Objective: As urbanization accelerates and climate change intensifies, the risks of various urban disasters have increased considerably, hindering the transition toward urban resilience. As the core carrier of urban operation, urban infrastructure plays a pivotal role in enhancing urban resilience. Analyzing the interactions and correlation characteristics within its flood control system is of great importance for risk identification and resilience building in urban systems. Previous research on urban resilience has primarily focused on statistical methods, with limited research analyzing the characteristics of urban resilience from a probabilistic perspective. The Copula joint probability method offers distinct advantages in exploring multidimensional nonlinear relationships in complex systems and conducting comprehensive probability assessments of interfactor interactions. Methods: This study developed an evaluation index system for urban infrastructure flood resilience (UIFR) based on the resistance R–adaptability A–recovery C framework. The study used survey data from 2000 to 2022 and employed a Copula-theory-based joint probability estimation approach to characterize the correlation patterns among UIFR dimensions and the comprehensive probability of multifactor joint effects. In addition, the Copula function model was applied to investigate the joint distribution probability of UIFR in the Beijing–Tianjin–Hebei region of China, which is frequently affected by floods. The eXtreme gradient boosting-SHapley additive explanations (XGBoost–SHAP) model was integrated to explore the driving effects on UIFR. Results: The results showed that: 1) from 2000 to 2022, the UIFR in the Beijing–Tianjin–Hebei region exhibited fluctuating changes with an overall upward trend. Beijing recorded the highest UIFR, followed by Tianjin, while Hebei Province ranked the lowest. 2) For the two-dimensional joint distribution probability, the probability of achieving high resilience for the A–C pair in Beijing was 0.761 9, whereas the highest probabilities of high resilience for the R–A pair in Tianjin and Hebei were 0.116 9 and 0.149 7, respectively; the lowest probability of high resilience for the A–C pair in Hebei was 0.063 4. 3) For the three-dimensional joint distribution probability, the probabilities of achieving high resilience for the R–A–C combination in Beijing, Tianjin, and Hebei were 0.717 2, 0, and 0.066 2, respectively. The unbalanced development among UIFR dimensions impaired the overall level of the UIFR. 4) The harmless domestic waste treatment rate C5, per capita urban electricity consumption R4, and the share of the labor force A4 had remarkable impacts on the UIFR in the Beijing–Tianjin–Hebei region, with strong interaction between C5 and other factors. The complex interactions among factors jointly influenced the UIFR in the Beijing–Tianjin–Hebei region. Conclusions: The UIFR joint probability method based on Copula functions demonstrates strong linear correlations between theoretical and empirical probability values of each Copula function, enabling effective two-dimensional and three-dimensional joint probability assessments of urban resilience. This method offers a new perspective on the relationships between internal connectivity and different aspects of resilience in UIFR. In addition, this study identifies key risk factors affecting the UIFR improvement in the Beijing–Tianjin–Hebei region, proposing additional targeted policy recommendations for future urban resilience initiatives. This study provides data support and theoretical guidance for joint risk assessment within infrastructure flood resilience systems and the coordinated development of urban resilience.
Objective: Global climate change and rapid urbanization have intensified compound flood risks in coastal cities. A critical research gap persists regarding cascading failures triggered by instability in high-standard flood control infrastructure. Prevailing assumptions often underestimate the failure probability of robust defense systems; however, historical events, such as the 2015 dike overflow along the Jiangnan Canal, demonstrate that extreme events can overwhelm even 50–200-year flood defenses. This study examines the Wuxi Yundong Enclosure, a representative high-standard flood control zone in the Yangtze River Delta. Focusing on the "rainstorm–dam Breach–urban flooding" disaster chain, this study aims to quantitatively reveal the superposed magnification mechanism of economic losses under beyond-standard rainfall, providing a scientific basis for resilient urban flood management. Methods: An integrated "hydrodynamic simulation–loss quantification–attribution analysis" framework was developed. A multiscale coupled hydrodynamic model was established to link a Taihu Basin-scale model (boundary conditions), a 1D river network model (canal dynamics), and a 2D inundation model (surface flooding). Validation against the June–July 2016 flood event yielded a Nash–Sutcliffe efficiency of 0.85 and R2 of 0.99. A hydrographically triggered breach mechanism was simulated: failure occurred when water levels exceeded the dike crest by 10 cm for 10 min, forming a trapezoidal breach (top width 12 m, bottom width 8 m, height 3 m). Twelve scenarios were designed across four return periods (50-, 100-, 150-, and 200-year) covering rainstorm-only, breach-only, and compound conditions. Economic losses for seven asset types (agriculture, buildings, equipment, and others) were quantified by spatializing high-resolution land-use data and applying depth–damage curves. Novel indices—the composite increment coefficient and amplification factor (AF)—were proposed to isolate the marginal impact of dike breaching. Results: The simulations revealed significant nonlinear amplification in compound scenarios. (1) Hydrological Effect: At the 200-year return period, the maximum inundation depth reached 2.44 m, a 134.7% increase over the linear sum of single-disaster scenarios. (2) Economic Magnitude: Total direct losses in the compound scenario surged by 221.7% relative to linearly superimposed single disasters. High-loss zones (> 30, 000 CNY per 900 m² pixel) expanded by 179.1%. Critically, over 50% of total losses occurred within a narrow 25 h window coinciding with peak rainfall and dike breach. (3) Asset Heterogeneity: Amplification varied significantly by asset type. Due to "plateauing" depth–damage curves, AF values for plantations, fisheries, and equipment remained below 100%. Conversely, buildings exhibited the highest sensitivity, with AF increasing by ~6.9% per decade of return period. (4) Attribution: The net superposition increment for total assets turned positive near the 85-year return period. Buildings accounted for the vast majority of incremental loss, contributing 94.2% of the total superposition at the 200-year return period. The total asset loss growth rate was approximately 3613.8 million CNY per 10-year increase in return period. Conclusions: This study demonstrates that high-standard flood control zones face catastrophic risks during compound disasters with structural failure. The economic impact of the "rainstorm–dam Breach–urban flooding" chain vastly exceeds the sum of individual hazards. The identification of buildings as the primary driver of loss magnification underscores the necessity for targeted structural reinforcements and risk-transfer strategies. Despite focusing on direct losses and excluding indirect disruptions, the framework provides a robust methodology for extreme flood risk analysis. These findings offer critical insights for optimizing flood storage strategies, refining emergency protocols, and updating engineering standards to incorporate compound disaster chains, thereby enhancing urban flood security in deltaic megacities globally.
