Data-driven rock mechanics has long been constrained by a chronic deficit: coordinated laboratory measurements of uniaxial compressive strength (UCS) and its physically related properties are not only limited in number but also unevenly distributed across rock types. To address this constraint, the present study proposes a diversity-tuned physics-informed generative adversarial network—termed PI-GAN—designed to produce synthetic UCS, porosity, and bulk-density data that preserve the fundamental mechanical relationships inherent to rock materials. The framework integrates adversarial learning with an intact-rock strength-envelope constraint, category-specific UCS–porosity–density correlation losses, moment and covariance matching, feature-envelope penalties, and a pairwise-distance diversity term. The model was calibrated on 944 complete records drawn from the P3 petrophysical database, grouped into dense crystalline, volcanic, pyroclastic, and sedimentary rock categories. The trained PI-GAN generated 944 synthetic samples, increasing the usable dataset to 1,888 records. The synthetic data preserved the dominant correlation structure, with the overall UCS–porosity correlation shifting from − 0.598 to − 0.675 and the UCS–density correlation changing from 0.566 to 0.650. Category-specific Fisher z-tests showed non-significant differences for 11 of 12 correlation pairs, with an average absolute correlation difference of 0.036. Distributional validation yielded a mean Kolmogorov–Smirnov statistic of 0.181, while diversity diagnostics produced a mean pairwise-distance ratio of 0.703 and a near-duplicate rate of 0.0
When natural disasters occur, damage to roads and bridges not only disrupts transportation networks and endangers public safety but also affects the efficiency of rescue and evacuation work. Therefore, in the disaster response stage, the ability to quickly assess the situation in disaster-affected areas is extremely important. With the rapid development of technology, remote sensing provides a more efficient method for creating disaster maps compared to traditional methods. However, when dealing with large-scale data, there are still limitations in traditional image analysis methods based on remote sensing data. To solve this problem, this study proposes a binary deep learning framework to determine whether the bridges or roads in remote sensing data have been damaged, and provides an automated method for image analysis based on remote sensing data. According to the results, the EfficientNet-b1-based deep learning model excelled with an accuracy of 94.09
The increasing frequency and intensity of typhoons in recent years due to climate change has heightened the need for rapid post-disaster assessment. Inspection of post-typhoon damage is required to assess damages to critical infrastructure and facilitate disaster management. Unmanned aerial vehicles (UAVs) are commonly used for the recognition of post-typhoon damage. However, existing UAVs require either to be manually teleoperated or to follow a pre-programmed flight path, which is significantly limited in terms of rapid or adaptive post-typhoon inspection. To address this critical limitation, this study introduces a novel agentic AI and aerial robotics-enabled model designed for autonomous inspection operations. We propose a three-layered conceptual framework that closes the perception–cognition–action loop, thereby enabling goal-directed mission reasoning. This framework is operationalized as the Agentic AI and UAV-enabled Autonomous Inspection and Assessment System (AAUIAS)—a comprehensive cyber-physical system architecture. Within this architecture, the core agentic AI layer independently performs world modeling, real-time damage assessment, and dynamic path replanning. The proposed model and system were evaluated through high-fidelity simulations of a typhoon-impacted urban environment. Results demonstrate that AAUIAS reduces overall inspection time by 34.5
Rapid post-earthquake health assessment of shear wall buildings is critical for effective disaster mitigation and swift rehabilitation, particularly in earthquake‑affected urban areas. This study developed a novel artificial intelligence (AI)-based approach to predict key structural health parameters—maximum displacement, maximum inter‑storey drift (MISD), base shear, and maximum overturning moment—without requiring physical measurements of structural response. Instead of instrumented monitoring, the model relies solely on earthquake ground motion records that are readily available online, enabling fast and reliable predictions. A robust AI model was trained on a curated dataset generated from validated finite element simulations, attaining a high regression accuracy of 97.528
The accelerating digitalization of the construction sector has brought into sharp focus the inefficiencies and labor-intensive nature inherent in traditional craftsmanship for historical architecture. In response, this study proposes an AI-driven generative design workflow that translates semantic inputs into 2D imagery and 3D models, enabling the systematic learning and replication of stylistic features from a quintessential southern Chinese architectural ornament—the Lingnan stucco relief. This approach moves beyond the conventional reliance on individual expertise and heuristic empirical methods. The structural performance of the AI-generated designs was evaluated through combined material testing and numerical simulations. Concurrently, hybrid additive manufacturing techniques were explored for rapid physical prototyping. Collectively, this integrated framework illustrates the feasibility of revitalizing traditional craft practices within a contemporary intelligent paradigm, offering a technological pathway for the preservation and renovation of historic built heritage.
