AI in civil engineering最新文献

This study developed an integrated numerical and data-driven framework for predicting the free-vibration characteristics of thin-walled curved box-girder bridges, a widely used yet mechanically complex structural form in modern bridge engineering. A computationally efficient one-dimensional thin-walled beam finite element method (FEM) was implemented in MATLAB, explicitly incorporating torsional, distortional, and warping effects, which are critical for accurately representing the dynamic behavior of curved girders. The proposed model was rigorously validated against detailed ANSYS shell-element simulations and published experimental data, demonstrating close agreement in both natural frequencies and corresponding mode shapes. A systematic parametric study was conducted to evaluate the influence of key design variables, including curvature radius, span length, boundary conditions, diaphragm layout, and cross-sectional geometry, on the first three modal frequencies. This process generated a comprehensive dataset, which then served as the basis for developing multivariate linear regression models. The resulting models yielded explicit predictive equations with excellent accuracy, with R² values exceeding 0.999 and root mean square error (RMSE) not greater than 0.31 Hz. The principal contribution of this work lies in its hybrid methodology, which effectively combines physics-based FEM with data-driven regression modeling. This dual approach not only deepens mechanistic insight but also delivers practical utility. The derived closed-form expressions offer engineers an efficient preliminary design tool, significantly reducing the dependency on computationally intensive finite element simulations during early design phases.

Accurate prediction of long-term settlement under complex traffic loads remains a pivotal challenge for the safety and durability of transportation infrastructure. While explicit models for settlement calculation have been advanced to handle general three-dimensional stress states, a major practical hurdle lies in determining reliable model parameters. Parameter inversion offers a viable path to high-fidelity estimates, yet conventional inversion techniques often fall short in accuracy. Ensemble learning methods can improve data precision by synthesizing predictions from multiple intelligent models; however, commonly used soft voting strategies tend to overlook both systemic bias across base models and the distinct contribution of each predictor. To address this, this study proposes a Particle Swarm Optimization-Back Propagation Neural Network-Random Forest (PSO-BPNN-RF) inversion model that incorporates a refined soft voting method. Coupling this inversion model with a three-dimensional explicit settlement calculation framework for complex traffic loading enables high-precision parameter identification. The proposed approach is subsequently applied to parameter inversion for an explicit model of the Xiaoshan Airport taxiway, demonstrating strong generalization capability and superior accuracy.

This study proposes a comprehensive performance evaluation and intelligent decision support system for the maintenance and seismic retrofitting of aging transportation infrastructure, aimed at enhancing structural safety, extending service life, and optimizing life-cycle costs. The research focuses on reinforced concrete (RC) bridge columns commonly found in urban elevated railway systems in Japan, addressing key issues such as strength degradation, insufficient ductility, and inadequate seismic performance. Using static nonlinear analysis, the residual load-bearing capacity and damage state of the columns were evaluated, and a comprehensive performance index system was established. To enhance structural resilience while minimizing operational disruption, a space-efficient seismic reinforcement method characterized by high spatial adaptability was adopted, making it particularly suitable for dense urban environments. The decision-making process is underpinned by the Adaptive Integrated Digital Architecture Framework (AIDAF), which establishes a closed-loop system integrating data acquisition, performance assessment, parameter optimization, and feedback validation. By incorporating machine learning (ML), specifically the random forest (RF) algorithm, into the AIDAF framework, a data-driven retrofitting system was developed. Feature importance analysis identified key variables, including steel plate thickness, rebar diameter, and spacing. The ML-enhanced system reduces design iteration time and facilitates rapid evaluation of multiple reinforcement configurations. The predictive accuracy of the model was validated using an in-service railway viaduct, confirming its effectiveness. Furthermore, the study recommends integrating explainable AI techniques to improve transparency and regulatory acceptance. The findings demonstrate that the proposed ML-AIDAF framework is technically feasible, economically viable, and scalable for real-world infrastructure retrofitting projects.

