AI in civil engineering最新文献

This study investigates the application of established open-source machine learning tools, specifically CatBoost, XGBoost, LightGBM, and TensorFlow, which are based on Forest and Radial Basis Function Networks, to predict and analyze the structural behavior of concrete arch dams. Utilizing the Karun-I dam as a case study, the research assesses the performance of various machine learning frameworks. The results demonstrate that Random Forest-based methods achieve superior prediction accuracy and computational efficiency in comparison to Radial Basis Function Networks. Additionally, the analysis emphasizes the critical influence of lake levels as the primary factor impacting dam displacement, as revealed through feature importance evaluation. Overall, this research underscores the promising potential of machine learning in enhancing structural health monitoring for large dams, offering significant insights that contribute to the improvement of safety measures and operational efficiency in dam management.

Civil engineering relies on data from experiments or simulations to calibrate models that approximate system behaviors. This paper examines machine learning (ML) algorithms for AI-driven decision support in civil engineering, specifically construction engineering and management, where complex input–output relationships demand both predictive accuracy and interpretability. Explainable AI (XAI) is critical for safety and compliance-sensitive applications, ensuring transparency in AI decisions. The literature review identifies key XAI evaluation attributes—model type, explainability, perspective, and interpretability and assesses the Enhanced Model Tree (EMT), a novel method demonstrating strong potential for civil engineering applications compared to commonly applied ML algorithms. The study highlights the need to balance AI’s predictive power with XAI’s transparency, akin to the Yin–Yang philosophy: AI advances in efficiency and optimization, while XAI provides logical reasoning behind conclusions. Drawing on insights from the literature, the study proposes a tailored XAI assessment framework addressing civil engineering's unique needs—problem context, data constraints, and model explainability. By formalizing this synergy, the research fosters trust in AI systems, enabling safer and more socially responsible outcomes. The findings underscore XAI’s role in bridging the gap between complex AI models and end-user accountability, ensuring AI’s full potential is realized in the field.

Predicting concrete compressive strength with limited data remains a critical challenge in civil engineering. This study proposes a novel framework integrating Model-Agnostic Meta-Learning (MAML) with SHAP (Shapley Additive Explanations) to improve predictive accuracy and interpretability in data-scarce scenarios. Unlike conventional machine learning models that require extensive data, the MAML-based approach enables rapid adaptation to new tasks using minimal samples, offering robust generalization in few-shot learning contexts. The proposed pipeline includes structured preprocessing, normalization, a neural network-based meta-learning core, and SHAP-based feature attribution. A curated dataset of 430 samples was used, focusing on 28-day compressive strength, with input features including cement, water, aggregates, admixtures, and age. Compared to standard models like XGBoost and Random Forest, the MAML framework achieved superior performance, with MAE = 3.56 MPa, RMSE = 5.55 MPa, and R² = 0.913. SHAP analysis revealed nonlinear interactions and dominant factors like water-cement ratio, curing age, and aggregate content. Statistical validation via the Wilcoxon Signed-Rank Test confirmed the significance of the model’s improvements (p < 0.05). Furthermore, SHAP insights closely align with domain knowledge and mix design principles, enhancing model transparency for practical application. This work demonstrates the applicability of meta-learning in civil engineering and provides a scalable, interpretable solution for strength prediction in real-world, data-limited conditions.

Rivers provide irreplaceable resources for human life, and the problem of water scarcity has attracted serious attention worldwide. In this study, Kashkan River located in Loristan Province of Iran was studied using data obtained from the database of Iran Water Resources Company (IWRC). Three distinct machine learning (ML) models – Regression Tree (RT), Random Search Regression Tree (RSRT), and Bayesian Optimization Regression Tree (BORT) – were utilized to enhance water resource management practices. The primary model used was RT, a method that uses Bayesian optimization and stochastic search algorithms to provide an accurate estimate of the maximum flow within a river. The two hybrid models, RSRT and BORT, were introduced to improve the model performance. Through a comprehensive comparison and analysis of the results generated by these models, valuable insights were gained. Among the three models, the RSRT model demonstrated superior performance and accuracy metrics in streamflow (SF) modeling, closely aligning its results with a DR line of 1, indicating an optimal fit. The BORT and RT models also achieved excellent results, with their performance being on par with that of the top-performing RSRT model.

