As infrastructure ages and demands on our built environment increase, efficient and accurate assessment of concrete structures is more critical than ever. This session will explore how artificial intelligence (AI) and machine learning (ML) are transforming the way engineers and inspectors perform reconnaissance and inspection of concrete infrastructure. Attendees will learn how advanced data analytics, computer vision, and predictive modeling are being applied to automate and enhance traditional inspection methods. From drone-based imaging and crack detection to condition assessment and deterioration forecasting, AI and ML are opening up new possibilities for faster, safer, and more consistent evaluations.
Learning Objectives:
(1) Understand the fundamentals of AI and machine learning and how they apply specifically to the inspection and assessment of concrete structures;
(2) Identify current technologies and tools - including drones, sensors, and computer vision systems - used in AI-driven reconnaissance of concrete infrastructure;
(3) Evaluate real-world case studies that demonstrate the effectiveness of AI and ML in detecting defects, predicting deterioration, and improving inspection efficiency;
(4) Explore the benefits and limitations of integrating AI and ML into traditional inspection workflows, including considerations for data quality, validation, and implementation in field settings.
Simulated Ground Motions as Synthetic Data for Structural Health Monitoring and Rapid Assessment
Presented By: Claudio Perez
Affiliation: University of Miami
Description: The scarcity of strong motion records in seismically active regions presents a significant challenge for developing and validating structural health monitoring procedures. This paper explores the use of simulated ground motions as synthetic input data for evaluating and training both physics-based and data-driven models used in post-earthquake assessment. Ground motions are drawn from the PEER-LBNL Simulated Ground Motion Database (SGMD), which provides spatially dense, validated time histories generated via high-performance physics-based simulations of large-magnitude earthquake scenarios. We discuss integration of SGMD with the BRACE2 web platform, enabling virtual deployment of predictors under hypothetical ground motions at arbitrary bridge sites. Results demonstrate how simulated datasets can expose vulnerabilities and failure modes in predictive models that may remain undetected under limited empirical data. This work supports the case for incorporating synthetic data pipelines into structural health monitoring frameworks and outlines key considerations for doing so effectively.
Enhancing Bond Strength Prediction at UHPC-NC Interface: A Data-driven Approach with Augmentation and Explainability
Presented By: Nima Khodadadi
Affiliation: University of California, Berkeley
Description: Existing concrete structures often struggle to reach their intended service life due to aging, increased loading demands, and natural disasters, necessitating repair, strengthening, or replacement. Ultra-high-performance concrete (UHPC), known for its exceptional strength and toughness, presents a promising solution for repairing and strengthening normal concrete (NC) structures. When applied in this context, UHPC is expected to be a viable and effective material. A reliable interfacial bond between the two materials (the NC substrate and UHPC overlay) is critical to the overall performance of the composite structure. In this study, a data-driven approach is proposed to predict the bond performance at the UHPC–NC interface with high accuracy. To address the challenge of limited experimental data, a data augmentation model is introduced, combining Kernel Density Estimation (KDE) and Tabular Generative Adversarial Networks (TGAN). The optimal model is selected from six decision tree-based ensemble learning methods, evaluated under two strategies: “synthetic training – real testing” and “real training – real testing.” Finally, a parametric analysis is conducted using SHapley Additive exPlanations (SHAP) to assess the importance and sensitivity of different features related to bond strength. The results demonstrate that the proposed KDE–TGAN model effectively captures the distribution of the original dataset, enhancing both the accuracy and robustness of the bond strength prediction models. Moreover, the model’s ability to interpret feature importance and sensitivity provides valuable insights for bond strength prediction. Therefore, the proposed data augmentation approach offers a reliable framework for modeling experimental data with limited samples in structural engineering applications.