Cracks and water seepage are common structural safety hazards in excavation and pit support system. Traditional methods usually rely on a lot of manpower and material resources, and there are some problems in the monitoring process such as low efficiency, long time, incomplete data collection and insufficient accuracy, which cannot meet the needs of modern engineering construction. In recent years, the construction industry has gradually changed to the trend of intelligence and automation, and machine vision has entered the field of vision. It can not only effectively reduce labor costs, but also improve the overall accuracy of monitoring. However, previous machine learning framework usually uses a two-stage monitoring method, which takes a long time including the collection and process of data separately. This paper focuses on pit support systems and provides an overview and comparison of the application of machine vision and machine learning technologies. Furthermore, a real-time defect detection method based on the improved YOLOv8 algorithm, which can process the collected crack data and water seepage pictures, give the physical characteristics of the crack, and mark the location of water seepage, has been proposed and verified. Additionally, a practical project in Huzhou serves as a case study, where the established method has been applied. The actual implementation shows that the model also has good robustness under complex foundation pit conditions.

Due to climate change, permafrost regions are undergoing rapid evolution, posing a serious threat to roads, pipelines, foundations and other geotechnical infrastructure. Conventional methods for monitoring and predicting permafrost degradation have limitations in spatial coverage, temporal resolution and environmental dynamic adaptability. In recent years, the development of machine learning (ML) has opened up a new way to simulate the complex interaction between thermal state, soil properties and atmospheric variables in cold regions. This paper reviews the emerging applications of ML technology, from supervised learning models such as Random Forests (RF) and Support Vector Machines (SVM), to deep learning frameworks such as Convolutional Neural Networks (CNN), in predicting the thawing depth of permafrost, the evolution of ground temperature and the phenomenon of thermokarst. We systematically classify the application of ML according to the input data types (remote sensing, in-situ sensors, satellite climate data) and geotechnical output variables (thermal conductivity, soil strength, bearing capacity), and discuss the practice of combining ML with physical process model to enhance the interpretability and generalization ability. This review pays special attention to the risk of soil weakening, foundation instability and infrastructure failure caused by permafrost melting in the Arctic and subarctic regions. Moreover, this paper points out the key challenges such as data scarcity, lack of cross regional mobility and lack of uncertainty quantification. By systematically integrating the latest research results, data sources, model architecture and evaluation indicators, this review provides a basic reference for researchers and practitioners engaged in climate adaptive geotechnical engineering. The research results highlight the potential of ML as a transformative tool in permafrost geotechnical engineering, thereby facilitating environmental monitoring, risk assessment and infrastructure planning.

This study presents an approach that demonstrates the capabilities of Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) in solving mechanics-related problems, particularly in the design of monolithic reinforced concrete slabs on a base. For the first time, a voxel-based representation of the studied object is proposed. In many cases, the design stage involves the inclusion of technological holes of various shapes, and the slab surface may have complex geometry. Determining the stress–strain state (SSS) using closed-form solutions under such conditions is highly labor-intensive or even unattainable. This paper presents an alternative approach using a 3D CNN with a U-Net architecture, Deep Convolutional Generative Adversarial Nets (3D-DCGAN), and an Improved GAN (I-GAN). This method enables accurate prediction of shrinkage stresses and displacements in slabs more efficiently than the finite element method (FEM). The paper highlights the promising potential of neural networks in structural engineering.

This study presents a deep learning framework for non-destructive evaluation of concrete compressive strength using high-resolution microstructural images. Unlike traditional destructive testing, this approach enables efficient large-scale and continuous strength monitoring. The proposed model combines: (1) CAE for efficient feature extraction (achieving 80% dimensionality reduction without significant information loss); (2) Transformer-based self-attention mechanisms to dynamically weight critical image regions, enhancing interpretability; and (3) LSTM networks to capture temporal strength evolution during curing, improving forecasting accuracy by 15%. The framework is trained and tested on a hybrid dataset integrating UCI concrete strength data with high-resolution microstructural images. Nested cross-validation coupled with Bayesian optimization ensures robust performance evaluation and hyperparameter tuning. Comparative analyses demonstrate superior performance over baseline CNN and traditional ML models, with 20% reduction in MAE (3.7 MPa vs. 4.6 MPa), 18% lower RMSE (4.9 MPa vs. 6.1 MPa), and 7% higher R² (0.87 vs. 0.81). The model also reduces prediction time by approximately 20%. This scalable solution offers high accuracy, robustness, and generalizability for real-time concrete strength monitoring in infrastructure projects, advancing intelligent image-based non-destructive testing beyond conventional destructive methods.