This paper investigates the application of pre-trained Vision-Language Models (VLMs) for describing images from civil engineering materials and construction sites, with a focus on construction components, structural elements, and materials. The novelty of this study lies in the investigation of VLMs for this specialized domain, which has not been previously addressed. As a case study, the paper evaluates ChatGPT-4v’s ability to serve as a descriptor tool by comparing its performance with three human descriptions (a civil engineer and two engineering interns). The contributions of this work include adapting a pre-trained VLM to civil engineering applications without additional fine-tuning and benchmarking its performance using both semantic similarity analysis (SentenceTransformers) and lexical similarity methods. Utilizing two datasets—one from a publicly available online repository and another manually collected by the authors—the study employs whole-text and sentence pair-wise similarity analyses to assess the model’s alignment with human descriptions. Results demonstrate that the best-performing model achieved an average similarity of 76% (4% standard deviation) when compared to human-generated descriptions. The analysis also reveals better performance on the publicly available dataset.

Wastewater-based epidemiology (WBE) is emerging as an effective tool to provide early warnings of potential disease outbreaks within communities through detecting the presence of pathogens in wastewater before clinical cases are reported. Nevertheless, quantitative prediction of future clinical case is challenging as uncertainties of dynamic shedding and disease transmission patterns can lead to complex correlation between wastewater viral concentration and clinical cases. Such complexities, augmented by factors such as viral variant, public behavioral change, etc., make it challenging to develop empirical models or data-driven models to provide accurate prediction of disease case for public health policy makings. To address this gap, this study developed an iterative data-driven framework utilizing Long-Short Time Memory (LSTM) neural networks for multi-timestep real-time predictions of future clinical cases based on WBE. The proposed LSTM model structure integrates both wastewater and historical clinical data as inputs. The prediction framework enables the update of LSTM model as more WBE dataset become available to enhance its adaptability to evolving pandemic stages. This framework was applied for real-time forecasting of COVID-19 clinical cases based on dataset of Ohio Wastewater Monitoring Project from July 2020 to October 2023. The developed iterative LSTM models were proven to achieve excellent performance in making clinical case predictions at different stages of COVID-19 pandemic. Early warning threshold of viral surge was defined by moving percentile method and results showed that the model achieved over 90% accuracy in future clinical case prediction and therefore demonstrated high reliability in pre-warning of potential disease outbreaks. This framework was also found to possess strong transferability across diverse geographic regions. The impacts of social policies and events on model predictions as well as the ramification of this model for future pandemics warning are discussed.

This research examines the application of eight different machine learning (ML) algorithms for predicting the compressive strength of high-performance concrete (HPC). Achieving precise predictions is crucial for enhancing structural reliability and optimizing resource usage in construction projects. The analysis utilized the “Concrete Compressive Strength” dataset, sourced from UC Irvine’s publicly available ML repository. The models evaluated include Gradient Boosting Regressor (GBR), Extreme Gradient Boosting Regression (XGBoost), Random Forest (RF), Support Vector Regression (SVR), Artificial Neural Network (ANN), Multilayer Perceptron (MLP), Lasso, and k-Nearest Neighbors (KNN). To enhance performance, critical data preprocessing steps were undertaken, which involved feature scaling, cleaning, and normalization. Hyperparameter tuning via Grid Search (GS) and K-fold cross-validation further optimized the models. Among those analyzed, XGBoost and GBR achieved the highest predictive accuracy, with R² values of 93.49% and 92.09% respectively, coupled with lower mean squared error (MSE), mean absolute error (MAE), and root mean squared error (RMSE). SHapley Additive exPlanations (SHAP) analysis revealed cement content and curing age as the most significant factors affecting compressive strength. Validation against experimental data confirmed the reliability of XGBoost and GBR through consistent prediction patterns and close alignment with empirical measurements. The results establish ML as an effective approach for HPC strength prediction, offering advantages in computational efficiency and accuracy over conventional analytical methods.

