Hong Kong University of Science and Technology
Hong Kong University of Science and Technology Institutional RepositoryNot a member yet
162821 research outputs found
Sort by
Machine Learning under Scarcity: Addressing Data Scarcity in Tabular Models and Computational Scarcity in Statistical Estimation
No full textEnhancing Quantitative Precipitation in Mesoscale NWP Models: A Multifaceted Approach Using Data Assimilation, Deep Learning, and Practical Applications
No full text
Temperature-dependent soil water retention and hydraulic conductivity model for biopolymer-treated soil
No full textWith the growing emphasis on sustainable development in geotechnical engineering, biopolymers are emerging as eco-friendly soil improvement agents. However, the influence of temperature on the hydraulic performance of biopolymer-treated soils remains poorly understood. This study develops a novel model to describe the water retention and saturated hydraulic conductivity (ks ) of biopolymer-treated soils under varying temperature conditions. The model integrates the temperature-dependent effects on biopolymer water absorption, swelling behavior, and hydrogel viscosity. Laboratory experiments were conducted to validate the model. As temperature increased from 20 to 60◦C, gravimetric water content of biopolymer decreased by 49%, leading to a reduced swelling effect. In parallel, biopolymer viscosity declined by 34%–50% between 5 and 65◦C. Elevated temperature diminished soil hydraulic performance. Saturated water content decreased by 4.9% at 1% biopolymer dosage, as temperature rose from 20 to 60◦C. Both air-entry value and residual water content also declined. Soil ks decreased by nearly four orders of magnitude with a 1% biopolymer dosage at 25◦C. However, this effect weakened at elevated temperatures, as ks increased by 4.3-fold when temperature rose from 5 to 65◦C. The proposed model accurately reproduced these results, offering a reliable tool for designing sustainable geotechnical systems under varying climatic conditions.</p
Tracing paths, pruning noise: Toward robust IP geolocation via topology-guided shaping and refinement
No full textAccurate IP geolocation is essential for a wide range of applications, yet real-world networks pose significant challenges due to noisy measurements, dynamic topologies, and incomplete information. This study proposes TRFGeo (Topology-guided with Reliable Filtering Geolocation), a robust geolocation framework that enhances structural reliability and predictive accuracy through three synergistic modules: network-aware graph construction, topology-guided link refinement, and adaptive message filtering. TRFGeo jointly optimizes graph structure and representation learning via self-supervised edge modeling and role-aware message propagation. We evaluate TRFGeo on three large-scale datasets from New York, Los Angeles, and Shanghai, encompassing over 310,000 IP addresses with 24-dimensional network measurements and attribute features ranging from 6 to 27 dimensions across datasets. Experimental results demonstrate that TRFGeo consistently achieves the lowest geolocation errors across all three datasets, with median errors reduced to 0.868 km in New York, 1.626 km in Los Angeles, and 3.577 km in Shanghai. The framework also maintains superior performance under structural noise and missing-link perturbations, consistently outperforming five state-of-the-art baselines. These results validate the effectiveness of topology-guided graph reconstruction and adaptive message filtering in enhancing the robustness of IP geolocation systems.</p
Adaptive Optimization of Active RIS-assisted ISCPT Network: A Hybrid MoE Scheme
No full textThis paper investigates the active reconfigurable intelligent surface (RIS)-assisted integrated sensing, communication, and power transfer (ISCPT) networks, where rate-splitting multiple access (RSMA) scheme is employed to serve multiple downlink communication users. To promote the energy efficiency (EE) of such a system, we formulate an EE maximization problem by jointly optimizing the beamforming matrix, the sensing matrix, the active RIS matrix, the power splitting (PS) ratio vector, and the common rate allocation vector. Due to the non-convexity of the problem, we first design a successive convex approximation scheme with alternating optimization method (named SCA-AO) to solve it. As SCA-AO operates in an iterative manner, which is with relatively high computational complexity, we then design a mixture of experts (MoE)-based deep reinforcement learning (DRL) scheme with smooth