NHERI Computational Symposium

February 5-7, 2025

Machine Learning Applications in Earthquake Engineering

Poster Presentations


Omar Abukassab

Graduate Student Researcher
University of Texas at Austin

Regional Seismic Landslide Prediction using XGBoost

Co-Authors: Ellen Rathje (The University of Texas at Austin) and Krishna Kumar (The University of Texas at Austin)

Abstract: Regional seismic landslide prediction models serve the purpose of spatial predictions of future landslides triggered by earthquake shaking as a function of topographic characteristics, ground shaking intensities, and geologic conditions. In our study, we develop a regional seismic landslide prediction model using the eXtreme Gradient Boosting (XGBoost) machine learning algorithm, based on the landslide inventory from the 1994 Northridge earthquake. The binary prediction model utilizes several predictor variables that incorporate the influence of topographic parameters (topographic slope, surface roughness), ground motion parameters (PGV), landcover features (tree cover), and geologic properties (soil type and soil thickness). An important feature of our model building was the use of Poisson-disk sampling to randomly select training cells while maintaining a minimum radius distance r between cells, which avoided spatial correlation in the training data. Evaluated on the independent testing data, the prediction model has a high true positive rate (TPR=94%) and a low false positive rate (FPR=19%). Comparisons with predictions from a physics-based regional prediction model and a global empirical prediction model reveal that our XGBoost model captures more positive cases while minimizing false predictions, and it remains resilient to the over-prediction issue that emerged in the other models. These findings serve as the stepping stone for further research about the use of machine learning algorithms to develop a generalizable global seismic landslide prediction model.

Prakash Gaire

Graduate Student Researcher
University of Utah

Effectiveness of Machine Learning Models in Predicting Engineering Demand Parameters Across Diverse Structural Datasets

Co-Authors: Luis Ibarra (University of Utah)

Abstract: This study examines the effectiveness of Machine learning (ML) models in predicting collapse status and maximum interstory drifts across a variety of datasets of steel moment resisting frames (SMRFs). Previous studies have utilized datasets of a single building or limited sets of SMRF buildings with very similar characteristics. Although some regional studies address the performance of large number of buildings, they often overlook the structural features directly controlling the building’s behavior, focusing mainly on dimensional parameters, age, and location, among other general features. To date, ML models based on structural features, which are expected to provide a better prediction, have not been tested across diverse databases to assess their efficiency when tested on systems that may depend on features not initially investigated.

This study developed ML models that incorporate uncertainties in nonlinear system and ground motion parameters. The models were initially trained using a dataset of regular buildings with different heights and same floor layout, which were designed according to modern US design codes. The trained ML models were then applied to different datasets of SMRFs, revealing reduced efficiency in prediction of engineering demand parameters. The use of ML models on different datasets showed that features missing in the trained dataset may lead to poor predictions in the new datasets. If these features lack significance in the trained dataset, the ML shortcomings can only be solved by improving the training database.

Jorge Macedo

Assistant Professor
Georgia Institute of Technology

Leveraging Machine Learning and HPC for assessing penetration data

Co-Authors: Luis Vergaray (Georgia Tech) and Srinivas Vivek (Georgia Tech)

Abstract: This presentation focuses on the inverse problem of assessing soil's state using penetration data (e.g., data from the penetration testing). Specifically, machine learning based procedures that leverage high performance computing simulations are presented for assessing soil's state given penetration information.

Haifeng Wang

Assistant Professor
Washington State University

Vision-Based Motion Measurement: A Gaussian Splatting Approach

Co-Authors: Hongyi Liu (Washington State University)

Abstract: Vision-based motion measurement offers significant benefits in terms of cost-effectiveness and efficiency compared to traditional contact-based sensors. However, its accuracy is generally limited to the pixel level, as pixels are the smallest units in an image. To achieve motion detection at the sub-pixel level, it is crucial to consider variations in pixel brightness. Existing brightness-based methods like Lucas-Kanade and phase-based techniques can attain sub-pixel accuracy but frequently encounter difficulties with parameter selection. This study introduces a Gaussian splatting approach designed to provide robust sub-pixel motion measurement, with parameters that are automatically adapted to the target image. Numerical verification confirms that this method delivers high accuracy and robustness.

Mohsen Zaker Esteghamati

Assistant Professor
Utah State University

How well can design-oriented machine learning models capture different performance-based earthquake engineering outputs?

Abstract: Recent advancements in machine learning (ML) have shown promise as efficient surrogates in performance-based earthquake engineering (PBEE), from dynamic analysis to fragility assessments and loss evaluations. Despite the potential, the literature presents diverse data, models, and features, making it challenging to assess ML models’ efficiency across PBEE stages. Additionally, the black-box nature of advanced ML algorithms limits their ability to provide design-oriented insights linked to building mechanics, hindering their adoption in PBEE design processes.

This study investigates “design-oriented” ML models that map design and geometry-related parameters to response, damage, and loss using a consistent database of finite element models. Unlike existing approaches, these models leverage only design parameters as inputs, allowing for direct design guidance. The study has three objectives: (1) compare ML models' accuracy across PBEE stages to assess sensitivity to training data and suitability for specific steps, (2) benchmark common ML algorithms across PBEE stages to identify optimal choices, and (3) apply explainable AI techniques to investigate the potential variation in model explanations across different outputs and algorithms.

The result shows that while ML models can generally map design parameters to various PBEE outputs, the highest accuracy was observed for drift responses, median fragilities, and component-based loss metrics. Moreover, the most accurate algorithm remained consistent across different PBEE outputs. Although the selected feature sets varied depending on the output and algorithm, critical features such as building height, number of stories, fundamental period, and the minimum moment of inertia of beams were found to significantly influence all outputs.

Sonia Zehsaz

Graduate Student Researcher
Oregon State University

Probabilistic Seismic Demand Models for Aging Bridges: A Machine Learning Perspective

Co-Authors: Sabarethinam Kameshwar (Louisiana State University) and Farahnaz Soleimani (Oregon State University)

Abstract: This study introduces contemporary Machine Learning (ML)-based Probabilistic Seismic Demand Models (PSDMs) aimed at assessing a crucial parameter in bridge engineering: the column drift ratio as an Engineering Demand Parameter (EDP). It also highlights the potentially significant effects of modeling-related parameters, such as aging and deterioration, on the seismic response and vulnerability estimates of deteriorating bridges. Unlike traditional methodologies relying on specific statistical techniques, and fixed functional forms, ML-based models present a more efficient, robust, and reliable alternative for quantifying EDPs and developing fragility models for bridges. These ML-based models often do not necessitate extensive computations and time-consuming processes. Moreover, traditional methodologies often exhibit a prevalent lack of comprehensive understanding regarding the influence of various modeling uncertainty parameters, particularly those related to aging factors, on the seismic response and seismic fragility estimates of bridges. The advantage of this approach lies in its capacity to capture the individual and cumulative effects of aging and deterioration without explicitly modeling the time parameter. Instead, the ML based PSDMs focus on the condition of columns and bearing assemblies, critical components susceptible to aging-related issues. The findings of our study suggest a heightened seismic demand response in corroded columns but minimal, relatively consistent changes in demands placed on specific elements like fixed and expansion bearings, when accounting for aging factors such as rebar degradation, bearing pad deterioration, and the weakening of bridge dowels. The results of this study can provide a significant step forward in enhancing performance-based seismic assessment of aging bridge systems.