To enhance customized terrestrial communications and support diverse independent emergency services, uncrewed aerial vehicle (UAV) network slicing is regarded as a feasible solution. However, UAV network slicing suffers from UAV's limited resources and imperfect channel state information (CSI) due to channel errors. In this paper, we propose a robust resource optimization framework with channel uncertainty for UAV network slicing, by introducing a hybrid quantum neural network (HQNN)-based multi-agent deep reinforcement learning (MADRL) algorithm. Specifically, to cope with limited UAV resources, inspired by Open Radio Access Network (RAN), we decompose users' baseband processing functions and dynamically distribute them across different UAV sub-networks. Then, considering channel uncertainty, we formulate a quantile-based delay minimization problem under an outage probability constraint. This problem jointly optimizes function splitting (FS), user association, and UAV power allocation. To obtain robust resource optimization, we customize a HQNN-based multi-agent dueling double deep Q-network (HQMAD3QN). It perfectly integrates the mature techniques of MADRL for processing multiple elements and the advanced computational capabilities of HQNNs coupling the quantum neural network with the convolutional neural network. Notably, we utilize a Monte Carlo (MC)-based data augmentation module to address the difficult outage probability constraint. Numerical results show that the proposed method outperforms state-of-the-art approaches in terms of service delay, robustness, constraint violation, and computational complexity.

RAN slicing has been widely studied for providing ultra-reliability low-latency communication (URLLC), enhanced mobile broadband (eMBB), and massive machine type communication (mMTC) services in 5G. However, the existing RAN slicing networks have not been explored to support the hybrid services, such as massive URLLC (mULC) and ubiquitous eMBB (uMBB) services. In this paper, we propose a novel rate splitting multi-access (RSMA)-enabled RAN slicing system to facilitate the runtime support of mULC and uMBB services. Firstly, three typical slices, i.e., URLLC, eMBB, and mMTC slices are constructed. Then, a multi-connection scheme is proposed by using RSMA technology, i.e., the users can be connected with two typical slices to obtain mULC and uMBB services. Specifically, the transmitted data of each mULC/uMBB user will be split into the common mMTC data and the private URLLC/eMBB data, which will be encoded into the corresponding traffic flows and served by corresponding slices. Next, a system-wide utility optimization problem is proposed to optimize heterogeneous requirements for mULC and uMBB services by joint user grouping, bandwidth allocation, and power control. Finally, a two independent agent DDPG (2IADDPG) algorithm is customized to solve the formulated problem, wherein two independent agents are responsible for independent decision-making. The reported numerical results show that the RSMA scheme outperforms the benchmarks, and in the meanwhile our proposed 2IADDPG algorithm can achieve faster convergence rate compared with the multi-agent DDPG algorithm and other comparison algorithms.

Integrating Unmanned Aerial Vehicles (UAVs) into Mobile Edge Computing (MEC) systems offers a promising solution to enhance coverage and flexibility in 6G edge networks. However, the dynamic and resource-limited nature of UAVs introduces strong cross-layer coupling between UAV-Access Point (AP) coordination and AP-user scheduling, making efficient resource allocation highly challenging. To address this issue, this paper proposes a unified two-timescale hierarchical Deep Reinforcement Learning (DRL) framework that jointly optimizes channel utilization and transmission latency. Unlike conventional hierarchical decompositions that separate large- and small-timescale decisions, the proposed method models UAV-AP interactions as an integrated optimization process to preserve global consistency and avoid local optima. A dual-layer architecture is designed, where the UAV-level agent performs large-timescale bandwidth allocation and the AP-level agents conduct fine-grained small-timescale scheduling. To mitigate the exponential complexity caused by nested DRL updates, an asynchronous cross-timescale coordination mechanism is introduced, enabling stable convergence and efficient learning. Simulation results demonstrate that the proposed framework achieves superior performance in channel utilization, delay reduction, and robustness under dynamic traffic conditions compared with single-agent, prediction-based, and heuristic baselines.