The design of dynamically resilient concrete materials remains a complex, fragmented process that depends on iterative modelling, expert judgment, and poorly integrated workflows spanning structural analysis, material formulation, and seismic performance evaluation. To address this challenge, we develop an autonomous multi-agent system (MAS) that optimizes the life cycle design of fiber-reinforced concrete (FRC) for earthquake resilience. The system focuses on shear walls, beams, and columns that constitute the seismic load bearing envelope of multi-storey buildings. Rather than functioning as a general-purpose structural tool, the MAS is designed to intelligently generate FRC specifications, tailoring fiber type, geometry, and reinforcement ratios according to seismic inputs and project constraints that are automatically parsed by the system. The framework integrates agentic artificial intelligence (AgenAI) via an Agent2Agent (A2A) protocol with physical artificial intelligence (PhysAI) implemented through physics informed stochastic models for micromechanical prediction and energy dissipation damage analysis. A modified maximum entropy principle models fiber matrix interactions under heterogeneous uncertainty, enabling unbiased optimization of energy dissipative properties. In parallel, a CNN based computer vision pipeline employing a 3D U-Net architecture performs accurate segmentation and interpretation of micro-CT data. The architecture comprises five specialized AI agents built on Qwen-30B-A3B and Phi-4-14B large language models (LLMs), with context grounding provided by agentic retrieval augmented generation (ARAG) for domain specific decision making. Experimental validation shows that the system reduces human intensive design cycles by 75
The upward movement of an existing tunnel caused by excavation of a nearby foundation pit has emerged as a pressing issue in urban rail transit safety. Conventional data-driven prediction methods often lack physical interpretability, and their performance tends to deteriorate when only a small number of training samples are available. To fill this gap, the present study proposes a physics-guided gradient boosting regression tree (PG-GBRT) framework. This architecture integrates domain knowledge from geotechnical engineering into a machine learning model to achieve reliable tunnel deformation predictions. The framework consists of two hierarchical layers. In the first layer, a physics-based empirical formula captures the fundamental geometric relationships. It draws on three theoretical foundations: Gaussian settlement trough theory, the link between excavation depth and volume loss, and three-dimensional geometric effects. The second layer is a residual machine learning model designed to handle the complex nonlinear interactions that the physical component cannot adequately represent. The empirical formula adopts a multiplicative structure. Four key geometric parameters enter this formulation: excavation depth, vertical clear distance, crossing length, and intersection angle. The gradient boosting regression tree then learns systematic deviations from what the physical formula predicts. To validate the framework, we compile 154 engineering cases from major cities across China. On the training set, PG-GBRT achieves an R2 of 0.9023, an RMSE of 0.6123 mm, an MAE of 0.4567 mm, and a MAPE of 8.23
Post-fire curing techniques for thermally damaged concrete are widely discussed, yet no dedicated study has quantified the extent of strength recovery. This research introduces an empirical model built upon a comprehensive experimental database that accounts for key input variables: exposure temperature, re-curing method, re-curing duration, and water-to-binder (W/B) ratio. The model was developed using 464 collected datasets. Statistical evaluation through ANOVA and regression analysis identified temperature as the dominant factor influencing strength recovery (F-value = 109.37), followed by re-curing type, re-curing duration, and W/B ratio. An optimization analysis further determined the conditions that maximize strength recovery: temperatures of 150–250 °C, hybrid re-curing methods, re-curing periods of 94–182 days, and W/B ratios between 0.30 and 0.45. The proposed model predicts the strength recovery percentage with demonstrable accuracy. It was subsequently refined into a more conservative formulation suitable for design codes, incorporating a conservatism parameter of 53.2
Abstract Accurately predicting the uniaxial compressive strength of ultra-high-performance concrete is vital for optimizing material design and ensuring cost-effective construction. However, existing predictive models often fail to capture the nonlinear relationships inherent in complex UHPC datasets and suffer from extended training times, thereby limiting their accuracy and practical utility. From the perspective of artificial intelligence, this study proposes an enhanced predictive framework that integrates a kernel extreme learning machine with a globally strengthened whale optimization algorithm. The proposed framework introduces three novel mechanisms—adaptive inertia weight, variable helix search, and optimal neighborhood perturbation—designed to improve the algorithm’s global search capability, convergence speed, and solution stability. From the engineering application perspective, the proposed model is applied to three UHPC datasets to predict uniaxial compressive strength. Prior to modeling, feature normalization is employed to reduce the impact of dimensional inconsistency among input variables. Experimental results demonstrate the model’s superior predictive precision, achieving $$R^2$$ R 2 values of 0.8823, 0.8906, and 0.8951 across the datasets. These findings confirm that the integration of advanced AI optimization techniques significantly enhances the performance of predictive models in civil engineering, offering a reliable and efficient approach for ultra-high-performance concrete strength estimation in design and construction workflows.
