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
Site-specific projection of rainfall patterns under climate change by joint sparse representation
No full textAs climate change alters global rainfall patterns, many regions are facing increased intensity and frequency of rainfall events. These changes pose significant risks to civil infrastructure, which was often designed based on historical data and may no longer be resilient. Rainfall-induced failures can lead to severe, life-threatening consequences. Local factors, such as topography and elevation, greatly influence rainfall variability, making site-specific projections essential for effective risk assessment of infrastructure. However, current rainfall projections from General Circulation Models (GCMs) have coarse spatial resolutions (e.g. 100 km), which are inadequate for assessing risks at specific sites, such as slopes near railways, where the relevant scale is often tens to hundreds of metres. This study proposes an innovative method that integrates historical rainfall records with GCM projections using a joint sparse representation (JSR) framework to project future rainfall patterns at specific sites. This approach combines regional trends from GCMs with local data to maintain regional consistency while accurately reflecting local characteristics. A temporal downscaling step further enhances the resolution for engineering applications. The method is demonstrated using real rain gauge data from Hong Kong.</p
Signaling for a Fluid Antenna System With Uniform Correlation
No full textFor a single-antenna fluid antenna system (FAS) with N fixed antenna locations or ports, analytical expressions for the error performance of four types of signaling schemes for digitally modulated data transmission and reception are found. The signaling schemes are: (1) M-ary phase-shift keying with coherent reception, (2) two-sided M-ary amplitude-shift keying (M-ASK) with coherent reception, (3) one-sided M-ASK with noncoherent reception, and (4) one-sided M-ASK with partially noncoherent reception. The N ports are evenly distributed on a line. The FAS error performance is analyzed when the ports are subject to identically distributed Rayleigh fading with uniform correlation between the channels at the ports. A series expression for the characteristic function (c.f.) of the random complex channel coefficient magnitude squared of the chosen FAS port, in terms of a series of elementary c.f.s, where each elementary c.f. is the c.f. of a sum of N independent gamma random variables, is obtained. Series expressions for the symbol error probabilities as weighted sums of functions in closed form are presented. Closed form asymptotic high signal-to-noise ratio results for error performance are also derived. Numerical are presented to show the variation of the error performance with respect to the spacing between the first and Nth ports and N.</p
Causal Inference Via Style Bias Deconfounding for Domain Generalization
No full textDeep neural networks (DNNs) often struggle with out-of-distribution data, limiting their reliability in real-world visual applications. To address this issue, domain generalization methods have been developed to learn domain-invariant features from single or multiple training domains, enabling generalization to unseen testing domains. However, existing approaches usually overlook the impact of style frequency within the training set. This oversight predisposes models to capture spurious visual correlations caused by style confounding factors, rather than learning truly causal representations, thereby undermining inference reliability. In this work, we introduce Style Deconfounding Causal Learning (SDCL), a novel causal inference-based framework that explicitly addresses style as a confounding factor to enhance domain generalization in image modalities. Our approaches begins with constructing a structural causal model (SCM) tailored to the domain generalization problem and applies a backdoor adjustment strategy to account for style influence. Building on this foundation, we design a style-guided expert module (SGEM) to adaptively clusters style distributions during training, capturing the global confounding style. Additionally, a backdoor causal learning module (BDCL) performs causal interventions during feature extraction, ensuring fair integration of global confounding styles into sample predictions, effectively reducing style bias. The SDCL framework is highly versatile and can be seamlessly integrated with state-of-the-art data augmentation techniques. Extensive experiments across diverse natural and medical image recognition tasks validate its efficacy, demonstrating superior performance in both multi-domain and the more challenging single-domain generalization scenarios.</p
Multi-objective optimization for sustainable dimethyl oxalate synthesis: A plant-wide framework balancing economic benefits and carbon emissions