Objective: Extreme rainstorms are increasingly threatening urban flood safety under climate change and rapid urbanization. Because rainfall is a direct driver of urban flooding, existing studies usually describe flood-producing rainstorms using meteorological indicators such as rainfall amount, duration, intensity, and coverage. However, in complex urban catchments, rainfall spatiotemporal heterogeneity interacts with highly variable underlying surface and drainage conditions, making meteorological indicators alone insufficient to explain flood responses. Current studies still insufficiently account for the hydrological properties of rainfall areas and the effects of observation spatial resolution on rainfall characterization. Therefore, we develop a hydrometeorological-integrated method for quantifying the spatiotemporal characteristics of rainfall and combine rainfall upscaling with urban flood simulation to examine how rainfall characteristics and underlying surface conditions jointly affect flood processes while identifying key features of urban flood-producing rainstorms. Methods: The Qing River basin in Beijing, a typical highly urbanized catchment of approximately 200 km2 characterized by fragmented impervious surfaces and dense drainage networks, was selected as the study area. Rainfall fields were obtained from two X-band dual-polarization weather radars with a temporal resolution of 3 min and an original spatial resolution of 75 m. Radar quantitative precipitation estimation was conducted based on specific differential phase information and evaluated using observations from meteorological stations. The rainfall product was converted into a 100 m regular grid and spatially upscaled to 300 m, 500 m, 1 km, 3 km, and 10 km resolutions. Seven representative rainfall events with different rainfall amounts, intensities, movement paths, and spatial patterns were selected to construct rainfall scenarios. A hydrometeorological-integrated method incorporating eight temporal indicators related to rainfall duration, amount, intensity, and variability, along with eight spatial indicators related to rainfall spatial variability, impervious-area rainfall, rainstorm coverage, and rainfall-area distance to drainage elements, was established to quantify the spatiotemporal characteristics of rainfall. Urban flood processes were simulated under different rainfall scenarios using a grid-based urban hydrological model that represents runoff generation and routing over pervious areas, impervious areas, drainage pipes, and river channels. Peak flow bias and high-flow accuracy were used to evaluate flood impacts, whereas correlation and random forest importance analyses were used to identify the key rainfall characteristics controlling urban flood responses. Results: The results show that decreasing rainfall spatial resolution considerably changed the expression of rainfall spatiotemporal characteristics. When the resolution was decreased from 100 m to 10 km, rainfall amount, rainfall intensity, and spatiotemporal variability were generally underestimated, rainfall duration was prolonged, and rainfall coverage and rainfall-area distances generally increased. At the 10 km resolution, the median underestimation of peak rainfall intensity reached approximately 30%, and rainfall spatial variability decreased by approximately 40%. These changes further affected urban flood simulations and resulted in peak flow biases of up to −40%. Random forest analysis showed that temporal rainfall characteristics contributed approximately 64% and 70% to peak- and high-flow formations, respectively, indicating their dominant role in shaping urban flood responses. Among all indicators, the maximum 30-min rainfall intensity had the greatest influence on flood peaks, with an importance score close to 30%. Meanwhile, spatial indicators such as rainfall-area distance and rainfall over impervious areas had non-negligible effects, with importance scores of approximately 10%. Conclusions: In this study, we develop a hydrometeorological-integrated method for quantifying rainfall spatiotemporal characteristics and identifying urban flood-producing rainstorms. Results show that rainfall observation resolution directly affects rainfall intensity, variability, coverage, and location characteristics, thereby considerably affecting urban flood simulations. Rainfall intensity, particularly the maximum 30-min intensity, dominates flood peaks and high-flow processes, whereas rainfall-area hydrological properties and distances to drainage systems remain critical. The proposed framework extends rainfall characterization beyond conventional meteorological indicators and supports more accurate rainfall monitoring and urban flood prediction.
Objective: Building fire safety design commonly relies on computational fluid dynamics simulations to obtain fire temperature fields. However, simulating the complete fire process in complex scenarios is time-consuming and cannot readily meet rapid assessment requirements. Existing methods for predicting fire temperature fields with conditional generative adversarial networks mostly rely on static mappings from building conditions to single-frame temperature fields, without fully accounting for the temporal dependencies between temperature fields at adjacent time steps as a fire develops. To address this issue, this study develops a temporal conditional generative adversarial network (TC-GAN), incorporating historical-state constraints into the pix2pix model. This study further optimizes the model to mitigate error accumulation in autoregressive prediction, thereby enabling rapid dynamic prediction of temperature-field evolution in single-story building fires. Methods: A fire dataset was constructed using Fire Dynamics Simulator. It comprised 35 single-story building layouts, each configured with two fire-source locations, resulting in a total of 70 fire scenarios. Given the building layout, fire-source location, and prediction time, TC-GAN introduces the temperature field at the previous time step as a historical-state input and, during inference, continuously generates temperature-field sequences in an autoregressive manner. To mitigate error accumulation, three optimization strategies, namely, multi-step loss, scheduled sampling, and a video discriminator, were compared. Model performance was comprehensively evaluated using the normalized root mean square error (NRMSE), the structural similarity index measure (SSIM), temperature-evolution curves at key monitoring points, and the intersection over union (IoU) and area error of high-temperature regions above 70 ℃. Results: The results showed that TC-GAN with only historical-state constraints was affected by autoregressive error accumulation and exhibited lower predictive performance than pix2pix, whereas the model with scheduled sampling achieved the best performance. On the test set, the model with scheduled sampling reduced average NRMSE by 39.3% and increased average SSIM by 24.6% compared with pix2pix. Temperature-field visualizations showed that this variant captured the temperature diffusion process in complex building spaces more consistently, reducing the local high-temperature accumulation, spatial-pattern fragmentation, and abnormal diffusion observed in the pix2pix predictions. The predicted curves at key monitoring points likewise showed closer agreement with the reference temperature trends. For regions with temperatures above 70 ℃, the model with scheduled sampling increased the average IoU from 0.357 4 to 0.561 1 and decreased the average absolute area error from 0.276 5 to 0.212 2, indicating improved predictive capability for capturing the location and extent of high-temperature regions. Conclusions: Historical-state constraints supply temporal evolution information for building-fire temperature-field prediction, but effective error-suppression strategies are required to mitigate autoregressive error accumulation. Under the conditions investigated in this study, TC-GAN combining scheduled sampling with historical-state constraints achieved the best overall predictive performance, outperforming the conventional pix2pix model in temperature-field accuracy, spatial structural consistency, temporal evolution at key monitoring points, and prediction of high-temperature regions. This method enables rapid prediction of temperature-field evolution in single-story building fires, reduces the time and computational costs of conventional numerical simulation, and provides a potential technical pathway for building fire safety design and rapid fire-risk assessment.