Large language model (LLM) inference at the network edge is a promising serving paradigm that leverages distributed edge resources to run inference near users and enhance privacy. Existing edge-based LLM inference systems typically adopt autoregressive decoding (AD), which only generates one token per forward pass. This iterative process, compounded by the limited computational resources of edge nodes, results in high serving latency and constrains the system's ability to support multiple users under growing demands.To address these challenges, we propose a speculative decoding (SD)-based LLM serving framework that deploys small and large models across heterogeneous edge nodes to collaboratively deliver inference services. Specifically, the small model rapidly generates draft tokens that the large model verifies in parallel, enabling multi-token generation per forward pass and thus reducing serving latency. To improve resource utilization of edge nodes, we incorporate pipeline parallelism to overlap drafting and verification across multiple inference tasks. Based on this framework, we analyze and derive a comprehensive latency model incorporating both communication and inference latency. Then, we formulate a joint optimization problem for speculation length, task batching, and wireless communication resource allocation to minimize total serving latency. To address this problem, we derive the closed-form solutions for wireless communication resource allocation, and develop a dynamic programming algorithm for joint batching and speculation control strategies. Experimental results demonstrate that the proposed framework achieves lower serving latency compared to AD-based serving systems. In addition,the proposed joint optimization method delivers up to 44.9

Fluid antenna systems (FAS) have emerged as a promising enabling technology for sixth-generation (6G) wireless networks. By dynamically switching ports according to channel characteristics, FAS can exploit spatial diversity and significantly enhance wireless performance. However, most existing studies focus on maximizing instantaneous throughput while neglecting the long-term cost associated with frequent port switching. This work aims to jointly optimize long-term throughput and port-switching cost in FAS, considering spatio-temporally correlated channels and minimum-rate requirements for each user. To address this challenge, we formulate the dynamic port selection problem as a Markov decision process (MDP) and employ a dueling DQN framework to obtain a near-optimal policy. A Transformer encoder is incorporated to capture spatial-temporal dependencies among ports, and a candidate pre-selection mechanism is introduced to reduce training complexity. Simulation results show that the proposed method achieves up to 96% of the per-slot exhaustive-search optimum while significantly reducing decision latency, demonstrating high efficiency and scalability for large-scale FAS deployments.

With increasing demands for ultra-reliable, low-latency applications and next-generation network services, integrating artificial intelligence and machine learning (AI/ML) into Open Radio Access Network (O-RAN) components has become a critical research focus. However, realizing the full potential of AI/ML in O-RAN presents unresolved challenges due to the absence of system-level mechanisms for dynamic resource allocation and limited coordination among the functionally separated components. The paper addresses some of these challenges by proposing a bottleneck-based deep learning-driven resource allocation approach that employs a Gated Recurrent Unit (GRU)-based forecasting model to proactively identify and mitigate bottleneck resources, enabling the system to adapt to fluctuating user demands and varying network conditions, and guiding task reallocation through policy-driven decisions. Our approach combines the capabilities of the Non-Real-Time (Non-RT) and Near-Real-Time (Near-RT) RAN Intelligent Controllers (RICs) across the cloud-edge continuum. Since edge computing nodes often have limited resources and are more expensive compared to cloud infrastructure, components of the Near-RT RIC are deployed at the edge, while Non-RT RIC components are placed in the cloud. We implement this framework in both xApp and rApp forms, fully compliant with O-RAN specifications, and conduct extensive performance evaluations using real-world network data in an extended Kubernetes environment, demonstrating the integration of Near-RT RIC at the edge and Non-RT RIC in the cloud. Comprehensive performance evaluations conducted on the O-RAN Software Community (OSC) testbed demonstrate significant improvements in network efficiency, scalability, and latency, as the proposed approach significantly outperforms existing methods by reducing resource utilization by 14%-40%, reducing task delay by 21.6%-44.0%, and achieving an admittance ratio improvement ranging from 6.48% to 16.6% compared to other approaches.