The scarcity of experimental data poses a significant challenge in predicting the self-healing capacity of bacteria-driven concrete. To address this issue, we explored the use of synthetic data generation to augment the limited available dataset. By creating a synthetic dataset derived from real-world data, we substantially expanded the original data volume. We then trained and evaluated multiple machine learning (ML) models, encompassing both probabilistic and ensemble methods, for predicting self-healing capacity. Our comparative analysis revealed that ensemble methods, specifically the random forest (RF) algorithm, achieved the highest performance with an accuracy and F1-score of 0.863, surpassing the probabilistic models. Furthermore, when applied to real-world cases, the models maintained high predictive accuracy. This work confirms the value of synthetic data for enhancing the accuracy and reliability of predictive models in civil engineering, especially in data-scarce contexts. Our findings underscore the potential of machine learning and artificial intelligence to transform concrete research and highlight the role of synthetic data in overcoming common data limitations.

This research is based on the idea that certain cognitive-intensive tasks typically carried out by structural engineers—such as choosing how to effectively arrange a building’s structure—can be successfully automated. In this article we investigate two techniques widely used in the field of Artificial Intelligence, namely Monte Carlo Tree Search (MCTS) and Genetic Algorithm (GA). Following a tabula rasa approach, according to which no hints nor external data are used as a reference for navigating the search space, we demonstrate how structural designs of two-dimensional multi-storey reinforced concrete structures can be generated, with no human intervention, by implementing and combining the two techniques from a reinforcement-learning perspective. The design tasks assigned to the developed software agents concern civil structures under static and seismic loads, and the basis for comparison is represented by a combination of construction cost and greenhouse gas emissions associated with the making of the structures. In the article, based on the main concepts of Computational Mechanics, a theoretical framework is introduced, which allows to represent both structures and design tasks in a simple, analytical way. The process of gamification, to which MCTS is often associated, is then described, so that structural design is reduced to the concepts of state, actions and payoff.. Finally, case studies are presented in which different agents—based respectively on GA, MCTS, and a combination of both—are tested. The results suggest that hybrid approaches, where GA operates first followed by MCTS, are the most effective.

The expansion of rail transport infrastructures necessitates accurate and efficient soil surveys to ensure long-term stability and performance, particularly in regions prone to soil heaving. This study aimed to demonstrate the potential of non-destructive spectral analysis combined with Agentic Artificial Intelligence for automating the identification of soil heaving potential, providing a transformative approach to soil assessment in railway construction. A robust AI-agent was developed to predict soil heaving potential across temperature regimes (ranging from 0°C to -5°C and back), enabling characterization of the relative acoustic compressibility coefficient (β) based on the physical and mechanical properties of the soil. The main objective was to develop a framework that integrated spectral reflectance data with machine learning algorithms to predict soil heaving potential and reduce the reliance on traditional invasive methods. The experimental setup employed digital techniques to process and record longitudinal and transverse acoustic pulse signals reflected from piezoelectric sensors mounted on soil specimens. The processed signals were automatically transferred via a USB adapter to a PC for further analysis by the AI-agent. Acoustic diagnostics of the soils were performed using Fast-Fourier Transform (FFT) Spectral Analysis, followed by correlation of waveform spectra with heaving deformation. The AI-agent utilized a hybrid architecture combining Convolutional Neural Network (CNN), Support Vector Machine (SVM), and Random Forest (RF) algorithms to address the complexities of heterogeneous soil data and multifaceted prediction tasks—including heaving classification and deformation regression—while mitigating overfitting. Soil heaving potential was accurately predicted by the AI agent, with minor variations attributed to equipment sensitivity.