Deep Learning for Earthquake Engineering: Temporal Convolutional Networks for Structural Response Prediction and Ground Motion Reconstruction
Presented By: Kwok Pang
Affiliation:
Description: Temporal Convolutional Networks (TCNs) have shown significant potential in seismic applications, particularly for predicting structural responses and reconstructing ground motion histories. Initial studies highlight their ability to accurately forecast linear structural responses and infer ground motion from response data, demonstrating their practical utility. However, modeling complex nonlinear behaviors, such as damping variations and period elongation, exposes limitations in the TCN framework. To address these challenges, we propose an enhanced TCN that incorporates residual blocks to improve feature representation and combines time-domain and frequency-domain loss functions to capture both temporal and spectral characteristics of structural dynamics. Additionally, the model incorporates an Exponentially Weighted Moving Average feature to emphasize recent data while accounting for historical trends. Validation experiments reveal that the enhanced TCN delivers a better accuracy in predicting nonlinear dynamics. This study underscores the adaptability and scalability of TCNs, establishing a robust foundation for future advancements in seismic response prediction, including the integration of physics-informed neural networks
Physics Informed Machine Learning for Reduced Data Modeling of Novel Reinforced Concrete Wall Behavior
Presented By: Stephanie Paal
Affiliation: Texas A&M University
Description: Concrete walls play a critical role in high-rise construction, primarily by resisting shear forces and enhancing the overall rigidity of structural systems. As the demands of the construction industry evolve, these walls are continuously adapted to meet both architectural constraints and structural performance requirements. One approach to improving their efficiency involves exploring novel wall geometries that optimize performance under specific loading conditions. However, fully characterizing the behavior of new wall configurations and updating design standards requires significant time and computational resources. To address this challenge, methods that reduce the data required for accurate evaluation are highly beneficial.
One such approach is Physics-Informed Machine Learning (PIML), a technique that integrates established physical principles into the machine learning (ML) process, enhancing model accuracy and interpretability with limited data. In this study, physics-based knowledge derived from ACI 318-19 is embedded into the learning process of two distinct machine learning algorithms, genetic expression programming and random forest. The effectiveness of physics injection is assessed by comparing model performance with and without physics knowledge injection. Given the limited dataset available for C-shaped and G-shaped concrete walls, this analysis evaluates the impact of incorporating domain knowledge into ML models to improve predictive reliability and generalization for these shapes. This demonstration using existing wall shapes serves as a case study, illustrating the potential for future applications to novel wall configurations.
Effect of Dataset Representation Bias on Generalizability of Machine Learning Models in Predicting Flexural Properties of Ultra-High-Performance Concrete (UHPC) Beams
Presented By: Jinxin Chen
Affiliation: Stevens Institute of Technology
Description: Machine learning (ML) offers transformative potential in structural design through the high efficiency in exploring optimal solutions within vast design spaces. However, concerns persist among structural engineers regarding the generalizability and reliability of ML models trained on datasets with biases. This study investigates the influence of dataset representation bias on the generalizability of ML models in predicting the flexural properties of ultra-high-performance concrete beams. The research addresses three primary objectives: (1) developing a novel metric to quantify representation bias of continuous datasets, (2) devising an algorithm to create datasets with controlled bias levels, and (3) evaluating the effect of dataset representation bias on the generalizability of ML models. Key contributions include the first comprehensive analysis of representation bias in structural prediction models, the introduction of a Monte Carlo Bias Estimation method for evaluating dataset bias, the development of an Adaptive Bias Sampling Algorithm for dataset generation, and the modification of Latin Hypercube Sampling to ensure uniform dataset distribution. Findings reveal that dataset bias significantly undermines the generalizability of ML models, and the proposed methods offer effective strategies for assessing and mitigating dataset bias, thereby enhancing the generalizability of ML models.
