We develop an efficient variational inference algorithm for posterior estimation. Experiments on the real-world MIMIC-III dataset demonstrate that our model significantly outperforms state-of-the-art methods in predicting medication usage, validating the efficacy of jointly modeling the graph structure and the MNAR mechanism.

A systematic study was conducted on commercial Al–7Si–0.35Mg cast alloy modified by a multielement microalloying strategy: fixed additions of Sr and Ti–B, 0.3 wt.% La or Ce, and a graded Cu addition (0–3.5 wt.%). Microstructure and mechanical properties were characterized using optical microscopy, scanning electron microscopy, energy‐dispersive X‐ray spectroscopy, electron backscatter diffraction, X‐ray diffraction, Vickers hardness testing, and room‐temperature tensile testing. The results show that the Sr–Ti–B–La/Ce composite modifier effectively rounds/ spheroidizes corners of eutectic Si and reduces the area fraction of large Si particles (>300 μm 2 ) from 47.8% to 21.0%, while markedly increasing the proportion of finer Si particles in the 1–50 μm 2 range. Secondary dendrite arm spacing and mean grain radius were refined by ≈54% and 38.5%, respectively (average grain radius decreased from 351 to 216 μm). With increasing Cu content the alloy behavior transitions from predominantly solid‐solution strengthening to second‐phase particle strengthening controlled by Al 2 CuMg. At about 1.5 wt.% Cu, a fine dispersion of Al 2 CuMg particles forms and the mechanical properties reach an optimum, with the ultimate tensile strength of 271.9 MPa and elongation (EL) of 6.7%. When Cu exceeds ~2.5 wt.%, Al 2 CuMg phases coarsen and form continuous networks, causing ductility to fall sharply (EL down to 2.94% at 3.5 wt.% Cu) despite a continued rise in hardness with Cu content (maximum ~109 HV). Fractographic analysis indicates a uniform ductile dimple morphology at the optimal composition, whereas coarse second‐phase particles induced by excess Cu act as preferred crack‐initiation sites. Based on multiscale characterization, this work reveals synergistic interactions between Sr–Ti–B–La/Ce and Cu in terms of interfacial adsorption, solute redistribution, and precipitation kinetics. An optimized composition represented by 0.3 wt.% La/Ce + 1.5 wt.% Cu is proposed, providing insights into the microstructure and engineering guidance for the alloy design and industrial application of high strength‐and‐toughness cast Al–Si–Mg alloys.

Allosteric regulation is a central mechanism for modulating biological functions and offers an attractive strategy in drug discovery, particularly for targets considered challenging or "undruggable." However, the discovery of allosteric inhibitors is hindered by poorly defined binding sites and the lack of effective screening approaches. Here, we present a dual DNA-encoded library (DEL) screening strategy that integrates reversible DEL and covalent DEL (CoDEL) technologies to identify novel allosteric inhibitors of the SARS-CoV-2 RNA-dependent RNA polymerase (RdRp). Using this approach, we discovered the first covalent allosteric inhibitors of RdRp, which engage a previously uncharacterized pocket on the nsp8 subunit and form a covalent bond with Cys114. Subsequent SAR studies and biochemical assays confirmed the allosteric mechanism and elucidated structural determinants of activity. This work highlights the power of integrating reversible DEL screening with CoDEL screening for ligand discovery and establishes a generalizable strategy to identify covalent allosteric modulators for therapeutically important targets for therapy or active probe design.