Service-based architecture (SBA) has long been recognized as a key enabler of the 5G core network in supporting a wide variety of use cases with challenging requirements, both in the Internet of Things (IoT) and across vertical industries. However, the evolution toward 6G reveals inherent limitations of existing SBA-based core networks, which are primarily designed for static service provisioning. As a consequence, artificial intelligence (AI) is often introduced as an external add-on rather than being natively embedded into the core network architecture, fundamentally limiting the ability to support intelligent, adaptive, and service-aware networking. Furthermore, AI is increasingly becoming a native capability, rather than an external “plug-in” function, in nextgeneration networks. In this context, SBA and AI are natively integrated into an open-source core network (OpenCN). First, we present a novel holistic-service-based OpenCN architecture that incorporates native AI as well as a management and orchestration (MANO) plane. The servicebased MANO plane is designed according to two core principles: (i) decoupling monolithic OpenCN components into independent functions and resources, and (ii) reconfiguring the required functions and resources to construct a customized OpenCN instance tailored to users’ specific requirements. In this manner, each user can be provisioned with a personalized service. Second, we evaluate the performance of the proposed OpenCN framework using a smallscale test network. The results demonstrate that OpenCN achieves higher bandwidth utilization, more efficient resource allocation, and lower energy consumption compared with benchmark solutions. Finally, we discuss several promising research directions for next-generation OpenCNs.

With the proliferation of large language models (LLMs), cloud-based LLM serving mechanisms may cause network congestion and high serving delay. Edge computing offers a solution to alleviate backhaul pressure and reduce serving delay by deploying LLMs on edge servers and providing LLM inference services in users’ proximity. However, user accuracy requirements vary over time, and mismatches between these requirements and the deployed LLMs at the edge may lead to inefficient resource usage and increased serving delay. To address this, we formulate a joint LLM caching, inference task scheduling, and network resource allocation problem to minimize LLM serving delay under unknown time-varying user accuracy requirements. To solve the problem, we first derive closed-form solutions for optimal computation and communication resource allocation under any LLM caching and task scheduling policies. Then, we employ an improved branch-and-bound algorithm to obtain optimal task scheduling policies under any LLM caching strategies. Finally, we propose an improved double deep Q-network (DDQN)-based algorithm to determine the LLM caching decisions. It incorporates a state coding and action aggregation (SCAA) mechanism within the deep neural networks (DNNs) of the traditional DDQN. The SCAA-DNNs involve an input-layer gating mechanism to encode users’ request states for LLMs and a two-layer output architecture that dynamically aggregates LLM caching actions to generate the corresponding state-action values, thereby improving learning efficiency and accelerating convergence in large discrete action spaces. Experimental results show that the proposed scheme could rapidly converge and reduce average user delay by up to 20.8% compared to benchmarks.

In ultra-dense multi-access edge computing (MEC), efficient task offloading and resource allocation are critical to meeting stringent delay and energy constraints. However, the rapid growth in terminal devices poses a significant challenge to traditional deep reinforcement learning (DRL)-based methods, which struggle to maintain efficient task offloading and resource allocation due to increased computational complexity and decision latency. To address this challenge, in this paper, we propose a mean field theory (MFT)-guided DRL approach which leverages the statistical characteristics of a large population of terminal devices to reduce the computational complexity of decision-making. Firstly, we formulate the joint task offloading and resource allocation problem as a mean field Markov decision process (MFMDP) with the objective of minimizing the overall system energy consumption while satisfying the task delay requirements and resource constraints. Secondly, by leveraging the inherent structure of the MFMDP, which represents the system dynamics using state and action distributions rather than joint action spaces, we achieve significant dimensionality reduction and improved scalability. Thirdly, we prove that the Q-function of MFMDP satisfies the fixed point theory, and considering the continuity of task offloading and resource allocation variables, we develop an MFT-guided deep deterministic policy gradient (MFT-DDPG) algorithm to solve the proposed problem. Experimental results show that MFT-DDPG significantly outperforms conventional DRL baselines in convergence speed and scalability. Specifically, for 50 terminal devices, the training time of proposed MFT-DDPG is reduced by up to 86.1%, 92.1%, and 94.6% compared to those of MADDPG, QMIX, and DDPG, respectively. In comparison with the Edge Server Computing Only (ECO) scheme and Device and Edge Server Collaborative Computing (DECC) scheme, the proposed method consistently yields lower energy consumption and delay across varying numbers of terminal devices, transmission power, and edge server computing resources compared to benchmark solutions, demonstrating its robustness and effectiveness in ultra-dense MEC scenarios.