clipping function (named MoE-SCF). In comparison, SCA-AO is able to achieve higher solution accuracy, while MOE-SCF has a shorter online execution response time. In order to integrate the advantages of both presented SCA-AO and MoE-SCF simultaneously, we further propose a hybrid MoE (H-MoE) scheme, where both the SCA-AO and the MoE-SCF are employed as expert strategies, and an opportunistic activator (OPA) is designed to dynamically select the best strategy generated by all expert combinations according to the performance evaluation function. Simulation results demonstrate that the proposed H-MoE promotes the system's EE by about 18.14% compared to traditional MoE, with similar response time. Additionally, compared to the SCA-AO, H-MoE significantly decreases the response time by approximately 56.17%, while only marginally compromising the EE performance by less than 3.1%.</p
A Flipped-Voltage-Follower-Based LDO With Scaled-Down Current Buffer Compensation for Wide Load Current Range
No full textThe flipped voltage follower (FVF) based low dropout regulator (LDO) has been developed rapidly due to its fast transient response, high power supply rejection (PSR) and reduced complexity. The conventional FVF loop encounters stability challenges under heavy load conditions when employing dominant pole compensation with a large output capacitance. This paper proposes a solution to extend the load current range by reducing the size of the current buffer and adding a small feed-forward capacitor (Cc). In addition, a transient current enhanced buffer is used to reduce output voltage spikes. The proposed LDO has been verified in a standard 0.18-μm CMOS process and occupies an active area of 0.047 mm2. Experimental results show that this LDO can deliver 300 mA load current at 200 mV dropout voltage, with a current efficiency of 99.99%. The undershoot voltage under maximum load current change in 100ns edge time is 35.8 mV with a 0.95 μs recovery time.</p
Charged adsorbents for iodine capture
No full textRadioactive iodine isotopes, such as 129I and131I, present in nuclear waste are not only highly volatile but also tend to bioaccumulate in marine organisms, ultimately posing severe health risks to humans through the food chain. This critical challenge has spurred the development of advanced materials for the effective capture and safe storage of radioactive iodine. Charged adsorbents show great promise for iodine uptake through a range of noncovalent interactions including electrostatic forces, hydrogen bonding, anion-π interactions, and halogen bonding with iodine species, thereby improving adsorption performance. These charged materials demonstrate strong affinities for iodine species. This review summarizes the recent progress of various charged adsorbents, including metal-organic frameworks (MOFs), ionic liquids (ILs), porous aromatic frameworks (PAFs), porous organic polymers (POPs), covalent organic frameworks (COFs), macrocycles and molecular cages. Finally, the article discusses emerging trends and future prospects for charged adsorbents aimed at capturing radioactive iodine.</p
Mechanically assisted Li<sup>+</sup>-conduction in crown ether-covalent organic frameworks for lithium metal batteries
No full textMechanically interlocked molecules (MIMs) enable controlled motions like rotation and shuttling, ideal for molecular machines. Heteroatom-containing MIMs, such as crown ethers, exhibit host–guest interactions, coordinating Li+ for transport. Crown ethers are integrated into nitrogen-rich 2D covalent organic frameworks (COFs) to create a high-performance quasi-solid-state electrolyte (Li+@Crown-COF) for lithium metal batteries. This electrolyte achieves exceptional ionic conductivity (3.2 × 10−3 S cm−1) and a Li+ transference number (0.60) at room temperature (r.t.). The mechanically assisted Li⁺ conduction, driven by crown ether motion within the COF's porous framework, enhances ion transport and stabilizes the lithium anode, suppressing dendrite growth. Electrochemical tests show excellent cycling stability, with full cells using an LiFePO4 cathode retaining 95% capacity after 600 cycles at 0.5C and r.t. At 60 °C and 2C, the cell maintained 85% of its initial capacity after 300 cycles, with 99.99% Coulombic efficiency. Solid-state nuclear magnetic resonance and computational studies confirm mechanical motions and strong Li⁺ binding to COF's nitrogen and oxygen sites. This MIM-COF design, leveraging the chemical novelty of mechanically interlocked systems, paves the way for safe, stable, and high-energy-density LMBs.</p