Poster Presentations:
Undergraduate Student Researcher
University of Texas at Austin
AI-Driven Digital Twins for Multi-Phase Debris Flow Simulation in Natural Disasters
Co-Authors: Justin Bonus (UC Berkeley)
Abstract: Natural disasters such as tsunamis, landslides, and storm surges mobilize debris, exacerbating hazards to built environments. Predicting debris dynamics in these complex flow events is challenging due to the heterogeneous nature of materials and uncertainties in their physical properties. This study introduces an innovative approach to address these challenges by integrating artificial intelligence, machine learning (ML), and high-performance numerical simulations. Our goal is to enhance the understanding of these hazards through a digital twin surrogate modeling workflow, using experimental facilities as a foundation to expand on practicality (i.e., available time, funding, etc.). We introduce a prototype digital twin of the Hinsdale Wave Research Facility's Large Wave Flume at Oregon State University (OSU LWF), developed with the Taichi Material Point Method (MPM), though not limited to this simulation framework, to study wave-debris dynamics. Our high-performance digital twin workflow consists of three distinct phases. The first phase generates a structural point cloud using OpenAI's Point-E model or derived from lidar scans, which is then integrated into the OSU LWF simulation. The second phase employs the Taichi MPM for classical numerical simulations of large-deformation, multi-material dynamics. Finally, the third phase outputs lightweight surrogate models from the MPM simulations using the Graph Network Simulator (GNS), a versatile graph neural network (GNN) software package. This approach allows us to extend our workflow to various experimental facilities and potentially beyond wave-debris studies, leveraging high-performance computing (HPC) and GNNs to efficiently represent complex physics and facilitate uncertainty quantification in debris hazard events.
Graduate Student Researcher
Oregon State University
Uncertainty Quantification in Seismic-induced Building Debris Tsunami Evacuation
Co-Authors: Mehrshad Amini (University of Rhode Island), Andre R. Barbosa (Oregon State University), and Daniel Cox (Oregon State University)
Abstract: This study aims to quantify uncertainty in predicting seismic damage to buildings and associated debris at a community scale. Traditionally, the HAZUS methodology offers expected values for the debris fraction, resulting in deterministic outcomes for structural and nonstructural debris volumes. In this study, we extend the HAZUS methodology to quantity and propagate uncertainties in the model. Results are compared using two building inventory databases: (1) the National Structure Inventory (NSI), and (2) a local tax assessor inventory. The NSI is a publicly available building inventory developed by the U.S. Army Corps of Engineers, whereas local tax assessor data are unique for each community. This study is applied to the coastal city of Seaside, Oregon, which is vulnerable to earthquake and tsunami hazards from the Cascadia Subduction Zone. The seismic damage and associated debris volume are estimated at the parcel level across seven (7) mean recurrence intervals, ranging from 100-yr to 10,000-yr. Additionally, we compare the outcomes of this new approach with those of the traditional HAZUS methodology. The impact of the damage on debris volume is assessed on the tsunami life-safety, using IN-CORE, which is an open-source community resilience modeling environment. The insights gained will allow decision-makers, engineers, and scientists to better understand the uncertainty in seismic debris of buildings, which can improve the effectiveness of seismic and tsunamis risk management and response strategies.
Graduate Student Researcher
University of Southern California
CelerisAi: A Python-Taichi-Based Nearshore Wave Modeling Framework for Integrated AI Applications
Co-Authors: Patrick Lynett (University of Southern California)
Abstract: This presentation introduces PyCeleris, a novel nearshore wave modeling software developed in Python-Taichi. Building upon the Celeris Advent framework, CelerisAi offers a high-performance, real-time solution capable of simulating wave dynamics in coastal environments. By leveraging Taichi's high-performance parallel programming capabilities(CPU/GPU), CelerisAi enables efficient computations on personal laptops, and it is also suitable for large-scale simulations on HPCs.
A key advantage of CelerisAi is its seamless integration with machine learning and artificial intelligence environments. This integration allows for hybrid models combining numerical simulations with data-driven approaches. As a proof of concept, we demonstrate how CelerisAi can be used to train a neural network to predict wave effects based on boundary and initial conditions. This approach could reduce the memory footprint required for storing simulation results while simultaneously training a neural network on a large dataset of scenarios.
Beyond wave prediction, CelerisAi can serve as a foundation for addressing various coastal engineering problems, including inverse problems, sediment transport, and wave run-up. The framework's flexibility allows for exploring different neural network architectures, such as reinforcement learning for incorporating agent-based models into coastal infrastructure design or variational autoencoders for identifying patterns in coastal processes.
By combining numerical modeling with artificial intelligence, CelerisAi offers a promising approach to advancing the understanding of coastal dynamics and hazards.