With the rapid development of artificial intelligence and the dramatic growth of communication services, the sixth-generation (6G) wireless network needs to handle communication tasks more flexibly and efficiently, significantly exacerbating the challenge of resource allocation. For the access network scenarios in 6G networks, the existing single-layer reinforcement learning resource allocation algorithms are hard to satisfy the diverse demands of users due to the complex and variable state space. Therefore, we propose a reinforcement learning-based two-timescale resource allocation scheme, aiming to jointly enhance the quality of service and system resource utilization. The proposed method comprises an upper-layer controller that allocates network resources to lower-layer controllers on a large time scale. Then, lower-layer controllers refine the resources based on user service types on a smaller time scale. To implement the proposed two-timescale allocation scheme, we propose a two-layer reinforcement learning framework consisting of a deep deterministic policy gradient (DDPG) and a dueling deep Q network (Dueling-DQN). Furthermore, recognizing that coupling multiple reinforcement learning processes may slow down algorithm convergence, we employ asynchronous training, transfer learning, and prediction-based action space simplification to expedite the model’s convergence speed. Finally, we build a prototyping network to verify the performance of the proposed small-timescale and the large-timescale allocation algorithms. Our proposed scheme demonstrates significant improvements in both resource utilization and quality of service compared to existing schemes.

In this letter, we investigate an opportunistic fluid antenna multiple access (O-FAMA) system in a multiuser downlink channel. A multi-antenna base station (BS) communicates to many user equipments (UEs) on the same time-frequency resource block without precoding. Each UE decides on its own whether it should access the channel at any given time. If active, it utilizes solely its fluid antenna to overcome the inter-user interference. This setup is designed to combine decentralized opportunistic scheduling and FAMA for sum-rate maximization. Specifically, the joint optimization of opportunistic UE scheduling and port selection is prohibitively complex even if it is performed in a centralized fashion. To address this challenge, we propose a low-complexity mean field reinforcement learning approach, for each UE to select its best port and make scheduling decision based on the mean field behavior of other UEs and its own partially observable signal-to-interference-plus-noise ratio (SINR) at the ports. Numerical results show that the proposed method reduces training time by 95.7% when but improves sum-rate by 12-14% over MA SD3-DerivNet.

Network slicing promotes the development of different industries by dividing multiple logical networks on the same physical network to provide the customized services. However, the complex environment of network slicing deployment makes it difficult to obtain the accurate mathematical models, so that the traditional rule-based heuristic algorithms are difficult to process them efficiently. Therefore, we combine the Graph Convolutional Networks (GCN) and the Deep Deterministic Policy Gradient (DDPG) algorithms and propose the GCN-DDPG (G-DDPG) algorithm in this paper to solve the end-to-end network slicing deployment problem, while taking into account the constraints of Virtual Network Function (VNF) placement, VNF sharing, tolerable latency, and node and link resources limitations. First, the end-to-end network slicing deployment optimization is formulated as a problem of maximizing the weighted sum of system resource utilization and acceptance rate. Second, the physical network features extracted by GCN are combined with the state information of end-to-end network slicing requests as the state space of the optimization problem, and a G-DDPG algorithm is proposed to solve it. Finally, the simulation results demonstrate that our proposed method superior to benchmark solutions in terms of the resource utilization of the system, acceptance rate of end-to-end network slicing.

For the first time, we transplant the Quantum Approximate Optimization Algorithm (QAOA) from quantum computing labs to the radio links that connect smart meters inside power grids, reframing Non-Orthogonal Multiple Access (NOMA) resource allocation as a combinatorial task within a multi-carrier downlink. The search for optimal pairs is then delegated to QAOA, whose superposition and entanglement resources explore the solution space far more compactly than any classical heuristic. An Improved-QAOA (I-QAOA) algorithm is also proposed, which dynamically adjusts user grouping based on channel state information, significantly enhancing spectral efficiency. Simulation results demonstrate that I-QAOA achieves higher spectral efficiency compared to traditional channel-based and random pairing methods, even at low transmission power, and maintains superior performance at high power levels, offering a promising approach for optimizing smart grid communication networks.