Authenticated key agreement (AKA) between vehicles and roadside units (RSUs) is a fundamental component for securing Vehicular ad-hoc networks (VANETs). However, most existing AKA schemes rely on conventional cryptographic primitives, such as discrete logarithm and large integer factorization, rendering them vulnerable to emerging quantum attacks. Although several post-quantum solutions based on lattice-hard problems have been proposed, their heavy dependence on a centralized trusted authority exposes them to single-point-of-failure risks. To address these issues, we propose a decentralized authenticated key agreement scheme grounded in the security assumptions of the Ring Learning With Errors (RLWE) and Inhomogeneous Small Integer Solution (ISIS) problems. The proposed scheme achieves strong resistance against quantum adversaries while preserving the privacy of communicating entities. Furthermore, a threshold-based voting mechanism is employed to mitigate the inherent key exposure problem within network infrastructure. By distributing authentication authority across multiple edge server nodes, the scheme significantly enhances system robustness against single-node failures and collusion attacks. Comprehensive security analysis and performance evaluations demonstrate that the proposed scheme satisfies diverse security requirements with acceptable computational and communication overhead, thereby confirming its practicality for deployment in real-world VANETs environments.

With the increasing prevalence of diffusion-based malicious image manipulation, existing proactive defense methods struggle to safeguard images against tampering under unknown conditions. To address this, we propose Anti-Inpainting, a proactive defense approach that achieves protection comprising three novel modules. First, we introduce a multi-level deep feature extractor to obtain intricate features from the diffusion denoising process, enhancing protective effectiveness. Second, we design a multi-scale, semantic-preserving data augmentation technique to enhance the transferability of adversarial perturbations across unknown conditions. Finally, we propose a selection-based distribution deviation optimization strategy to bolster protection against manipulations guided by diverse random seeds. Extensive experiments on InpaintGuardBench and CelebA-HQ demonstrate that Anti-Inpainting effectively defends against diffusion-based inpainters under unknown conditions. Additionally, our approach demonstrates robustness against various image purification methods and transferability across different diffusion model versions.

Against the background of new quality productivity, knowledge graphs, and AI technology have opened up new paths for higher education reform, promoting the transformation of traditional teaching models to “smart classrooms.” Taking the “Botany” course as the research object, this paper explores the path of constructing a blended online-offline teaching model using AI + knowledge graphs to address the problems of a fragmented knowledge system, static teaching resources, and a single learning path in traditional teaching. Based on the technological characteristics and educational application value of AI and knowledge graphs, a “Knowledge Graph + AI Scenario Practice + BOPPPS” model is proposed: a knowledge graph is constructed by extracting knowledge points and sorting out knowledge relationships to help students consolidate basic knowledge; scenario-based practice tasks are released through an AI course assistant to stimulate students’ learning interest, and full-process blended teaching activities of “pre-class—in-class—post-class” progressive exploration are carried out, aiming to enhance students’ independent learning, communication, and teamwork abilities. Research results show that this blended teaching model can effectively realize the systematic organization and visual presentation of botanical knowledge, significantly improving students’ learning efficiency and deep learning capabilities.

State space models (SSMs) have recently emerged as a powerful framework for long sequence processing, outperforming traditional methods on diverse benchmarks. Fundamentally, SSMs can generalize both recurrent and convolutional networks and have been shown to even capture key functions of biological systems. Here we report an approach to implement SSMs in energy-efficient compute-in-memory (CIM) hardware to achieve real-time, event-driven processing. Our work re-parameterizes the model to function with real-valued coefficients and shared decay constants, reducing the complexity of model mapping onto practical hardware systems. By leveraging device dynamics and diagonalized state transition parameters, the state evolution can be natively implemented in crossbar-based CIM systems combined with memristors exhibiting short-term memory effects. Through this algorithm and hardware co-design, we show the proposed system offers both high accuracy and high energy efficiency while supporting fully asynchronous processing for event-based vision and audio tasks.