Objective: Rigid polyurethane foam (RPUF) is widely used in building insulation and other applications, yet its inherent flammability and substantial release of toxic smoke during combustion severely limit its safe use across various scenarios. Conventional flame-retardant strategies often cause environmental pollution and impair mechanical properties, whereas bio-based flame retardants alone exhibit limited effectiveness. Herein, based on the intumescent flame-retardant theory, a synergistic flame-retardant system was constructed using melamine (MEL) as the gas source and silk fibroin (SF) as the charring agent. This study systematically examines the effects of different MEL/SF mass ratios on the pyrolysis, combustion, and smoke-toxicity characteristics, and elucidates the underlying synergistic mechanism to provide a design reference for developing high-efficiency, low-toxicity bio-based flame-retardant RPUF. Methods: MEL/SF/RPUF composites with MEL/SF mass ratios of 0, 5∶2, 5∶1, 15∶2, and 10∶1 were fabricated via a one-step foaming method at a total additive loading of 10 wt%. Thermogravimetric analysis was used to examine the thermal decomposition behavior. Apparent activation energies were calculated via the Friedman differential and Starink integral methods, and the Malek method was applied to identify the most probable reaction mechanisms. Cone calorimetry was conducted to record the heat release rate, smoke production, and CO yield, from which the fire growth index and toxicity potential index were derived to quantify fire hazards. The residual chars were characterized by scanning electron microscopy, energy-dispersive X-ray spectroscopy, and Raman spectroscopy to determine their micromorphology, elemental composition, and graphitization degree. Thermogravimetry–infrared spectrometry was additionally employed to identify the gaseous pyrolysis products. Results: The pyrolysis process of MEL/SF/RPUF composites can be divided into three stages. Increasing MEL content lowered the onset decomposition temperature, and the apparent activation energy followed a three-stage trend of "rising–fluctuating–high-level stabilization." The MEL/SF/RPUF-4 sample (MEL/SF = 15∶2) consistently showed the highest average activation energies across all stages, with final-stage values reaching 147.59 kJ·mol−1 (Friedman method) and 140.84 kJ·mol−1 (Starink method), indicating superior thermal stability. As MEL content increased, the reaction mechanism in the first half of pyrolysis gradually transitioned from the F2 model to the A2 model, whereas the second half deviated from all theoretical models, exhibiting complex mixed-control kinetics. At an MEL/SF ratio of 15∶2, the composite achieved the lowest mass loss rate and the highest residual mass (21.58%), which was 1.5 times that of the SF-only modified group. The peak heat release rate decreased by 23.36%, the peak smoke release rate dropped by 63.69%, and both the fire growth index and the toxicity potential index reached their minimum values. The char layer at this ratio was the most compact, smooth, and highly graphitized, with a carbon content of 75.9% and nitrogen retention of 11.3%. The early decomposition of MEL released NH3, which diluted the gas phase while promoting char formation in SF, demonstrating a pronounced gas–solid synergistic flame-retardant mechanism. Conclusions: The MEL/SF system exhibits a pronounced gas–solid synergistic flame-retardant mechanism, with the MEL/SF ratio of 15∶2 identified as the optimal formulation. The composite at this ratio achieves the best overall performance in thermal stability, heat suppression, smoke reduction, and toxicity mitigation, offering a design reference for developing high-performance, environmentally friendly bio-based flame-retardant RPUF.
Objective: The large-scale adoption of electric vehicles (EVs) has become a critical strategy for decarbonizing the transportation sector. However, uncoordinated EV charging behavior widens the peak–valley load difference of distribution networks compared with the base load without EVs, posing formidable challenges to grid operation and security. Studies on EV charging scheduling optimization overlook the effective coupling between carbon factors and electricity prices, preventing carbon costs from guiding user charging behavior through price. Methods: To address the above issue, this study proposed a dynamic pricing-based multi-objective EV charging scheduling model that incorporates carbon price and a carbon pass-through coefficient. The model considered the grid load peak–valley difference, operator revenue, and user charging cost as the objectives. Meanwhile, regret theory was adopted to characterize boundedly rational user decision-making behavior in choosing charging time periods and charging stations. To address the high-dimensional nonlinearity and multi-constraint characteristics of the model, as well as the tendency of existing solution algorithms to fall into local optima, this study developed adaptive weighting and disturbance factors based on the crested porcupine optimizer (CPO) to solve the above multi-objective nonlinear optimization problem. Taking the Nanshan District of Shenzhen city, China, as the case study with the IEEE 33 bus system, simulation experiments were conducted to validate and calibrate the rationality of the proposed dynamic pricing mechanism, the superiority of the improved CPO (ICPO), and the patterns of carbon price and carbon pass-through coefficient on the optimization performance of EV charging scheduling. Results: The results of this study were as follows. 1) The influence of carbon price and carbon pass-through coefficient on the multi-objective optimization of the system exhibited remarkable threshold effects, with a critical threshold identified at a carbon price of 70 RMB/t and a carbon pass-through coefficient of 0.3. Below this threshold, the peak–valley load difference remained relatively small, and user costs were low. Above this threshold, the user costs increased substantially, and the peak–valley load difference increased in some scenarios. The optimal parameter combination was a carbon price of 60 RMB/t and a carbon pass-through coefficient of 0.2, which achieved synergistic optimization between grid security and user benefits. 2) The ICPO demonstrated excellent optimization performance under the proposed carbon–electricity coordinated scheduling strategy. Compared with the genetic algorithm, particle swarm optimization, and the original CPO, the ICPO reduced the peak–valley load difference by 27.5%, 19.2%, and 18.4%, respectively. It also reduced user charging costs by 5.6%, 16.4%, and 6.6%, respectively. Furthermore, the final solution quality is improved by 14.9%, 10.3%, and 9.4%, respectively. 3) The number of charging users and user satisfaction exhibited a remarkable unimodal distribution pattern as the regret weight combination varied. Within the price regret weight range of 0.4–0.6, both indicators remained at relatively high levels, demonstrating a stable superior performance zone. Conclusions: The dynamic pricing model developed in this study can more effectively guide users toward coordinated charging behavior. The results demonstrate that an appropriate combination of carbon price and carbon pass-through coefficient enables the synergistic optimization of grid security and user benefits. The ICPO exhibits superior performance in reducing peak–valley load difference and user charging costs, highlighting the effectiveness of the proposed adaptive weighting and disturbance factor strategies in high-dimensional, multi-constraint optimization scenarios. A moderate configuration of regret weights is crucial for balancing user participation and system scheduling performance, whereas extreme weight settings significantly suppress users' willingness to charge and reduce satisfaction. This study provides scientific support for government carbon market policy design and for electricity pricing strategies by system operators. Future research could extend the proposed framework by incorporating a coupled transportation–distribution network model to jointly optimize travel routes and charging scheduling, thereby further enhancing its practical applicability.