Abstract Deep learning (DL) is increasingly used to support image-based post-earthquake damage assessment of local structural elements in buildings, yet the existing literature remains inconsistent in task formulation, dataset design, model evaluation, performance enhancement and field deployment readiness. This paper presents a systematic literature review conducted in accordance with the “Preferred Reporting Items for Systematic Reviews and Meta-Analyses” (PRISMA) framework, covering 46 primary studies published between 2015 and 2025 on DL-based structural element level post-earthquake damage assessment. Using a transparent protocol for study identification, screening, eligibility assessment, quality appraisal, and data extraction, the review synthesises the literature through a task-aware taxonomy comprising four DL configurations: damage classification, damage segmentation, damage detection and estimation of quantitative damage severity. The review examines how these task configurations relate to damage assessment targets, local structural elements, DL architectures, data acquisition methods, data modalities, data quality, preprocessing pipelines, evaluation protocols, and model enhancement strategies, including transfer learning, data augmentation, synthetic data generation and multimodal fusion. The synthesis reveals key gaps in the under representation of several structural elements and failure mechanisms, limited use of multimodal and realistic field data, inconsistent task specific evaluation protocols, weak evidence of model generalisation to unseen data, and the absence of structurally grounded damage assessment criteria. To improve dependable field deployment, the review highlights the need for data efficient models, stronger multimodal learning, stricter evaluation on unseen field data, and integration with real-time engineer facing assessment workflows, eventually supporting development of disaster-resilient build environments.
Particle shape and gradation govern force chain networks and interparticle contacts in dense granular systems, thereby determining macroscopic shear strength. To investigate their influence on the peak internal friction angle (φd), triaxial tests were conducted on four binary mixed granular materials with contrasting morphologies, including Nanhai calcareous sand (GZS), Fujian standard sand (FS), glass beads (GB), and glass sand (GS). Ensemble learning algorithms were then employed to model φd from shape and gradation parameters. Results show that particle shape dominates φd: irregular particles yield systematically higher φd, whereas gradation plays a secondary role. Fine particle content (FC) induces non-monotonic effects. Angular particles (GZS, GS) exhibit an inverted V-shaped trend, with peak φd at a threshold FC. In contrast, subrounded particles (FS, GB) display a V-shaped trend, with minimum φd at the threshold FC. For subrounded particles, the threshold FC correlates strongly with the size ratio (SR). Heatmap analysis further identifies aspect ratio (AR) as the shape parameter most strongly correlated with φd. XGBoost and multilayer perceptron (MLP) models were developed to predict φd from shape and particle size inputs. Both models achieve good fit, confirming the relevance of the selected features. Due to its superior handling of small-sample data and multi-feature interactions, XGBoost significantly outperforms MLP in both accuracy and stability, offering a robust tool for predicting shear strength in granular materials.
The increasing frequency of wildfires, particularly in wildland-urban interface zones, has elevated the hazard posed by post-wildfire debris flows. Following combustion, surface and near-surface soils often become hydrophobic, a phenomenon that occurs predominantly on sand-based hillslopes. Rainwater and eroded soil accumulate downslope, frequently evolving into destructive debris flows. Soil hydrophobicity exacerbates erosion, distinguishing post-wildfire debris flows from natural debris flows in terms of intensity, duration, and destructive potential. Therefore, understanding the timing and conditions under which such debris flows initiate is critical. This initiation results from coupled effects among several key parameters: rainfall intensity (RI), slope gradient (δ), water entry value ( Ψ_wev ), and grain size (D50). Machine Learning (ML) techniques have become increasingly valuable in geotechnical engineering due to their ability to model complex systems without predefined assumptions. This study applied multiple ML algorithms: multiple linear regression (MLR), logistic regression (LR), support vector classifier (SVC), K-means clustering, and principal component analysis (PCA) to predict and classify outcomes from laboratory experiments that model field conditions using a rain device on various soils in sloped flumes. The results indicate that while MLR performs adequately in predicting total discharge, its accuracy in predicting erosion is limited, particularly for coarse sand. In contrast, LR and SVC achieve satisfactory accuracy in classifying failure outcomes, supported by clustering and dimensionality reduction techniques. Sensitivity analysis reveals that fine sand is highly susceptible to erosion, especially under low-intensity, prolonged rainfall. Furthermore, the first ten minutes of high-intensity rainfall emerge as the most critical period for discharge generation and slope failure. Collectively, these findings underscore the potential of machine learning to enhance post-wildfire hazard assessment and inform emergency response planning.