Compute-in-memory (CIM) systems implemented with resistive random access memory (RRAM) crossbars are a promising approach for accelerating deep neural network (DNN) computations. However, it is noteworthy that RRAM-based CIM systems are susceptible to computational errors. Unlike digital computation, the nature of analog computing introduces the risk of error accumulation throughout the computation process. Various techniques have been proposed to help deal with the errors in CIM systems, among which, training methods to create noise-tolerant CIM-based DNNs (CIM-DNNs) models that are insensitive to weight variations are the most promising due to their simplicity and low implementation cost. Although promising empirical results of variation-aware training (VAT) showcasing DNN models with high tolerance to device nonidealities have been demonstrated, there remains a significant gap in the understanding of noise tolerance properties in VAT-trained CIM-DNNs and how to improve VAT based on these understandings. The exploration of these theoretical aspects represents an area requiring further investigation and research. This work endeavors to explore the fundamental properties of noise tolerance in DNNs for CIM systems. We encapsulate our contributions into three key points. First, we identify factors that influence DNNs' performance when subjected to noise through a series of training experiments. Second, we offer both theoretical insights and practical demonstrations illustrating how VAT operates to yield solutions with heightened resistance to noise during the training process. Finally, leveraging these insights, we provide guidelines for implementing VAT to obtain optimal noise tolerance in CIM-DNNs. Our objective is to establish a theoretical foundation for VAT, and building on these insights, we aim to offer general and straightforward guidelines for DNN training, experimenting with factors such as hyperparameter choices for optimizers and weight clamping. Ultimately, our aim is to contribute to general and practical solutions for the development of reliable CIM systems. Our studies focus on analyzing how noise injection and different optimizers affect the convergence dynamics during training to reach a more noise-tolerant solution through VAT. Future studies could incorporate advanced regularizers reflecting the flatness of the solution into the cost function, which may be necessary for models beyond DNNs studied here. Combined together, these techniques can potentially lead to practical solutions for the development of reliable CIM systems.

Biphasic calcium phosphate (BCP) composites are widely employed for bone repair, with osteogenic performance governed by both the HA:β-TCP ratio and crystallinity. While compositional tuning has been extensively investigated, the role of crystallinity as a multiscale regulator of composite properties remains insufficiently explored. Here, BCP composites with controlled crystallinity were fabricated by adjusting calcination temperature, and systematically characterized across multiple length scales. Reducing crystallinity decreased calcium ion release but enhanced surface roughness, hydrophilicity, and protein adsorption, while maintaining mechanical competence sufficient for load-bearing environments. These physicochemical changes synergistically promoted cytoskeletal extension, integrin-mediated signaling, and osteogenic gene expression in bone marrow stromal cells. Among the tested BCP composites, BCP2, with moderately low crystallinity, showed the best osteogenic outcomes both in vitro and in vivo, due to its balanced combination of favorable surface properties and moderate ion release. In contrast, BCP1, with the lowest crystallinity and superior surface properties, exhibited slightly lower osteogenesis than BCP2, likely due to insufficient calcium release. BCP3, although highly crystalline with abundant calcium release, showed the lowest osteogenic performance, possibly due to insufficient stimulation of surface characteristics during cell-material interactions in vitro and in vivo. Transcriptomic profiling further confirmed that BCP2 activated integrin-mediated cytoskeletal pathways to drive osteogenesis. Overall, this study identifies crystallinity as a key tunable design parameter that indirectly regulates osteogenesis through its coupled effects on surface and ionic cues, providing strategic insights for the rational development of BCP composites for bone regeneration.

The rapid proliferation of Internet of Medical Things (IoMT) devices has introduced new security challenges, particularly susceptibility to reconnaissance (recon) attacks that often precede more severe intrusions. In this paper, we present a lightweight, preprocessing-aware baseline for detecting recon attacks in IoMT networks using Logistic Regression. Leveraging the CICIoMT2024 dataset, we investigate the effects of various data cleaning strategies, including missing value imputation, duplicate removal, and outlier filtering, on classification performance. We evaluate both full and reduced feature sets, focusing on traffic characteristics such as packet load, byte rate, and flow duration. Experiments are conducted on three dataset variants with increasing levels of preprocessing, and model performance is assessed across binary (benign vs. recon) and multi-class (benign +4 attack types) settings. Our findings show that Logistic Regression, when combined with appropriate preprocessing and feature selection, achieves competitive accuracy in binary classification, reaching up to 99 %. However, performance degrades significantly in multi-class tasks due to data imbalance. To enhance the interpretability and portability of the system, we further developed a mobile application using MIT App Inventor that visualizes key traffic features and model predictions in real time. This study provides a transparent and reproducible benchmark for recon detection in IoMT networks, highlighting key considerations in preprocessing, feature selection, and model limitations that inform future work on secure, resourceconstrained medical systems.