Objective: Although digital twin technology has matured rapidly, prevailing transportation digital twins remain largely descriptive and reactive, focusing on static model replication and real-time data monitoring while lacking a closed-loop architecture capable of enabling end-to-end autonomous operation across the system's full lifecycle. Critically, current implementations omit persistent memory for experience retention and adaptive learning mechanisms for continuous evolution, leaving them fragile in complex, dynamic, or previously unseen operational contexts. Furthermore, dominant research paradigms optimize isolated components without cross-granularity coordination, thereby failing to realize system-level autonomy. To bridge these critical gaps, this study proposes a foundational design methodology for digital twins of autonomous transportation system (ATS), explicitly engineered to embed cognition, memory, and evolutionary intelligence into the core architecture. Methods: This study proposed a digital twin system design methodology for the autonomous operation of transportation systems by integrating autonomous operational mechanisms and digital twin technology with existing transportation systems, thereby providing support for realizing ATS. First, based on autonomous systems, a "perception, cognition, decision-making, control, execution, memory, and learning" autonomous operational mechanism was proposed to simulate human cognitive processes, which is then integrated into transportation systems to define the concept of ATS. Next, leveraging the concept of digital twin systems, a collaborative architecture comprising the physical, digital, and interactive spaces of transportation systems was constructed, highlighting the importance of digital space design in achieving ATS. To integrate autonomous mechanisms into transportation systems and address their inherent complexity, this study applied the autonomous operational framework to design digital twin systems in the digital space at multiple levels and scales, enabling autonomy. A multilevel architecture encompassing road networks, road segments, and transport equipment was designed to enable coordinated management across macroscopic, mesoscopic, and microscopic granularities. The study integrated multiscale modeling and simulation capabilities across spatial, temporal, data, and logical dimensions, supporting a complete decision-making process from system-level situation awareness to individual vehicle behavior prediction. By leveraging distributed computing across cloud, edge, and end devices, the system achieved distributed deployment and efficient execution of cognitive, decision-making, control, memory, and learning functions, enabling autonomous evolution and learning. Finally, the proposed system was validated through application scenarios including intelligent obstacle avoidance, intersection conflict coordination, and peak-hour route planning. Results: Based on the proposed design methodology for digital twin systems supporting autonomous transportation operations, three representative scenarios were evaluated. In intelligent obstacle avoidance scenarios, after detecting obstacles, vehicles rapidly evaluated multiple avoidance and driving options through coordinated decision-making among vehicle-, road segment-, and road network-level digital twins, generating optimal avoidance strategies, thereby improving decision accuracy and control responsiveness. In intersection scenarios, coordinated signal and vehicle control enabled by digital twins allowed the system to better adapt to changing traffic demands across different time periods, weather conditions, and unexpected events, enhancing both intersection throughput and vehicle efficiency. In peak-hour scenarios, collaborative path planning across these digital twins improved the autonomous operational capability and travel efficiency of transport systems during high-demand periods. Conclusions: The proposed methodology transcends conventional digital twin paradigms by embedding cognitive architecture, long-term memory, and self-improving learning capabilities directly into the system's design. It provides a theoretically grounded framework and an implementation-ready blueprint for building ATS that are not only safe and efficient at runtime but also resilient to uncertainty, capable of lifelong learning, and able to evolve system-level intelligence over time. This study lays essential groundwork for next-generation transportation systems that anticipate, adapt, and autonomously improve.
Objective: Drivability is one of the key performance indices of electric vehicles and is closely related to the dynamic characteristics of the drivetrain. In-wheel motor drives differ from centralized motor drives in that the mechanical half-shaft is shorter or even eliminated, while the motor mass and corresponding rotational inertia are added directly to the wheel. These changes, together with the mass and stiffness distribution of the drivetrain system, give rise to new longitudinal–torsional coupled vibration characteristics in the low-frequency range. This study aims to investigate the underlying mechanisms and characteristics of longitudinal and torsional vibrations in the low-frequency range of vehicles driven by in-wheel motors and to clarify the differences between in-wheel motor configurations and their centralized counterparts. Methods: A dynamics model was established to describe the longitudinal–torsional coupled vibration of the in-wheel motor drive system in the low-frequency range. For torsional vibration, the motor and wheel were represented as equivalent rotational degrees of freedom, while the unsprung mass and vehicle body were modeled as translational degrees of freedom for longitudinal vibration. To capture the coupling between rotational and longitudinal motions, the longitudinal tire force was linearized under the assumption of small slip and combined with a first-order relaxation model. Based on the established models, dynamic analyses were conducted for three representative drivetrain configurations: centralized motor drive, in-wheel motor drive with inner rotors, and in-wheel motor drive with outer rotors. The configurations were compared by evaluating their natural frequencies, mode shapes, frequency responses, and transient responses under tip-in conditions. For the in-wheel motor drive with outer rotors, the effects of motor mass, rotor inertia, suspension bushing stiffness, suspension bushing damping, initial wheel speed, and motor time constant were also evaluated. Results: The dominant low-frequency vibration of the centralized motor drive system was a torsional mode caused by half-shaft flexibility, with a frequency of approximately 7.5 Hz. In contrast, in the in-wheel motor drive systems, longitudinal vibration of the unsprung mass, with a frequency of about 10 Hz, dominated due to the removal of the half-shaft and the increased unsprung mass. For the configuration with inner rotors, a rotor torsional mode appeared near 20 Hz. For the configuration with outer rotors, a rotor–wheel coupled torsional mode was observed at 26.6 Hz. Under tip-in excitation, all three configurations exhibited noticeable transient vibrations. The centralized motor drive system showed a single low-frequency torsional oscillation, whereas the in-wheel motor systems exhibited higher-frequency coupled oscillations. Parametric sensitivity analysis of the in-wheel motor drive configurations indicated that increasing suspension bushing damping effectively suppressed the first longitudinal mode vibration. Increasing the motor time constant reduced the amplitude of both longitudinal and torsional vibrations, although at the expense of a slower torque response. Conclusions: The vibration characteristics of in-wheel motor driving vehicles in the low-frequency range differ fundamentally from those of centralized motor driving vehicles. In the latter, the dominant vibration is a torsional mode governed by half-shaft flexibility, whereas in the former, the dominant motion is a longitudinal mode associated with the unsprung mass. The proposed model effectively captured the vibration characteristics of all three drivetrain configurations and provides a theoretical basis for structural design, parameter matching, and motor control in in-wheel motor drive systems.
