Poster Presentations:
Graduate Student Researcher
Texas A&M University
Machine Learning for Improved Landslide Susceptibility Analysis with Limited Data
Co-Authors: Stephanie Paal (Texas A&M University)
Abstract: Landslides have always posed a serious threat to the communities that they impact, they directly threaten human life and infrastructure with their destructive power. Furthermore, landslides can indirectly cause serious harm through the destruction of vital infrastructure in isolated and oftentimes economically disadvantaged geographies. Landslide susceptibility analysis is conducted to inform decision makers and resident alike about the dangers they face and allow them to make informed choices about their safety. Recently machine learning (ML) has been investigated as a method for solving a multitude of civil engineering challenges. ML algorithms can analyze vast datasets, including topographic, geological, hydrological, and meteorological factors, to identify patterns and correlations that may not be immediately apparent to human experts.
As ML models require substantial amounts of relevant data to train an accurate model, communities in locations that lack this historical data are at a disadvantage when it comes to producing and utilizing landslide susceptibility predictions. To alleviate this issue transfer learning is proposed as a method to produce more accurate landslide susceptibility results without necessitating the additional collection of data in the desired geographical prediction region. Landslide data from more widely studied and documented areas can be leveraged by TL techniques and combined with the limited data from the desired area of implementation to provide a more accurate model. By leveraging existing data and advances in ML and TL communities in vulnerable areas, often with limited resources, can have access to better predictions that will allow them to plan accordingly.
Graduate Student Researcher
University of Southern California
Understanding HVSR Curve Characteristics for Site Characterization using AutoEncoders
Co-Authors: Chukwuebuka Nweke (University of Southern California)
Abstract: Horizontal-to-Vertical Spectral Ratio (HVSR) curves are important products for seismic site characterization and microzonation studies, as these curves contain explicit or implicit representations of three-dimensional site characteristics such as resonant frequencies and other subsurface features. Current HVSR approaches are limited by their reliance on simplistic interpretations of spectral peaks and troughs, which often fail to capture the full complexity of ground motion augmentation due to subsurface conditions. The application of machine learning (ML) techniques like deep learning and autoencoder architecture offer a potential solution for interpreting HVSR curves by learning data representations and identifying underlying patterns without explicit feature engineering. Autoencoders consist of an encoder, which compresses the input data into a lower-dimensional latent space, capturing essential features, and a decoder, which reconstructs the original data from this compressed representation. We apply autoencoders to a dataset of microtremor HVSR curves derived from ambient noise collected in the Salt Lake City region of Utah over a 30-day period. The resulting latent space representation of the HVSR features is assessed for association with observable HVSR ordinates, conventional site parameters, and spatial-temporal variables, where the associations are used to categorize/classify HVSR curves for use in site response modeling. The use of autoencoders for HVSR curve interpretation presents a novel approach that combines the strengths of machine learning with geophysical expertise. This method enhances the efficiency and accuracy of seismic site characterization, providing a powerful tool for geophysicists and engineers.
Graduate Student Researcher
University of Washington
High-Fidelity Dynamic Analysis of Pile Foundations: A Step-by-Step Procedure with Emphasis on Realistic Modeling and High-Performance Computing
Co-Authors: Pedro Arduino (University of Washington)
Abstract: In geotechnical engineering, precise analysis of dynamic pile foundations is crucial. This study presents a comprehensive procedure for high-fidelity finite element simulations of pile foundations subjected to seismic events. The approach emphasizes realistic modeling and utilizes high-performance computing for efficiency.
The procedure commences by simulating the source rupture, a pivotal step in seismic event replication. Geophysics software, employing methods like green functions, finite elements, or finite difference techniques (e.g., shaker maker, Hercules, or SW4), facilitates the acquisition of site-specific ground motion data.
To tackle complexities in near-fault areas with distinct ground motions, the Domain Reduction method (DRM) is employed with high-performance computing. This enables regional-scale earthquake analysis simulations using truncated domains. The seismic source is strategically brought closer to the domain, curbing computational costs. Incorporating Perfectly Matched Layer (PML) elements boosts efficiency by absorbing outgoing waves, eliminating the need for domain extensions.
This research focuses on high-fidelity finite element simulations, encompassing 3D foundation analysis and nonlinear soil behavior including liquefaction. This is accomplished using advanced constitutive models, coupled FEM formulations, and a specialized embedded interface element representing pile behavior. The entire workflow is implemented in EEUQ, one of the SimCenter tools, and uses DesignSafe HPC resources to perform massive parallel simulations.
In summary, this study seeks to improve dynamic pile foundation analysis through a procedure that combines realistic modeling, advanced constitutive models, and high-performance computing. Innovative techniques like DRM, PML, and embedded interface elements enhance accuracy and efficiency, advancing our grasp of pile foundation seismic response.