Deepfakes pose significant societal risks, motivating the development of proactive defenses that embed adversarial perturbations in facial images to prevent manipulation. However, in this paper, we show that these preemptive defenses often lack robustness and reliability. We propose a novel approach, Low-Rank Adaptation (LoRA) patching, which injects a plug-and-play LoRA patch into Deepfake generators to bypass state-of-the-art defenses. A learnable gating mechanism adaptively controls the effect of the LoRA patch and prevents gradient explosions during fine-tuning. We also introduce a Multi-Modal Feature Alignment (MMFA) loss, encouraging the features of adversarial outputs to align with those of the desired outputs at the semantic level. Beyond bypassing, we present defensive LoRA patching, embedding visible warnings in the outputs as a complementary solution to mitigate this newly identified security vulnerability. With only 1,000 facial examples and a single epoch of fine-tuning, LoRA patching successfully defeats multiple proactive defenses. These results reveal a critical weakness in current paradigms and underscore the need for more robust Deepfake defense strategies. Our code is available at https://github.com/ZOMIN28/LoRA-Patching.

A blockchain and threshold fully homomorphic encryption (TFHE)-based federated learning framework is proposed to defend against model poisoning and privacy leakage. Standard federated learning (FL) remains vulnerable to inference and poisoning attacks during model aggregation and distribution. To address these threats, we present a secure and efficient FL framework that combines TFHE with Shamir’s secret sharing for privacy-preserving encrypted updates. A multi-layer verification mechanism consisting of verifiable secret sharing (PVSS) and cosine similarity checks is introduced to identify and filter out malicious updates. Furthermore, we replace the central aggregator with a blockchain-based mechanism to eliminate single-point trust. Extensive experiments conducted on MNIST and CIFAR-10 under IID and Non-IID settings demonstrate that our scheme maintains high model accuracy under 20%–40% label-flipping attacks. Compared with baseline methods such as threshold Paillier and PPVFL, our approach reduces encryption and decryption overhead by up to 28.4%, while achieving a lower accuracy drop of only 2.32% on MNIST and 7.44% on CIFAR-10 under 40% poisoning. These results highlight the robustness, efficiency, and practicality of the proposed scheme for real-world federated learning deployments.

Logical model is the key to realize the integration of data information and physics, and at present, most of the research focuses on concepts, and less research is carried out on the application scenarios of digital structure construction and 3D physical model association and digital twin modeling. In view of the substation logic model, through the description of the physical ability of the expression, the overall framework of the logic model, the structure of the technical model, the element composition, the coding form and other structural designs are proposed, and the logic model structure and classification method are proposed based on the different main wiring forms of the substation, relying on the specific engineering, the twin modeling application scenario of the logic model is developed, and the integration and association of the logic model and the 3D physical model are carried out according to the logic model digital platform, so as to realize the modeling application for the expression of multiple information in the design process. It provides the basis for the bidirectional data connection between the physical entity data and the virtual model in the subsequent substation.