Relational databases containing construction-related data are widely used in the Architecture, Engineering, and Construction (AEC) industry to manage diverse datasets, including project management and building-specific information. This study explores the use of large language models (LLMs) to convert construction data from relational databases into formal semantic representations, such as the resource description framework (RDF). Transforming this data into RDF-encoded knowledge graphs enhances interoperability and enables advanced querying capabilities. However, existing methods like R2RML and Direct Mapping face significant challenges, including the need for domain expertise and scalability issues. LLMs, with their advanced natural language processing capabilities, offer a promising solution by automating the conversion process, reducing the reliance on expert knowledge, and semantically enriching data through appropriate ontologies. This paper evaluates the potential of four LLMs (two versions of GPT and Llama) to enhance data enrichment workflows in the construction industry and examines the limitations of applying these models to large-scale datasets.

Concrete cracking poses a significant threat to the safety and stability of crucial infrastructure such as bridges, roads, and building structures. Recognizing and accurately measuring the morphology of cracks is essential for assessing the structural integrity of these elements. This paper introduces a novel Crack Segmentation method known as CG-CNNs, which combines a Clustering-guided (CG) block with a Convolutional Neural Network (CNN). The innovative CG block operates by categorizing extracted image features into K groups, merging these features, and then simultaneously feeding the augmented features and original image into the CNN for precise crack image segmentation. It automatically determines the optimal K value by evaluating the Silhouette Coefficient for various K values, utilizing the grayscale feature value of each cluster centroid as a defining characteristic for each category. To bolster our approach, we curated a dataset of 2500 crack images from concrete structures, employing rigorous pre-processing and data augmentation techniques. We benchmarked our method against three prevalent CNN architectures: DeepLabV3 + , U-Net, and SegNet, each augmented with the CG block. An algorithm specialized for assessing crack edge recognition accuracy was employed to analyze the proposed method's performance. The comparative analysis demonstrated that CNNs enhanced with the CG block exhibited exceptional crack image recognition capabilities and enabled precise segmentation of crack edges. Further investigation revealed that the CG-DeepLabV3 + model excelled, achieving an F1 score of 0.90 and an impressive intersection over union (IoU) value of 0.82. Notably, the CG-DeepLabV3 + model significantly reduced the recognition error for locating crack edges to a mere 2.31 pixels. These enhancements mark a significant advancement in developing accurate algorithms based on deep neural networks for identifying concrete crack edges reliably. In conclusion, our CG-CNNs approach offers a highly accurate method for crack segmentation, which is invaluable for machine-based measurements of cracks on concrete surfaces.

The accurate prediction of the deterioration of Continuously Reinforced Concrete Pavement (CRCP) is essential for the effective management of pavements and the maintenance of infrastructure. In this study, a comprehensive approach that integrates descriptive statistics, correlation analysis, and machine learning algorithms is employed to develop models and predict punchouts in CRCP. The dataset used in this study is extracted from the Long-Term Pavement Performance (LTPP) database and contains a wide range of pavement attributes, such as age, climate zone, thickness, and traffic data. Initial exploratory analysis reveals varying distributions among the input features, which serves as the foundation for subsequent analysis. A correlation heatmap matrix is utilized to elucidate the relationships between these attributes and punchouts, guiding the selection of features for modeling. By employing the random forest algorithm, key predictors like age, climate zone, and total thickness are identified. Various machine learning techniques, encompassing linear regression, decision trees, support vector machines, ensemble methods, Gaussian process regression, artificial neural networks, and kernel-based approaches, are compared. It is noteworthy that ensemble methods such as boosted trees and Gaussian process regression models exhibit promising predictive performance, with low root mean square error (RMSE) and high R-squared values. The outcomes of this study provide valuable insights for the development of pavement management strategies, facilitating informed decision-making regarding resource allocation and infrastructure maintenance. Future research could focus on refining models, exploring additional features, and validating results through real-world implementation trials. This study contributes to advancing predictive modeling techniques for optimizing CRCP infrastructure management and durability.