Objective: Large language models (LLMs) excel at text generation and complex problem solving but are prone to factual and logical errors in multidimensional reasoning due to internal knowledge limitations. Although retrieval-augmented generation (RAG) mitigates this problem by integrating external knowledge, it introduces knowledge noise, multisource conflicts, and a disconnection between retrieval and reasoning. Structured approaches such as chain-of-thought and tree-of-thought enhance reasoning via explicit intermediate steps, but they rely on internal knowledge without effective fact-checking. Recent methods that fuse RAG with structured reasoning lack coordination in situations involving conflicts between external and internal knowledge, and tree architectures focus locally without ensuring cross-level consistency. To address these issues, this study develops a framework that deeply integrates retrieval with tree-structured reasoning and enforces consistency constraints to guarantee factual accuracy and logical coherence. Methods: This study proposed the Consistent Thought Tree with Retrieval (CTTR) framework, featuring three core designs. First, dynamic retrieval embedding enabled context-aware knowledge acquisition at each tree node for on-demand external knowledge fusion. For each node, a retrieval query was built from the content and context, and the top-k documents were retrieved via cosine similarity. A reading-note mechanism processed the retrieved content through the LLM to ground each branch in accurate facts. Second, knowledge consistency verification resolved conflicts by comparing the semantic consistency among three results: those obtained from internal knowledge, external retrieval, and fused synthesis. Their embeddings were computed, and similarities were ranked; finally, the highest pair was selected as the output. Irrelevant branches were pruned to prevent noise propagation. Third, node consistency evaluation applied majority voting and confidence assessment among nodes of the same level, thereby enhancing reliability via multipath convergence and avoiding single-chain deviations. The algorithm processed each tree layer through retrieval, verification, optimization, and integration to produce the optimal result. Results: Experiments encompassed four tasks using GLM4-9B and DeepSeek-R1-Distill-Llama-8B: hallucination detection (TruthfulQA with 817 questions, HalluQA Chinese with 452 questions), code generation (HumanEval with 164 problems), the 24-Game (100 problems from 4nums), and creative writing (science fiction and historical narrative evaluated by GPT-4 scoring and human review). CTTR achieved 66.50%/57.78% on GLM4-9B and 91.31%/45.56% on DeepSeek-R1-Distill-Llama-8B for hallucination detection, averaging an improvement of 2.89% over the retrieval-augmented thought tree (RATT). Code generation reached 73.17%/87.20% and 75.61%/88.41% pass@1/pass@5, showing an improvement of 3.96%. Further, the 24-Game success rates of 11% and 24% indicated an improvement of 3.50%. Creative writing scores reached 8.96/8.81 and 9.57/8.45, reflecting an improvement of 1.24%. Ablation studies confirmed that removing knowledge selection (KS) or consistent nodes (CN) degraded the performance, with KS contributing the most; this validated the necessity of dual consistency. Resource analysis showed that compared to RATT, CTTR reduced token consumption by 11% and the time required to 247.67 s by pruning redundant retrievals. Poisoning attack tests demonstrated the robustness of CTTR against malicious noise injection. Conclusions: CTTR effectively addresses the need to maintain logical consistency and factual accuracy for complex problems. By embedding retrieval into tree-structured reasoning with dual consistency evaluations, it enhances hallucination detection, code generation, mathematical reasoning, and creative writing. The results confirmed that retrieval–tree fusion strengthens complex reasoning, and that dual consistency mechanisms are task-adaptive. KS is prominent in fact-intensive tasks, while CN is valuable in open-domain generation. CTTR improves accuracy while also optimizing resource efficiency. As a training-free framework deployable without fine-tuning, CTTR reduces application costs. Future work will explore the dynamic mechanisms that adaptively switch paths by task complexity, using direct responses for simple tasks and tree reasoning for complex ones to balance efficiency and accuracy.
Objective: In recent years, pretrained models have achieved remarkable progress in tasks such as speech recognition and speaker identification, significantly enhancing performance across a wide range of downstream applications, particularly when trained on large-scale speech datasets. However, the performance of pretrained models depends heavily on the quality of pseudolabels, and obtaining large quantities of audio data with true annotations is far more costly than acquiring unlabeled data. Therefore, improving the quality of pseudolabels generated from large-scale unlabeled data is critical. To address this challenge, this paper proposes a progressive clustering strategy to optimize pseudolabels produced by unsupervised clustering algorithms. Methods: The progressive clustering process involved the following steps. First, HuBERT was trained on an unlabeled dataset, and the trained model was used to extract speaker embeddings from the audio data. Next, unsupervised clustering methods were applied to these embeddings to generate initial pseudolabels for each audio sample based on the clustering results. For each pseudolabel category, the mean of all speaker embeddings within that category was calculated as its centroid. Cosine similarity was then computed between the centroids of different categories. When the cosine similarity met a predefined threshold, the corresponding pseudolabel categories were merged to obtain a final, high-quality pseudolabel set. Results: Four algorithms, namely K-means, agglomerative hierarchical clustering, Infomap, and Leiden, were used to cluster the unlabeled Librispeech dataset to generate pseudolabels. After applying the progressive clustering strategy, the F1-score, normalized mutual information, and cluster accuracy metrics of the pseudolabel set all improved, while the Coverage metric decreased. These results indicated that the quality of the pseudolabels improved significantly, demonstrating the positive effect of pseudolabel merging on clustering optimization. A slight decrease in purity was observed, which was attributed to the merging process inevitably combining some incorrectly assigned categories, thereby increasing intraclass complexity. This suggested that, in pursuit of higher category consistency and label accuracy, and in improving overall pseudolabel quality, some loss in cluster purity was unavoidable. Based on this strategy, a pretrained model incorporating joint speaker information modeling—speaker-related HuBERT (SR-HuBERT)—was trained. Experimental results showed that the model achieved a speaker identification accuracy of 85.95% and a speaker diarization error rate of 4.39% on speech processing universal PERformance benchmark speaker-related tasks, outperforming publicly available pretrained models. To further validate the generalization capability of the progressive clustering strategy, additional experiments were conducted on the VoxCeleb2 dataset, where pseudolabel quality also improved significantly after applying the strategy. The speaker representations extracted by SR-HuBERT were evaluated on VoxCeleb1-O, where the equal error rate metric decreased from 12.82% to 12.73%. These results indicate that the proposed strategy effectively enabled the model to learn more discriminative speaker representations. Conclusions: This study proposes a progressive clustering strategy for unsupervised clustering to address the issue of pseudolabel quality in unlabeled audio data. The experimental results demonstrate that this strategy significantly improved pseudolabel quality. More importantly, under the guidance of higher-quality pseudolabels, model performance is further enhanced.