Starvation is the most common physiological challenge for marine bivalves, especially in oligotrophic seas like the South China Sea. Its dominant oyster species, Saccostrea mordax, remains unclear regarding starvation tolerance and underlying mechanisms. Here, we treated S. mordax with starvation for four months. Compared with the control group, no significant decrease in survival was observed in starvation-treated S. mordax. A significant decrease in oyster respiration rate was observed from the 14th day, and on the 90th day, the activities of Superoxide dismutase (SOD) and Catalase (CAT), and the content of Malondialdehyde (MDA) in the gill of oyster decreased. Besides, we found that oysters had a slow process of morphological deterioration in different tissues (gill, digestive gland, muscle) under starvation treatment. The integration of transcriptomics analyses indicates that the metabolic capacity of oysters decreases under prolonged starvation, and the expression of genes in the solute carrier family that contribute to nutrient transport was down-regulated. Moreover, the down-regulation of the expression of immune-related genes, such as the IAP gene family, MYD88, and CDC42, indicated that the homeostasis between apoptosis and cell proliferation was disrupted, which was in agreement with the results of sections with the enlarged lumen of the digestive gland and apoptosis of digestive cells. On the contrary, most genes in the collagen family, that contribute to the stability of tissue structure, were up-regulated, which may contribute to long-term tolerance to starvation. Overall, our study reveals the physiological responses and mechanisms of starvation tolerance in S. mordax, which has the potential to contribute to the restoration of island reefs in oligotrophic seas.

The world’s energy is accelerating towards the stage of deep game, and there is a trend of adjustment, differentiation and reconstruction. Oil and gas resources are still the main source of core competitiveness in the oil and gas industry. As the same time, the production of carbonate reservoirs is limited by geological and engineering conditions, especially in areas with developed fractures. A thermal-hydraulic-mechanical-chemicalmulti-field coupled model is developed in this paper for exploring the acid fracturing productivity of fractured carbonate reservoir. The rock deformation, two-phase flow, heat transport and acid flow reaction are all solved in extended finite element method. The main controlling factors of productivity are studied by combining engineering geological condition in the fractured reservoir, and it involves fracture density, stress sensitivity, fracture morphology and acid fracturing mode. The mechanism of the main controlling factors of productivity are clarified in fractured reservoir. The acidizing modes under different fracture densities are demonstrated and the fracture strategy is determined.

Natural pH variability in coastal-estuarine systems exacerbates OAX events through frequent pCO2 spikes, posing severe threats to bivalves and ecosystems they support. While selective breeding has improved growth performance in oysters, its capacity to enhance tolerance to acidic stress remains poorly understood. Here, we evaluated the physiological performance of wild and recently selectively bred oyster variety (Guihao No. 1) under the simulation of recurrent OAX scenarios. In comparison to wild oysters, selectively bred oysters exhibited significantly higher survival rates, fast shell growth, and improved condition index. Energy metabolism suggests that selective breeding confers enhanced stress resilience in oysters by optimizing feeding capacity, increasing oxygen uptake, and reducing ammonia excretion rates. This metabolic efficiency supports more effective protein and glycogen turnover, as evidenced by elevated O:N ratios, and ultimately results in higher SFG. PCA analysis demonstrated that enhanced energy metabolism (CMA, NKA), antioxidant capacity (low MDA), and immune activity (high ACP, AKP) contributed to improved growth and resilience of selectively bred oysters when exposed to OAX, whereas wild oysters showed metabolic suppression and oxidative damage. These results highlight the role of selective breeding in promoting stress tolerance through optimized energy allocation and defense mechanisms, offering valuable guidance for climate-resilient oyster aquaculture in acidifying oceans.

In oil and gas production, reservoir heterogeneity causes plugging removal fluids to preferentially enter high-permeability zones, hindering effective production enhancement in low-permeability reservoirs. Traditional chemical diverting agents exhibit insufficient stability in high-temperature, high-salinity environments, risking secondary damage. To address these challenges, this study developed a water-sensitive self-thickening emulsion, targeting improved high-temperature stability, selective plugging, and easy flowback performance. Formulation optimization was achieved via orthogonal experiments and oil-water ratio adjustment, combined with particle size regulation and viscosity characterization. Core plugging experiments demonstrated the new emulsion system's applicability and diverting effects. Results showed that under 150 °C and 15 × 104 mg/L NaCl, the emulsion maintained a stable viscosity of above 302.7 mPa·s, with particle size D50 increasing from 31.1 μm to 71.2 μm, exceeding API RP 13A's 100 mPa·s threshold for acidizing diverters, providing an efficient plugging solution for high-temperature, high-salinity reservoirs. The injection pressure difference in high-permeability cores stabilized at 2.1 MPa, significantly enhancing waterflood sweep efficiency. The self-thickening mechanism, driven by salt-induced droplet coalescence, enables selective plugging in heterogeneous formations, as validated by core flooding tests showing a 40% higher pressure differential in high-permeability zones compared to conventional systems.