A grand challenge in modern neuroscience is to bridge the gap between the detailed mapping of microscale neural circuits and mechanistic understanding of cognitive functions. While extensive knowledge exists about neuronal connectivity and biophysics, how these low-level phenomena eventually produce abstract behaviors remains largely unresolved. Here, we propose that a model based on State Space Models, an emerging class of deep learning architectures, can be a potential biological model for analysis. We suggest that the differential equations governing elements in a State Space Model are conceptually consistent with the dynamics of biophysical processes, while the model offers a scalable framework to build on the dynamics to produce emergent behaviors observed in experimental neuroscience. We test this model by training a network employing a diagonal state transition matrix on temporal discrimination tasks with reinforcement learning. Our results suggest that neural behaviors such as time cells naturally emerge from two fundamental principles: optimal pre-configuration and rotational dynamics. These features are shown mathematically to optimize history compression, and naturally generate structured temporal dynamics even prior to training, mirroring recent findings in biological circuits. We show that learning acts primarily as a selection mechanism that fine-tunes these pre-configured oscillatory modes, rather than constructing temporal codes de novo. The model can be readily scaled to abstract cognitive functions such as event counting, supporting the use of State Space Models as a computationally tractable framework for understanding neural activities.

Multimodal fusion is susceptible to modality imbalance, where dominant modalities overshadow weak ones, easily leading to biased learning and suboptimal fusion, especially for incomplete modality conditions. To address this problem, we propose a Shapley-guided alternating training framework that adaptively prioritizes minor modalities to balance and thus enhance the fusion. Our method leverages Shapley Value-based scheduling to improve the training sequence adaptively, ensuring that under-optimized modalities receive sufficient learning. Additionally, we introduce the memory module to refine and inherit modality-specific representations with a cross-modal mapping mechanism to align features at both the feature and sample levels. To further validate the adaptability of the proposed approach, the encoder module empirically adopts both conventional and LLM-based backbones. With building up a novel multimodal equilibrium metric, namely, equilibrium deviation metric (EDM), we evaluate the performance in both balance and accuracy across four multimodal benchmark datasets, where our method achieves state-of-the-art (SOTA) results. Meanwhile, robustness analysis under missing modalities highlights its strong generalization capabilities. Accordingly, our findings reveal the untapped potential of alternating training, demonstrating that strategic modality prioritization fundamentally balances and promotes multimodal learning, offering a new paradigm for optimizing multimodal training dynamics.

Accurate detection of road and bridge changes is crucial for urban planning and transportation management, yet presents unique challenges for general change detection (CD). Key difficulties arise from maintaining the continuity of roads and bridges as linear structures and disambiguating visually similar land covers (e.g., road construction vs. bare land). Existing spatial-domain models struggle with these issues, further hindered by the lack of specialized, semantically rich datasets. To fill these gaps, we introduce the Road and Bridge Semantic Change Detection (RB-SCD) dataset. As the first benchmark to systematically target semantic change detection of roads and bridges, RB-SCD offers comprehensive fine-grained annotations for 11 semantic change categories. This enables a detailed analysis of traffic infrastructure evolution. Building on this, we propose a novel framework, the Multimodal Frequency-Driven Change Detector (MFDCD). MFDCD integrates multimodal features in the frequency domain through two key components: (1) the Dynamic Frequency Coupler (DFC), which leverages wavelet transform to decompose visual features, enabling it to robustly model the continuity of linear transitions; and (2) the Textual Frequency Filter (TFF), which encodes semantic priors into frequency-domain graphs and applies filter banks to align them with visual features, resolving semantic ambiguities. Experiments demonstrate the state-of-the-art performance of MFDCD on RB-SCD and three public CD datasets. The code will be available at https://github.com/DaGuangDaGuang/RB-SCD.

Diffusion Models (DMs) have achieved remarkable success in realistic voice cloning (VC), while they also increase the risk of malicious misuse. Existing proactive defenses designed for traditional VC models aim to disrupt the forgery process, but they have been proven incompatible with DMs due to the intricate generative mechanisms of diffusion. To bridge this gap, we introduce VoiceCloak, a multi-dimensional proactive defense framework with the goal of obfuscating speaker identity and degrading perceptual quality in potential unauthorized VC. To achieve these goals, we conduct a focused analysis to identify specific vulnerabilities within DMs, allowing VoiceCloak to disrupt the cloning process by introducing adversarial perturbations into the reference audio. Specifically, to obfuscate speaker identity, VoiceCloak first targets speaker identity by distorting representation learning embeddings to maximize identity variation, which is guided by auditory perception principles. Additionally, VoiceCloak disrupts crucial conditional guidance processes, particularly attention context, thereby preventing the alignment of vocal characteristics that are essential for achieving convincing cloning. Then, to address the second objective, VoiceCloak introduces score magnitude amplification to actively steer the reverse trajectory away from the generation of high-quality speech. Noise-guided semantic corruption is further employed to disrupt structural speech semantics captured by DMs, degrading output quality. Extensive experiments highlight VoiceCloak's outstanding defense success rate against unauthorized diffusion-based voice cloning. Audio samples of VoiceCloak are available at https://voice-cloak.github.io/VoiceCloak/.