Objective: Synthetic aperture sonar (SAS) achieves high-resolution underwater imaging by synthesizing a large virtual aperture through platform motion. However, system parameters still constrain SAS imaging resolution, especially under low-frequency detection conditions. Additionally, echoes from small underwater targets are usually weak and can be easily contaminated by background noise and interference, which further reduces imaging quality and target detectability. To improve imaging resolution and enhance weak target echoes, this study proposes a weighted regularization-based signal enhancement SAS method, termed RSE-SAS. This method exploits the sparse spatial distribution characteristics of small underwater targets and aims to improve target focusing while suppressing background noise. Methods: A multisubarray SAS signal model was first established, and the imaging region was discretized into candidate scattering grids. Rather than directly fusing the signals collected by multiple subarrays, a mapping space for the imaging region was constructed using theoretical multichannel echoes from different candidate scattering positions. The received echoes were then mapped into this space. As a result, the imaging problem was transformed into the estimation of sparse projection coefficients. Based on the sparse distribution prior of small targets, an L1-norm regularization term was introduced to formulate the imaging problem as a sparse reconstruction problem. An iterative shrinkage-thresholding algorithm based on the majorization–minimization framework was employed for optimization. To alleviate amplitude underestimation caused by the fixed threshold in conventional sparse reconstruction, a weighted regularization strategy was further introduced. Target-related coefficients with relatively large amplitudes were assigned smaller penalty thresholds. In contrast, noise- and interference-related coefficients were subjected to stronger penalties. Therefore, target echoes were preserved and enhanced while background components were gradually suppressed during iteration. Point- and extended-target simulations were conducted. DAS, RD, conventional CS-SAS, and the proposed RSE-SAS were then compared in terms of imaging resolution, target detection probability, output signal-to-noise ratio, convergence performance, and computational efficiency. Results: The simulation results showed that RSE-SAS achieved better target focusing and background suppression than the comparison methods. For point-target imaging, DAS and RD exhibited relatively broad mainlobes, whereas CS-SAS improved target focusing while retaining noticeable background noise. RSE-SAS further suppressed the background components while maintaining a concentrated target response. Similar improvements were observed for extended-target imaging. Quantitative analysis based on 100 Monte Carlo trials showed that the azimuth and range −3 dB mainlobe widths of RSE-SAS were 0.29 and 0.81 m, respectively, compared with 1.46 and 1.82 m for DAS. The target detection probability increases from 0.77 for DAS to 0.98 for RSE-SAS, while the output signal-to-noise ratio increased from 12.3 to 31.6 dB. Compared with CS-SAS, RSE-SAS reduced the azimuth and range mainlobe widths by approximately 35.2% and 12.9%, respectively. It also increased the detection probability by 0.05 and improved the output signal-to-noise ratio by 3.1 dB. Furthermore, the target-related coefficients stabilized after approximately 10 iterations. With the same data scale and computational platform, the average single-frame imaging time for RSE-SAS was 52 ms, indicating that weighted regularization incurred limited additional computational cost. Conclusions: The proposed RSE-SAS method combines the sparse spatial prior of small underwater targets with a weighted regularization strategy. By applying different penalty strengths to target- and noise-related coefficients, this method alleviates amplitude underestimation in conventional sparse reconstruction and realizes simultaneous target echo enhancement and background suppression. The results demonstrate that RSE-SAS provides improved imaging resolution, target detection performance, noise robustness, and focusing capability while maintaining acceptable convergence and computational efficiency. The proposed method therefore has potential for high-resolution imaging of small underwater targets. Future work will focus on validation using measured SAS data and further evaluation under more complex underwater propagation conditions.
Objective: Prestressed concrete structures often require active strengthening measures to improve structural behavior and increase stiffness when there is insufficient prestress. The strengthening design needs to efficiently handle the complex connection relationships between new strengthening members and the original structure, aiming to minimize component clashes during the design stage. A scientific evaluation of the corresponding construction duration is also crucial for assessing the scheme's feasibility. Targeting a newly constructed continuous box-girder structure with irreversible prestress defects arising during construction, this paper designs a combined strengthening scheme using internal support trusses and external tendons, along with process simulation and intelligent static/dynamic duration prediction, using building information modeling (BIM) and gradient boosting decision tree (GBDT) methods to increase its stiffness and resistance reserve. Methods: First, specialized bridge structural finite element analysis software was used to analyze the improvement efficiency of truss strengthening on the girder stiffness and internal forces, thereby verifying the scheme effectiveness. Second, a BIM parametric model of the strengthening project was established to complete tasks such as clash detection and construction process simulation. Furthermore, upon integrating BIM feature indicators, the GBDT algorithm was used to train static duration prediction models for the four subprojects involved in the strengthening project: support erection, installation of internal steel support trusses, tensioning and anchoring of external prestressing strands, and bridge deck pavement construction. Thereafter, the accuracy of the prediction results was analyzed. Finally, for the internal steel support truss installation subproject, a GBDT dynamic duration prediction model that accounts for random clash rates across construction segments was built. This model's performance was analyzed through numerical experiments, and its prediction accuracy was compared with that of a linear prediction model. Results: The results revealed the following: 1) The internal support trusses were effective in improving the box-girder stiffness and bending moment distribution. For the controlling mid-support negative moment region, the bending moments caused by the dead load and the live load (lane load) were reduced by an average of 12% and 14%, respectively, but there was no effect on the shear force near the supports. 2) BIM technology identified and counted more than 1 300 clashes between the steel and concrete components, guiding the optimization of the strengthening details and eliminating clashes at the design stage. 3) The GBDT algorithm demonstrated high fitting and generalization capabilities on the relevant training sets for each strengthening subproject, with the coefficients of determination ranging between 0.89 and 0.93. 4) The GBDT dynamic duration prediction model showed higher accuracy and reliability than the linear prediction. Furthermore, the minimum coefficient of determination increased from 0.66 to 0.95. In particular, under unknown clash rate conditions, the former also had smaller prediction errors for the remaining duration, with the maximum error decreasing from 5.1% to 3.6%. Conclusions: The BIM–GBDT method efficiently completes the forward design of this strengthening project and accurately predicts the project duration, thereby demonstrating the feasibility of the active strengthening scheme for a prestressed concrete structure. The methodological workflow described in this paper can serve as a reference for similar projects.