The Broad Learning System (BLS) provides an effective framework for nonlinear mapping, offering advantages over traditional deep neural networks through its expanded input node architecture. While BLS has demonstrated improved classification accuracy and reduced computational cost, its performance can be compromised by randomly initialized input weights and biases. To address this limitation, this study proposes an integration of metaheuristic optimization algorithms with BLS (termed MBLS). Five metaphor-free optimization methods and four algorithm-specific, parameter-free methods are independently employed to optimize BLS parameters, yielding nine hybrid models. These models are applied to four benchmark datasets for predicting the compressive strength (CS) of concrete structures. Although various machine learning (ML) and deep learning (DL) methods have been explored for this task, their practical utility is constrained by structural and computational complexity. In contrast, the proposed MBLS framework achieves both structural simplicity and computational efficiency. The predictive performance of the nine hybrid MBLS models, along with a multilayer perceptron artificial neural network (MLPANN), is evaluated on four real-world datasets using four performance metrics: mean absolute percentage error (MAPE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). To further enhance prediction accuracy, the training data are augmented with interpolated samples. Extensive experimental results, comparative analyses, and statistical tests confirm the effectiveness of the MBLS methods. Among them, the BL-BMR model consistently achieves the best overall performance, evidenced by the lowest average MAPE, RMSE, and MAE, and the highest R2 across datasets. Specifically, adopting BL-BMR forecasts yields MAPE improvements ranging from 8
Maintaining the structural integrity of reinforced concrete bridges necessitates the timely and accurate detection of surface defects. Conventional inspection methodologies remain labor-intensive, inherently subjective, and susceptible to human error, driving the need for automated assessment frameworks. This study introduces a multi-label defect classification model tailored for reinforced concrete bridge inspection, engineered to process imagery consistent with prevailing bridge inspection standards. The proposed framework is designed to simultaneously identify multiple co-occurring defects within a single image, addressing the practical reality of overlapping deterioration mechanisms. Leveraging the open-source Concrete Defect Bridge Image Dataset (CODEBRIM), three distinct ImageNet-pretrained deep neural network architectures were subjected to systematic hyperparameter optimization and fine-tuning to enhance classification performance across bridge-relevant defect categories. Beyond achieving high per-class accuracy, the optimized model attained a subset accuracy of 84.0
Landslides, as prevalent geohazards, exhibit complex and nonlinear evolutionary dynamics, frequently triggered by the coupled effects of reservoir water-level fluctuations and extreme precipitation. Such events are often characterized by abrupt, step-like deformation, posing significant challenges for accurate long-term displacement forecasting. To address the limitations of conventional models, including poor generalization, low robustness to chaotic disturbances, and insufficient capacity for nonlinear representation, we propose a hybrid deep learning framework termed GRU–TimeMixerKAN. This model synergistically integrates the sequential modeling capabilities of Gated Recurrent Units (GRU), the temporal-feature decoupling mechanism of TimeMixer, and the high-order nonlinear approximation power of the Kolmogorov–Arnold Network (KAN). Enhancements such as differencing-based detrending, sliding-window sampling, and automated hyperparameter optimization via Optuna are incorporated to further refine performance. The efficacy of the proposed model is evaluated using long-term displacement monitoring data from three reactivated reservoir landslides in the Three Gorges Reservoir Area (TGRA), with its performance benchmarked against nine state-of-the-art deep learning baselines. The results demonstrate that GRU–TimeMixerKAN consistently achieves the lowest Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), alongside competitive Symmetric Mean Absolute Percentage Error (sMAPE) and the highest Coefficient of Determination (R2). These findings underscore its superior capability in capturing displacement trends, responding to sudden changes, and generalizing robustly across diverse landslide cases. This study presents an effective and scalable methodology for advancing intelligent early warning and prediction systems for landslides.