Abstract Strong electron-electron interaction can induce Mott insulating state, which is believed to host unusual correlated phenomena such as quantum spin liquid when quantum fluctuation dominates and unconventional superconductivity through doping. Transition metal compounds as correlated materials provide a versatile platform to engineer the Mott insulating state. Previous studies mostly focused on the controlling of the repulsive interaction and bandwidth of the electrons by gating or doping. Here, we performed angle-resolved photoemission spectroscopy (ARPES) on monolayer 1T phase NbSe 2 , TaSe 2 , and TaS 2 and directly observed their band structures with characteristic lower Hubbard bands. By systematically investigating the orbital textures and temperature dependence of the energy gap of the materials in this family, we discovered that hybridization of the chalcogen p states with lower Hubbard band stabilizes the Mott phase via tuning of the bandwidth, as shown by a significant increase of the transition temperature ( T C ) at a stronger hybridization strength. Our findings reveal a mechanism for realizing a robust Mott insulating phase and establish monolayer 1T phase transition metal dichalcogenide family as a promising platform for exploring correlated electron problems.

Based on the theory of life cycle cost (LCC), the economic, safety and social benefits of 110kV outdoor HGIS and GIS substation are comprehensively evaluated in order to optimise the selection decision of power distribution equipment. The LCC mathematical model is constructed, and the whole life cycle cost (LCC) of the two options is quantitatively measured and comparatively analysed in terms of initial investment, operation, maintenance, failure and decommissioning disposal. The results show that the initial investment cost of the HGIS scheme is slightly higher than that of the GIS, but its operation cost is basically the same, the maintenance and fault repair cost is significantly lower than that of the GIS, and the decommissioning cost is also more advantageous. Within 40 years of operation, the whole life cycle cost of the HGIS scheme is lower than that of the GIS scheme by about 3 million RMB, which shows a better economic performance. The study shows that the application of LCC calculation method in the optimisation of general design scheme of power distribution equipment is of great value, which can provide a scientific basis for the economic decision-making of electric power engineering projects, improve the investment efficiency and operation efficiency of power grid construction, and promote the sustainable development of the industry.

Traditional pain assessment methods heavily rely on patient self-reporting, which is often unreliable because different people have different pain baselines, and some people may not even have verbal communication ability. To address this limitation, researchers utilize facial expressions to decipher the pain level. Unfortunately, the accuracy of such methods is challenged by many factors, such as the angle between the subject’s face and the camera, the interference from the subject’s surrounding environment, and the mixed motion of the subject. To improve the accuracy and automate the detection of pain levels, this work proposes a hardware-software collaborative framework that employs multiple emotion recognition models to perform explainable machine learning algorithms for pain detection. The emotion recognition based on the Deepface learning model facilitates us to identify the top five Action Units (AUs) that are highly related to pain. Our case studies with the MIntPain database show that the proposed method can cause the false negative rate to diminish to zero. Our experimental results also indicate that the Pearson correlation coefficient between emotion categories and the pain level is not a reliable metric to distinguish a specific pain level.