Objective: As the demand for large-scale underground projects in western China grows, long tunnels inevitably cross active faults, exposing them to the threats of fault dislocation and high-intensity earthquakes. Previous studies mainly treated these hazards separately or through simple linear superposition, overlooking the complex stress path dependencies and sequential mechanisms inherent in real geological hazard chains. To establish a solid theoretical foundation for the safety assessment of cross-fault tunnels in high-seismicity regions, this study explores the structural dynamic responses and cumulative damage mechanisms of a typical tunnel crossing a strike-slip fault under two extreme temporal sequences: a dislocation-earthquake sequence and an earthquake-dislocation sequence. Methods: To simulate a horseshoe-shaped tunnel crossing an active strike-slip fault with a 70°dip angle, a detailed three-dimensional numerical model was developed using FLAC3D finite difference software. The 455 m wide fault area was carefully subdivided into a 270 m wide fault zone, a 20 m wide central fracture zone, and a 165 m wide fault influence zone. The Mohr-Coulomb strain-softening constitutive model was used to accurately represent plastic strain accumulation. During dynamic time-history analysis, a 4.0 m quasi-static horizontal displacement was applied to simulate the dislocation phase. For the seismic phase, a Ricker wavelet was converted into equivalent stress loads on a static viscous boundary. To validate the numerical approach, a 1:50 scaled physical model test was conducted using a split-box shear device and specialized analogous materials. Subsequently, a comprehensive parametric sensitivity analysis examined how fault dip angles (60°, 70°, and 80°), fracture zone widths (10-30 m), and eccentric fracture zone placements (left, middle, and right) affected the lining internal forces. Results: The experimental and numerical results aligned well, successfully reproducing typical "Sˮ-shaped bending and high-strain concentration in the fracture zone. The finding showed: 1) The two temporal sequences produced fundamentally different failure mechanisms. In the dislocation-earthquake sequence, the initial dislocation caused significant plastic damage. The subsequent earthquake triggered stress redistribution and a severe damage superposition effect, leading to strain increases of 323.27% and 425.49% at the tunnel crown and invert, respectively, with the maximum invert strain reaching 70 146.7 με. 2) The earthquake-dislocation sequence was mainly dominated by the fault dislocation. The initial earthquake caused only 0.15 με of elastic strain. but the fault dislocation was responsible for the structural failure, raising the crown strain to 62 547.2 με and the maximum compressive stress to –8.29 MPa. 3) The fault dip angle was a key factor controlling the lining internal forces. Under the dislocation-earthquake scenario, increasing the dip angle from 60° to 80° reduced the peak bending moment from 266.39 MN·m to 128.91 MN·m—a 51.61% decrease. 4) Fracture zone width and location were secondary influences but significantly affected local internal force distribution. Conclusions: Evaluating fault dislocation and seismic loads separately or through simple linear superposition significantly underestimates structural damage risks. Fault dislocation drives the "S"-shaped deformation and high-strain failure. The timing of earthquake intervention greatly influences stress evolution: the dislocation-earthquake sequence causes stress redistribution after damage, while the earthquake-dislocation sequence intensifies stress concentration induced by the dislocation.
Significance: Single-crystal (SX) blades made of nickel-based superalloys represent a major technological breakthrough in the aerospace sector. As core, critical components, they enable advanced modern aero-engines to achieve high thrust and high thrust-to-weight ratios. The elimination of grain boundaries in SX blades provides excellent high-temperature performance. However, recrystallization (RX) can reintroduce grain boundaries, severely degrading in-service performance. Consequently, RX defects are a significant concern throughout the fabrication and service lifecycle of SX blades. Effective control of RX during fabrication is crucial for improving blade yield, and pre-existing and newly formed RX during service pose a severe threat to engine safety and reliability. Progress: This paper reviews decades of studies on RX in single-crystal nickel-based superalloys. Based on the manufacturing procedures of SX turbine blades and the causes of RX formation, RX is classified into seven types. Three categories can be identified during solution heat treatment (following prior plastic deformation): structural RX induced by directional solidification shrinkage, spot-like surface RX caused by mold shell debonding, and RX induced by plastic strain at core pins due to core-pin fracture. Four categories are triggered by extrinsic mechanical strain during subsequent thermal processes (e.g., welding, coating, and service): sandblasting-induced surface RX, machining-induced surface RX, shot-peening-induced surface RX, and tension–compression deformation-induced RX. This classification provides a clear framework for targeted RX studies and control. Combined with microstructural characterization, numerical simulation, and machine learning, this paper further elaborates on the nucleation and growth mechanisms, formation laws, and critical conditions for each RX type. Nevertheless, monitoring RX nucleation sites in SX superalloys remains challenging. Existing study primarily focuses on grain growth at the micrometer scale, with insufficient atomic-scale imaging observations and in-situ studies to reveal its intrinsic nature. The paper also summarizes advancements and limitations in RX nondestructive testing techniques (e.g., synchrotron X-ray diffraction) and control strategies (e.g., fillet transition design and magnetic field treatment). Significant challenges persist in controlling structural RX for multibody SX blades and in inspecting and controlling inner-cavity structural RX in double-wall, ultra-cooled SX blades. The restriction of RX on blade structural design and performance improvement has become a growing area of concern. A new breakthrough integrating multiple engineering technical elements is urgently required for regulating spot-like surface RX. For core-pin fracture-induced RX, quantitative data criteria for core-pin material specification and process implementation are lacking, and reliable data guidance for evaluating the influence of rolled core pins in different blade regions on in-service performance is unavailable. Even when indispensable extrinsic mechanical strain is introduced during blade fabrication following solution heat treatment, its latent RX tendency during subsequent thermal operations and high-temperature service must be thoroughly assessed. Conclusions and Prospects: Finally, this paper explores and envisions the identification of the universal physical essence and specific laws of RX formation, as well as the construction of a systematic RX database. This is achieved through comprehensive analytical approaches, including cross-scale and in-situ characterization, thereby providing effective guidance for the precise control of all types of RX defects.