No full textEthylene glycol (EG) serves as a primary raw material in the polyester industry, with syngas-to-dimethyl oxalate (DMO) conversion representing an advanced EG production method. However, this process encounters conflicting objectives between maximization of economic benefits and minimization of carbon emissions, particularly exacerbated by constraints and market prices. To address this challenge, we developed a multi-objective optimization framework for various working conditions: First, we establish a steady-state simulation system incorporating reaction kinetics and mechanisms to model the DMO synthesis process. Then, an innovative economy-carbon emission multi-objective optimization problem is formulated, where the ranges of pivotal operating parameters are determined by sensitivity analysis, and the response surface method is used to obtain the reference points under different conditions. Finally, the optimization problem is solved by the Pareto frontier (PF) estimation algorithm to solve the irregular PF problem, which arises from the complex nonlinear interactions between process variables under various working and price conditions. Under regular working conditions, we compare the knee point among the obtained Pareto solution set with the reference point, and the framework reduces carbon emissions by 19.63% (129.5 kmol/h) while increasing economic benefits by 1.38% (1253.1 yuan/h). Considering three typical conditions of sharp increase of DMC prices, limited production capacity and short-term negative profits, our framework identifies solutions that dominate the reference points and the original turning points in the obtained PF. The results have verified that this study is able to support the decision-making in providing solutions with a good balance between economy and carbon emissions under various working and price conditions.</p
Day-ahead power forecasting of self-cleaning nanocoated and conventional rooftop PV systems using SHAP-RFE-MCCV feature selection and deep learning
No full textAccurate day-ahead power forecasting of rooftop photovoltaic (PV) systems is critical for grid operation, energy trading, and smart building management. While self-cleaning nanocoatings can enhance PV energy yield by mitigating dust deposition and maintaining optical transmittance, their impact on forecasting performance remains largely unexplored. This study investigates the day-ahead forecasting behavior of nanocoated and conventional rooftop PV systems using five deep learning architectures: DNN, LSTM, 1D CNN, CNN-BiLSTM, and CNN-BiGRU. A SHAP-driven Recursive Feature Elimination with Monte Carlo Cross-Validation (SHAP-RFE-MCCV) framework was developed to identify the most relevant features from hundreds of lagged meteorological and power variables. Results indicate that the nanocoated PV system achieves a net cumulative power gain of 4.65% over 51 days relative to the uncoated system, corresponding to an average daily increase of 4.71%. This period covers the entire dataset used for forecasting, providing a representative assessment of coating benefits under varied irradiance conditions. While the coating enhances energy yield, sharper power variations lead to marginally higher prediction errors, reflecting the slightly increased forecasting difficulty. Among the models, DNN consistently attains the highest accuracy (R2: 0.9289–0.9496; MAE: 0.7051–0.8148), with LSTM also showing competitive predictive capability. The SHAP-RFE-MCCV framework effectively reduces input dimensionality by over 90% while preserving strong predictive accuracy across models (R2 > 0.92). The study demonstrates that nanocoating not only improves energy generation but also alters temporal power patterns and forecastability. The proposed feature selection method offers an efficient, interpretable solution for high-dimensional PV forecasting and insights for integrating rooftop PV systems into smart grid applications
Synergistic vacancy and fluorine doping in metallic 1T-CoTe<sub>2</sub> for high-performance sodium-tellurium battery cathodes
No full textSodium-tellurium (Na-Te) batteries have emerged as promising candidates for next-generation high-energy–density storage systems, owing to their high theoretical capacity and the natural abundance of sodium. However, their practical application is severely hindered by the shuttle effects of soluble sodium polytelluride (Na-pTe) intermediates, which leads to rapid capacity decay. In this study, we propose a synergistic defect-engineering strategy—incorporating Te vacancies and F doping—to transform metallic 1T-CoTe2 into an advanced cathode host for Na-Te batteries, as investigated through systematic first-principles calculations. The introduction of Te vacancies and F atoms (VTe-CoTe2-xF) significantly enhanced chemical adsorption of Na-pTe intermediates, effectively suppressing their dissolution and shuttle. Besides, VTe-CoTe2-xF markedly reduced energy barriers for the conversion reaction kinetics of Na-pTe intermediates, enabling high-rate capability. Moreover, it promoted uniform dispersion of Na atoms, increasing active sites and enhancing pseudocapacitive contribution. This work elucidates the atomic-scale mechanisms of the synergistic vacancy-doping effect provides a fundamental design principle for developing high-performance host materials based on two-dimensional metallic tellurides for advanced metal–chalcogen batteries.</p
SFVnet: Finite-volume informed U-net for compressible flow prediction with sparse data under ill-conditions