Circular RNAs (circRNAs) are non-coding RNAs with covalently closed loop structures that participate in various biological processes. However, the functions of many circRNAs remain unclear. Endothelial cell dysfunction, which involves abnormal ferroptosis, a unique form of regulated cell death, is a characteristic of various diseases. However, the mechanisms governing ferroptosis in endothelial cells are not fully understood. Here, we investigated the impact of a novel circRNA, circ-DAPK1, on ferroptosis in human umbilical vein endothelial cells (HUVECs) under high-glucose conditions. Our data showed that high-glucose conditions upregulate circ-DAPK1 expression in HUVECs. Overexpression of circ-DAPK1 induced ferroptosis in HUVECs, whereas depletion of circ-DAPK1 mitigated the ferroptosis triggered by high-glucose treatment. Inhibition of ferroptosis reversed the decrease in cell viability induced by high glucose or circ-DAPK1 overexpression. Using RNA immunoprecipitation analyses, we identified several ferroptosis-regulating proteins, including NAD(P)H dehydrogenase [quinone] 1 (NQO1) and insulin-like growth factor 2 mRNA binding protein 2 (IGF2BP2). Mechanistically, circ-DAPK1 interacts with NQO1, enhancing its ubiquitination and accelerating its degradation. NQO1 overexpression partially rescues HUVECs from high-glucose-induced ferroptosis. We also found that IGF2BP2 binds to the m6A site on circ-DAPK1. Depletion of IGF2BP2 in HUVECs reduced circ-DAPK1 expression and inhibited high-glucose-induced ferroptosis. These findings reveal the effects of the IGF2BP2-circ-DAPK1 axis in regulating ferroptosis in HUVECs under high-glucose conditions and extend our understanding of the mechanisms controlling ferroptosis in endothelial cells.

Fully homomorphic encryption (FHE) is a powerful cryptographic technique that enables computation on encrypted data without needing to decrypt it. It has broad applications in scenarios where sensitive data needs to be processed in the cloud or in other untrusted environments. FHE applications are both compute- and memory-intensive, owing to expensive operations on large data. While prior works address the challenges of efficient compute using dedicated hardware, expensive memory transfers still remain a major limiting factor. In this work, we propose a hierarchical near-DRAM processing (NDP) solution for FHE applications, called FHENDI, that harnesses the massive DRAM bank bandwidth. We observe various data access patterns in FHE that reveal distinct levels of parallelism: element-wise, limb-wise, coefficient-wise, and ciphertext-wise. FHENDI exploits these levels of parallelism to map FHE operations and data onto different hierarchies of our design, while addressing three major challenges with NDP for FHE: (i) the lack of bank-to-bank communication support, (ii) limited die-to-die bandwidth, and (iii) large memory access latencies. We resolve the first problem through a novel, conflict-free mapping algorithm built atop localized permutation networks that enables efficient element-wise and butterfly operations in FHE. The second problem is addressed by pipelining the execution of parallel bootstrap operations observed in compiled FHE workloads. Finally, we hide the memory access latency behind computation latency by exploiting a dual-banking scheme and subarray-level parallelism (SLP) of the DRAM banks. We evaluate FHENDI using representative workloads in the domains of privacy-preserving machine learning inference on CNNs and Transformers, database range query, and sorting, that are obtained using a compiler framework called HElayers. We compare FHENDI with a server-class CPU and GPU running the state-of-the-art HEaaN library, and an FHE accelerator ASIC, and report mean speedups of $2145.8 \times, 118.29 \times$, and $2.45 \times$, respectively.