Automatic Interpretation of Concrete Design Codes using a Domain-specific Large Language Model (LLM)
Presented By: Jinxin Chen
Affiliation: Stevens Institute of Technology
Description: Large language models (LLMs) hold promises for automating information-intensive tasks, but their applications in safety-critical design and construction domains remains limited. This research developed a domain-specific framework designed to deliver a ready-to-use application for automatic interpretation of concrete design codes, combining LLMs with retrieval-augmented generation (RAG) tailored to specific codes. The framework has three primary innovations: (1) an advanced retrieval mechanism combining context-aware hierarchical search and fact-grounding to minimize hallucinations, critical for safety-critical applications; (2) training-free RAG architecture enabling convenient deployment without costly training or fine-tuning; and (3) human-computer interactive interface where engineers dynamically adjust parameters (chunk size, search scope, retrieval breadth) to align outputs with specific requirements. Evaluations on 1,000 code queries demonstrate > 95% accuracy, outperforming error-prone general-purpose models (< 2% accuracy). The success of the system in automating compliance verification and structural design tasks establishes it as a deployable tool for intelligent assistance in safety-critical concrete structure design applications.
AI-Enhanced Models for Estimating Concrete Compressive Strength Using NDT Data
Presented By: Farid Moradi Marani
Affiliation: FPrimeC Solutions
Description: Core sampling for existing structures is a common practice to estimate the compressive strength of concrete in reinforced concrete structures. However, it can be impractical for many operational existing concrete structures (such as concrete structures in water treatment facilities) to extract core sample for the evaluation of strength. Additionally, core samples may not be a good representative of all areas under investigation, particularly in mass or large-scale concrete structures. Achieving a reliable estimation of compressive strength in these structures may require us to extract numerous concrete core samples. Non-destructive Testing (NDT) methods may offer a solution for evaluating the strength, because they are non-destructive, repeatable, and capable of covering a larger surface area without necessarily implying additional costs and time to the project. SonReb, a combined NDT method using Ultrasonic Pulse Velocity (UPV) and Rebound Hammer data, has been extensively used to estimate compressive strength. However, SonReb’s supporting empirical models have been essentially developed base on local data, which may not be applicable or reliable for general use by other engineers. This study presents a practical machine learning (ML) model for on-site concrete strength prediction. The idea was to train a comprehensive and global tool for engineers to estimate compressive strength. A large database was created from different projects along with new experimental test data in lab. UPV and Rebound Hammer tests were carried out on these samples along with compressive strength tests. The data was used to train three different ML models, then accuracy of each model was subsequently validated against data obtained from other existing structures. The results show that the proposed ML model provided the most reliable predictions of concrete strength with a mean absolute error of less than 10 %. Eventually, the R&D project findings suggest that proposed ML can be potentially
Smart Composite Reinforced Concrete System with Digital Twin-Based Monitoring and AI-Driven Condition Assessment
Presented By: Chengcheng Tao
Affiliation: Purdue University
Description: This study aims to develop a smart composite reinforced concrete system for real-time health condition monitoring using embedded sensors on the composite. The monitoring system can provide the health condition and risk information of the composite reinforcement. Data from smart sensors is paired with computational modeling results of composite-concrete system and AI-driven machine learning algorithms to assess and predict the condition of the composite reinforcement. The system detects the structural and materials failure and anomaly mechanism and predicts the associated risk in a wide range of engineering applications.
From Chaos to Insight: A Multimodal Framework for Post-Earthquake Reconnaissance Using Online Data
Presented By: Khalid Mosalam
Affiliation: University of California, Berkeley
Description: In the aftermath of major earthquakes, conducting reconnaissance efforts rapidly is essential for understanding the performance of the built environment and supporting informed decision-making. Meanwhile, virtual reconnaissance based on online data is increasingly gaining attention due to its feature of operating without on-site investigation. This study proposes a multimodal framework to leverage open-source online data to automatically generate structured post-earthquake reconnaissance briefings. The system design is informed by the existing virtual reconnaissance workflow of the Structural Extreme Events Reconnaissance (StEER) network, with key steps identified and automated. The framework integrates multiple APIs and AI components to monitor seismic events globally, scrape and process fragmented, noisy multimodal data from online sources, and extract and summarize valuable information. This work demonstrates the feasibility of using open-source online data and AI for automated virtual reconnaissance and highlights key challenges in developing a domain-specific, AI-assisted reporting framework.