ANNs Computer Vision Damage Methodologies Applied to Crack Evolution in Self-Healing Concretes
Presented By: Liberato Ferrara
Affiliation: Politecnico di Milano
Description: Modernizing the concrete industry requires the use of advanced technologies and digitalization in order to improve the durability and sustainability of our structures. Two key ideas in this direction are 1) the development of new high-performance concretes (UHDC) which undergo "selfhealing" and 2) autonomous systems for monitoring existing structures, based on computer vision. These two technologies, in fact, can potentially lengthen the service life of structures: self-healing offers better performing materials that have greater durability, while, thanks to computer vision, large-scale monitoring by using drones or other automated systems, can be possible. The goal in this paper is therefore to investigate these two issues and explore the potential of combining self-healing with computer vision and machine learning. All the research done aims to contribute to the construction one day of an automated system that starts from images and estimates structural safety. First of all, experimental data from previous research conducted at Politecnico di Milano on self-healing are collected and analyzed. The results of these analyses are used to derive qualitative empirical laws: they relate the visual crack closure in percentages with an estimation of the recovery of mechanical properties and permeability. The results demonstrate that analyzing the surface is meaningful also in terms of material characteristics. In the second part, the research focus on image processing. This process normally requires three steps: stitching of the microscope photos, recognition of the same crack section at different moments in time, and binarization of the image (to extract the open area). The commercial off-the-shelf softwares do not present problems stitching photos, while they fail to correctly align images taken at different times and require heavy human intervention for the segmentation of the crack.
From Chemistry to Code: A Data-Driven Reassessment of Available Alkali Tests for SCMs
Presented By: Pravin Saraswatula
Affiliation: Texas A&M Transportation
Description: Traditionally, ASTM C 311-based available alkali (AA) tests have been employed to assess supplementary cementitious materials (SCMs), especially fly ash, for their effectiveness in mitigating alkali-silica reaction (ASR), applying a 1.5% Na2Oeq threshold. However, recent developments, including the inconsistent correlation of test results with ASR expansion, high variability across laboratories, and the introduction of new & alternative SCMs (i.e., ASCMs such as natural pozzolans with higher alkali contents), have raised concerns about the test’s reliability and applicability. To address these challenges, recent research has integrated advanced data analytics with thermodynamic modeling to examine the role of ASCM composition, mineralogy, and reactivity in AA contribution to pore solutions. Bayesian machine learning models have further been developed to estimate the soluble alkali dissolution of ASCMs based on their bulk chemical composition, drawing from a comprehensive dataset of experimental data on mineralogy, degree of reaction, and other relevant factors. Although machine learning models primarily provide correlation insights rather than causation, these tools significantly enhance our understanding of alkali dissolution dynamics in ASCMs. The findings underscore the potential for AA testing to remain relevant in evaluating ASCMs reactivity while also highlighting the promising application of machine learning. As part of an evolving pore solution model, the predictive capabilities of these ML models offer direct estimates of soluble alkali contributions from ASCM compositions, establishing a data-driven approach that enhances the accuracy and reliability of ASR mitigation strategies.
Computer Vision for Accelerated and Automated Sorptivity Testing
Presented By: Hossein Kabir
Affiliation: University of Illinois Urbana-Champaign
Description: Reliable assessment of sorptivity—the rate at which water infiltrates unsaturated cementitious materials—is essential for predicting the long-term durability of structural systems. Conventional testing protocols such as ASTM C1585 are labor-intensive, rely on manual measurements, and require extended durations, limiting their practicality for quality control during construction. To address these challenges, we present two novel computer vision–based methods enhanced with artificial intelligence to automate and accelerate sorptivity evaluation. The first method, “droplet-based,” extracts contact angle dynamics of water droplets on cement paste surfaces to predict sorptivity from wetting behavior. The second “waterfront-tracking” method leverages an algorithm to monitor the visible wetting front and derive initial and secondary sorptivity values from image-derived wetted area ratios. Both approaches are trained on large image datasets and demonstrate high predictive accuracy (R² > 0.9), aligning well with conventional durability metrics such as freeze-thaw performance and surface resistivity indices. These AI-integrated methods offer accelerated, automated, and non-invasive tools for durability assessment and structural quality assurance.