Structures are prone to damage. Identification and localization of damage at its initiation stage are extremely helpful for ensuring safety, economy, and operational benefits. A machine learning (ML) approach is helpful for this purpose, provided that high-quality and sufficient damage signals are available from a variety of locations across the structure. Such signals, generated at damage initiation, are frequently obtained using a non-destructive testing (NDT) technique, such as acoustic emission (AE), which employs the pencil lead break (PLB) method. However, PLB is not possible at inaccessible locations of the structure. Therefore, synthetic experimental signals are required for such locations. Accordingly, the present study aims to generate synthetic experimental signals from numerically simulated AE signals using an artificial neural network (ANN). Here, parameters from numerical signals serve as inputs, and the corresponding parameters of experimental signals are outputs. The most relevant signal parameters are determined using the Pearson correlation coefficient (PCC). The developed model is found to perform very well, achieving an accuracy of around 99
Accurate prediction of compressive strength is essential for improving the performance and durability of Engineered Cementitious Composites (ECC) in construction applications. Traditional methods often fall short in accounting for the complex interactions between material properties, such as fiber type, matrix composition, and curing conditions. To address this challenge, this study presents an advanced ensemble learning framework based on a dataset of 313 ECC samples characterized by 18 key features. The ensemble model integrates three base learners, namely Random Forest (RF), Gradient Boosting Decision Trees (GBDT), and Support Vector Regression (SVR), along with a meta-learner selected from ten candidate models. The proposed ensemble model demonstrates significantly higher prediction accuracy compared to conventional approaches. The results show that the ensemble model achieves a coefficient of determination (R2) of 0.896, a root mean square error (RMSE) of 5.734, and a mean absolute error (MAE) of 4.505, substantially outperforming individual models. Among the evaluated meta-learners, Lasso Regression was identified as the optimal choice. Its regularization capability effectively mitigated overfitting and enhanced generalization, leading to a notable improvement in the final predictive performance of the stacking framework. Furthermore, SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDP) were employed for model interpretability and visualization. The analysis reveals that factors such as fiber elastic modulus, silica fume content, and fiber volume fraction significantly contribute to the enhancement of ECC compressive strength. This model provides practical insights for optimizing the design and application of ECC materials.
In response to the housing shortage in Canada, particularly in northern and remote communities, modular houses have emerged as a viable solution. These prefabricated structures offer speed, cost-efficiency, and flexibility. To enhance the durability and functionality of these modular homes, innovative construction techniques are being explored. A new bolted connection, utilizing high-strength long bolts, has been introduced for hollow structural sections (HSS), which can be designed using regression models trained by an experimentally validated finite element model (FEM). This study employs machine learning techniques, including neural networks, genetic regression, and decision trees, to detect the failure mode and predict the ultimate moment capacity of HSS moment connections under monotonic loading. A nonlinear validated FEM was developed using LS-DYNA software, and a matrix of 240 FEMs was generated to train and test the machine learning models, including a range of various design parameters such as the extended plate thickness, number of bolts, bolt arrangement, and bolt diameter. Five machine learning algorithms were used for classification and regression learning, with hyperparameter optimization applied to enhance their accuracy. Mathematical formulas for predicting the ultimate moment capacity were developed using genetic algorithm-based symbolic regression, trained on 70
Abstract This preliminary study introduces and evaluates a router-based multi-agent framework for automated foundation design calculations through intelligent task classification and expert selection. Three configurations were assessed: single-agent processing, multi-agent designer-checker architecture, and router-based expert selection, using baseline models including DeepSeek R1, ChatGPT 4 Turbo, Grok 3, and Gemini 2.5 Pro. Initial evaluation on 27 test cases with triple-trial execution shows promising performance: the router-based system achieved 95.00% for shallow foundations and 90.63% for pile design, representing improvements of 8.75 and 3.13 percentage points over standalone Grok 3, respectively, and outperforming conventional workflows by 10.0–43.75 percentage points. Grok 3 demonstrated superior standalone performance, indicating enhanced large language model (LLM) mathematical reasoning capabilities. The dual-tier classification framework successfully distinguished foundation types, enabling appropriate analytical approaches. While these preliminary results suggest router-based multi-agent systems as a promising approach for foundation design automation, the limited sample size necessitates comprehensive validation on larger, more diverse datasets before deployment recommendations. Safety–critical requirements necessitate continued human oversight in professional applications. This work provides a methodological foundation for future research in AI-assisted geotechnical engineering.
The construction industry faces the challenge of decarbonization. Integrating manufacturing principles and Artificial Intelligence (AI) offers a promising pathway to reduce CO2 emissions, specifically by integrating CO2-emission variables into AI-driven production schedules. However, transparency to users is essential, as human users remain ultimately responsible for production outcomes. This requirement can be met through Explainable AI (XAI), which aims to provide transparency for end users. However, defining an appropriate XAI approach requires understanding problem- and industry-specific variables. Using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) methodology, this study examines the state of the art in XAI literature to identify research gaps and formulate actionable recommendations. The study provides insights for developing an XAI approach to support the decarbonization of housing manufacturing explicitly. The key findings highlight the need for user-centric and industry-specific frameworks and the importance of clearly defining the XAI-AI relationship. Finally, this research synthesizes these findings into a roadmap to guide future research on XAI for the decarbonization of housing manufacturing.