No full textPhysics-informed neural network (PINN) is a promising methodology in scientific computing. However, predicting compressible flows poses a challenge for PINNs, since they struggle to accurately capture discontinuities arising in flow evolutions. In this work, a novel physics-informed deep learning framework, called sparse finite-volume informed U-net (SFVnet), is developed to predict compressible flow fields with sparse data under ill-conditions. The major contributions are as follows: (1) a new physical loss function is designed by incorporating finite volume discretized residuals and fusing predictions from multiple points within each cell, effectively improving the discontinuity-capturing ability compared to original PINN; (2) the model leverages interior sparse samples to reconstruct the full flow field without the input of initial/boundary conditions, which is particularly challenging for traditional FVM; (3) the trained model can extrapolate basic flow patterns beyond the training time window, which original PINNs fail to achieve. Furthermore, the proposed framework is distinguished by reconstructing region-of-interest flow fields by sampling data only within this region. A series of one-dimensional (1D) and two-dimensional (2D) benchmark cases, including the 1D Sod’s tube, 1D Lax’s tube, 2D Riemann problems, and double Mach reflection problem, demonstrate the prediction accuracy and robustness of the framework. Notably, this is the first physics-informed deep learning framework successfully applied to the double Mach reflection simulation with Mach number of 10. These results also indicate the potential of present framework for flow field reconstruction, data compression, and restoration.</p
Burning windjammer: Multi-rotor engineered photothermal agent with 87% photothermal conversion for antimicrobial treatment
No full textThe evolution of drug-resistant pathogens has significantly diminished the efficacy of antibiotics. Photothermal therapy has appeared as a promising approach for the curing of infectious diseases on account of its high therapeutic efficiency, high controllability, and broad-spectrum bactericidal effect. In this work, by subtly engineering the molecular rotors, a high-performance photothermal agent, 4TPE-TBZ, with a photothermal conversion efficiency up to 87 %, is elaborately designed and fabricated. Mechanism studies disclose that the strategic incorporation of multiple tetraphenylethylene units not only facilitates the molecular rotation but also expands intramolecular motional freedom, thereby effectively enhancing heat generation. In vitro studies demonstrate that 4TPE-TBZ nanoparticles can effectively eradicate the biofilms of both S. aureus and E. coli through the photothermal effect. Subsequent in vivo evaluations conducted on S. aureus-infected abscess and chronic joint inflammation models consistently reveal potent photothermal antibacterial activity alongside marked attenuation of inflammatory responses. This research provides valuable insights for the invention of high-performance photothermal agents that can be applied in effective antibacterial treatments.</p
Synthetic instance segmentation from semantic image segmentation masks
No full textIn recent years, instance segmentation has garnered significant attention across various applications. However, training a fully-supervised instance segmentation model requires both instance-level and pixel-level annotations, which are costly to obtain. In contrast, weakly-supervised instance segmentation methods, such as those using image-level class labels or point labels, often struggle to satisfy the accuracy and recall requirements of practical scenarios. In this paper, we propose a novel paradigm called Synthetic Instance Segmentation (SISeg). SISeg achieves instance segmentation results by leveraging image masks generated by existing semantic segmentation models. It is highly efficient as does not require additional training for semantic segmentation or the use of instance-level image annotations. In other words, the proposed model eliminates the need for extra manual effort or higher computational expenses. Specifically, we first obtain a semantic segmentation mask of the input image via an existing semantic segmentation model. Then, we calculate a displacement field vector for each pixel based on the segmentation mask, which can distinguish representations belonging to the same class but different instances, i.e., extracting the instance-level object information. Finally, the instance segmentation results are refined by a learnable category-agnostic object boundary branch. Extensive experimental results on two challenging datasets highlight the effectiveness of SISeg in achieving competitive performance compared to state-of-the-art methods, especially fully-supervised methods.</p
Reconfigurable Pixel Antennas Meet Fluid Antenna Systems: A Paradigm Shift to Electromagnetic Signal and Information Processing
No full textTraditionally, antennas and wireless communication technologies have been developed separately. While antennas focus on maximizing the received or transmitted signal strength, communication technologies optimize performance through coding, signal processing, and resource allocation. However, strong antenna signals do not guarantee high-quality communication due to factors such as channel fading and interference. Recently, the fluid antenna system (FAS) has emerged as a paradigm that treats the radiating aperture as a flexible, reconfigurable physical-layer resource and integrates it into the physical-layer design, broadening the scope of system and network optimization and inspiring next-generation reconfigurable antennas. This paradigm naturally couples electromagnetic signal and information processing (ESIP). A key enabler is the reconfigurable pixel antenna (RPA), which offers high degrees of reconfigurability via pixel-level switching. This article explores the integration of RPA into the FAS concept and highlights the unique ESIP opportunities and associated challenges. Experimental results are presented to demonstrate the significant potential of RPA-enabled FAS.</p