The development of 6G networks is driving the need for innovative resource orchestration solutions to meet the diverse and dynamic demands of the next-generation applications (APPs). As the number of connected devices increases, traditional network management approaches are insufficient to handle the complex and multiresource requirements of 6G, which include communication, computation, and storage capabilities across multidomain environments. To address these challenges, we introduce HyOrch, a novel approach for 6G-driven resource orchestration that employs hypergraph theory to model the intricate interactions and dependencies among network elements across multiple domains. HyOrch enables a comprehensive representation of these complexities, facilitating efficient resource orchestration across end devices (EDs), edge servers (ESs), and cloud servers (CSs). It employs a hierarchical distributed resource allocation mechanism that dynamically allocates resources based on real-time availability and application-specific requirements, ensuring optimal performance across the entire network. To validate the effectiveness of HyOrch, we conducted evaluations on a real-world testbed with both virtual and physical devices. The results show that HyOrch significantly outperforms existing approaches, improving resource efficiency by 31.54%-42.13% and reducing delay by 9.77%-39.12%, demonstrating its capability to address the evolving challenges of 6G network orchestration.

Global warming has significantly increased the frequency of marine heatwaves, causing severe and long-term impacts on tropical marine ecosystems and coral reef systems. As a key component of coral reefs, giant clams (Tridacna spp.) and their symbiotic algae are critical to ecosystem health, yet their complex regulatory mechanisms under environmental stress remain poorly understood. Our study revealed that while temperature fluctuations minimally affected the growth rates of giant clams, they significantly reduced survival probabilities. Physiological indicators demonstrated that giant clams mitigate oxidative stress caused by heatwaves through the regulation of antioxidant enzymes, such as catalase (CAT), glutathione reductase (GSH), and acid phosphatase (ACP), which play a key role in immune regulation and stress mitigation. For symbionts, with the elastic change of biomass, photosynthetic efficiency and chlorophyll content show dynamic changes, which is consistent with the pattern of symbionts' loss or death. Transcriptomic analysis further revealed distinct regulatory pathways for both giant clams and zooxanthellae under heatwave conditions, including nutrient metabolism, transport, energy distribution, and immunomodulatory responses. Notably, lipid metabolism genes were down regulated in host cells, suggesting reduced metabolic activity, while the MAPK (Mitogen-Activated Protein Kinase pathway) pathway was highly activated in clams, indicating enhanced immune responses to temperature stress. Conversely, the ABC (ATP-binding cassette transporter) transporters were significantly upregulated zooxanthellae, likely facilitating stress-related material transport. These findings highlight the need for integrated strategies targeting transcriptomic adjustments across symbiotic partners under stress-specific conditions, offering valuable insights into enhancing Tridacna resilience and addressing the broader impacts of climate change on marine ecosystems.

To support the explosive growth of wireless traffic, the emerging control- and user-plane separation (CUPS) paradigm allows control base stations (CBSs) to provide control plane (CP) coverage, while traffic base stations (TBSs) catering to varying mobile traffic and diverse Quality-of-Service (QoS) requirements. However, learning-based traffic prediction has yet to utilize the potential advantages provided by CUPS to facilitate energy saving. In this article, we introduce a novel base station (BS) sleep scheme for CUPS that employs a learning-based approach to determine the TBS's active/deactive states. First, traffic demands of TBSs are predicted by a novel data-driven learning approach. This approach leverages the receptive fields to explore the spatial characteristics of mobile traffic, and then employs bidirectional long short-term memory (Bi-LSTM) networks to extract context feature of mobile traffic. Based on the traffic forecasting, we formulate a long-term network energy efficiency maximization problem that optimizes the active/deactive states of TBSs. Moreover, we introduce a service penalty term into CP to mitigate potential network coverage vulnerabilities of TBS. Then, an improved twin delayed deep deterministic policy gradient (TD3)-based algorithm is employed to solve the above nondeterministic polynomial-time hard (NP-hard) problem. Extensive experiments using a real-world dataset demonstrate that the proposed scheme significantly outperforms the existing methods in terms of traffic forecasting accuracy. Additionally, the proposed energy efficiency maximization algorithm achieves superior performance than other benchmark schemes.