Objective: Hinge clearance in the aircraft rudder system reduces the motion accuracy and stability of the transmission structure, thus affecting the maneuverability, stability and flight safety of the aircraft. Therefore, it is necessary to perform accurate dynamic modeling of rudder systems with hinge clearance and analyze their nonlinear characteristics to guide design and optimization. Methods: The hinge-clearance contact force is modeled using the classical Hertz model and the modified Coulomb model implemented in Automatic Dynamic Analysis of Mechanical Systems (ADAMS). In this study, the rigid–flexible coupled dynamic model of a rudder system with hinge clearance considering a flexible rudder surface was developed using the virtual power principle and the mixed coordinate method. The fixed-step fourth-order Runge–Kutta integral solver was implemented in MATLAB for numerical simulations, and the correctness of the dynamic model was verified by comparing the obtained simulation results with ADAMS dynamic simulation results. Based on the dynamic model, numerical simulation experiments were performed to explore the influence of rigid–flexible coupling on the dynamic characteristics of the rudder system. Based on the theory of nonlinear dynamics, the bifurcation diagrams of a rigid body and a rigid–flexible coupled rudder system under different excitation frequencies were obtained. A specific frequency was selected to construct the phase trajectory diagram and the Poincaré map. The evolution of system dynamics caused by changes in the excitation frequency was analyzed, and the nonlinear coupling effect between the excitation frequency and structural flexibility was explored. The hinge clearance was increased to study its influence on the bifurcation diagrams of the rigid body system. Results: MATLAB numerical simulation results were overall consistent with the ADAMS dynamic simulation results. When only the first-order natural mode of the flexible rudder surface was considered, the response difference between the rigid–flexible coupled system and the rigid body system was very small. After the introduction of the flexible rudder, mainly the first two natural modes were excited. There was a sudden change in the bifurcation diagram of the rigid rudder system. The phase diagram and Poincaré map showed that the system was in a stable period-1 motion state in the low-frequency band, and that two attractors existed within the relative motion of the clearance. The system gradually transformed into a quasi-periodic motion state near the critical frequency of 21 Hz. Further, increased hinge clearance reduced the critical frequency of the rigid system. The rigid–flexible coupled rudder system resonated due to the excitation of the second-order mode, showing a resonance peak at 27.27 Hz, and the oscillation phenomenon was mitigated in the high-frequency band. Conclusions: All rigid rudder systems exhibit a critical excitation frequency that is associated with the clearance size. In rudder systems, the presence of a flexible structure reduces stiffness, and a coupling nonlinear effect is observed between the excitation frequency and structural flexibility. These research results clarify the coupling law between the excitation frequency and structural flexibility, which provides a reference for optimizing the structural design of the steering gear and improving the dynamic stability of the rudder system. It is necessary to further explore the coupling nonlinear effects between excitation frequency, structural flexibility and clearance size.
Objective: The resilience of coastal and offshore infrastructure against wave-induced scouring is a critical concern. Cemented granular materials (CGMs), which strengthen mass stability through interparticle bonds, present a potential solution. However, their design is limited by a lack of understanding of microscale fluid–solid interactions under dynamic hydraulic loading. This study aims to create a high-fidelity numerical framework to investigate these anti-scouring mechanisms, focusing on cementation distribution and two-phase flow. Methods: A coupled computational fluid dynamics and discrete element method (CFD-DEM) model was developed to simulate the complexities of two-phase flow (water and air) and irregular particle shapes. The model was based on a validated virtual domain/immersed boundary method. Realistic rubble shapes were modeled using clumped particles, where each irregular grain was represented by a rigid cluster of sub-spheres. The pore structure was analyzed via Voronoi tessellation, and a mesoscale bond model was implemented to define the spatial distribution of cementation. Bonds, generated as cylindrical elements connecting particle centers, were created if the Voronoi faces met a criterion based on a critical interparticle distance related to the cementation ratio (CR), which was the volume of cement over the total volume. These bonds followed a linear elastic model with tensile and shear strength limits; they broke once these limits were exceeded. To account for flow obstruction caused by cement within pores, cemented zones were modeled as an isotropic homogeneous porous medium using a Darcy–Forchheimer approach, with momentum sink terms added to the fluid equations. The DEM module (using LIGGGHTS) solved the translational and rotational motions of each clumped particle, considering contact, bond, and fluid forces. The CFD module (using OpenFOAM) solved the Navier–Stokes equations for the two-phase fluid. Coupling was achieved through momentum exchange, where fluid forces acted on particles, and particle motion constrained the fluid velocity via porosity-weighted correction. A dam-break wave scenario was simulated in a 3D numerical flume to scour a triangular rockfill mound composed of 292 clumped particles, representing a transient high-impact loading condition. Simulations were carried out for noncemented and cemented cases with CR values of 10%. Results: The model successfully replicated the fluid and contact force dynamics observed in noncemented rockfill. For the cemented cases, key findings included: the following: (1) Permeability control: cementation significantly reduced water infiltration. The noncemented mound exhibited deep seepage, while at CR = 5%, penetration was greatly decreased. At CR = 10% and 20%, water was mainly confined to surface flow, with minimal internal seepage. The effect saturated, with little difference between CR = 10% and 20%. (2) Micromechanical response: cementation bonds changed how internal forces were transmitted. They supported substantial tangential forces during wave action, and their progressive failure (via tensile/shear rupture) dissipated energy. This led to a decrease in the average normal contact force and an increase in tangential force, indicating dynamic rearrangement of force chains. (3) Structural integrity: despite similar peak fluid impact forces, cemented structures retained greater stability by stabilizing the granular skeleton and redistributing internal stresses through the bond network. Conclusions: An advanced CFD-DEM model is developed to reveal the micromechanisms behind CGMs' resistance to scour. Cementation improves performance not only by strengthening contacts but also by reducing permeability and enabling energy dissipation through bond failure. The effect of cement content is nonlinear; beyond around 10% CR, additional cement shows diminishing returns, which is important for cost-effective design. Future work will include more complex material models and realistic wave conditions to support engineering applications.