Analog compute in memory (CIM) with multilevel cell (MLC) resistive random access memory (ReRAM) promises highly dense and efficient compute support for machine learning and scientific computing. This article introduces analog to digital converter (ADC)-assisted bit-serial processing for efficient, high-throughput compute. Bit-serial digital to analog converters (DACs) and 8-bit binary-weighted multicycle sampling (BWMCS) ADCs perform analog vector-matrix multiplication (VMM) on MLC-based crossbar arrays. A direct drive <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${g}_{m}$ </tex-math></inline-formula>-boosted transimpedance amplifier (TIA) enables high-speed crossbar readout. We present a system on chip (SoC) prototype consisting of four self-contained ReRAM-based CIM macros and a reduced instruction set computer-five (RISC-V) host. The test chip is fabricated in 65 nm CMOS with foundry-integrated MLC ReRAM. We trained LeNet1 for handwritten digit classification and mapped the CNN weights differentially to 3-bit MLC ReRAM across multiple CIM macros. The classification accuracy loss is 1.6% when compared to the quantization-aware trained model. The measured raw and normalized peak efficiencies are 20.7 and 662 TOPS/W, respectively. The compute density is 8.4 TOPS/mm2.

Abstract Objectives/Scope The manuscript proposes an agent framework for elder (existed) well re-assessment. Through the allied application with different generative artificial intelligence (AIGC) models, the self-supervised feedback capability of the framework with a growing scheme, continuously stimulates the reasoning abilities of large AIGC models is proposed. With less rely on fine-tuning, the approach proposed aims to realize the precision improvement of curve re-construction, strata prediction, and potential layer interpretation for existed tight gas wells, especially in terms of dig deeper into even detail scenes. Methods, Procedures, Process The manuscript proposes a upgrade version of self supervised agent framework for purpose of assisting existed well re-evaluation among northwestern oil&amp;gas field of China. Based on explainable self-supervised agent group framework (ESSA) for putting-up academic mechanisms, business models and AI components altogether, an autonomous self-supervised agent framework (CISFA) along with generative AI large-model (AIGC-LLM) was introduced to construct copilot for re-evaluation activities. Upgraded from former contributions, a growing scheme with original designed attention spreading scheme is conducted in this manuscript, thereby growing "superior individuals" for in-depth analysis among specific tasks, namely, curve prediction, strata classification and potential useful layer recommendation. Originality The manuscript demonstrates the innovation of AI-native architecture with following originality: (1) The AI-native application jointly formed by the agent cluster and AIGC possesses the capability to address multi-scenarios cascading problems. (2) Through the multi-agent framework, a closed-loop agent cluster with growing scheme combined with attention spreading mechanism is constructed, which realizes adaptive-learning and improvement among collective intelligence through accurate model selection. This approach continuously optimizes the performance of re-evaluation of existed wells, significantly reducing the ongoing operational consumption of the model in practical applications. Results 50+25 existed wells among 2 various blocks, all with 30+ years, located in two distinct blocks were taken into account. Utilizing the framework introduced in this manuscript, we evaluated the accuracy of model selection, the effectiveness of well logging curve reconstruction, the classification of strata (oil, water, gas, sandstone, mudstone, and oil-water coexistence), and the capability of recommending potential layers. The experimental results demonstrate that the proposed framework possesses feasible agent based "Tool-using" capabilities. Through self-supervised prompting engineering provided by the agent framework proposed in this manuscript, it can continuously unleash the potential of AIGC models, resulting in satisfactory interpretations of potential layers, with additional agent growing scheme performed further accurate outcomes in detailed existed-well assessment tasks.

Enhancing photoluminescence (PL) efficiency in colloidal quantum dots is pivotal for next-generation near-infrared photodetectors, imaging systems, and photonic devices. Conventional methods, especially metal-based plasmonic structures, suffer from large optical losses, which limits their practical use. Here, we introduce a quasi-bound state in the continuum (quasi-BIC) metasurface on a silicon-on-insulator platform, tailored to provide high-quality factor resonances with minimized losses. Utilizing topological charge engineering and controlled in-plane asymmetry in silicon cylinder arrays, we developed a robust quasi-BIC capable of maintaining a high Q factor across a broad angular range, achieving an experimental Q factor of 3031 at normal incidence. This approach significantly enhances near-field interactions, achieving a ≤110-fold increase in PL for PbS quantum dots at 33 K and a 41-fold enhancement at room temperature. Our findings offer a scalable, cost-effective solution for enhancing light emission in advanced optoelectronic applications.