PhD Student University of California, Los Angeles
Fine-Scale Assessment of Landslide Risks in California
Co-Authors: Bozhou Zhuang (University of California Los Angeles), Debasish Jana (Colorado State University), Arjun Nair (University of California, Los Angeles), Tate Hom (University of California Los Angeles), Sriram Narasimhan (University of California Los Angeles), and Ertugrul Taciroglu (University of California Los Angeles)
Abstract: Landslides in California have become increasingly frequent, with reported events rising from an average of 49 per year during 2017–2022 to over 1,000 events in 2023. This increase is linked to frequent extreme rainfall associated with climate-change-driven wet–dry cycles, while recent wildfires further exacerbate landslide susceptibility by altering soil structure and vegetation cover. To investigate relationships among landslides, rainfall, geology, and transportation impacts, this study develops a framework for evaluating rainfall-induced landslide risk and transportation network vulnerability across California.
The framework integrates high-resolution (10 m × 10 m) datasets describing topography, geology, precipitation, and vegetation dynamics. California is discretized into approximately 4.8B grid cells with multi-year daily precipitation and Normalized Difference Vegetation Index (NDVI) time series. Data curation, spatial alignment, and time-series processing are performed through DesignSafe–TACC High Performance Computing resources. A 15-day rolling average is applied to NDVI to reduce daily artificial variability from atmospheric effects and viewing geometry, while 15-day accumulated precipitation represents short-term rainfall conditions relevant to landslide initiation. To capture rainfall variability beyond the historical record, 4,000 years of synthetic daily precipitation are generated using a generative AI model. Curated datasets are stored in the DesignSafe data repository and are planned to be staged and archived through the DesignSafe Data Depot. Ongoing work employs agentic large language models to extract geolocated landslide information from news and public reports. Vision–language models interpret landslide probability patterns from rainfall, NDVI, slope, and geology heatmaps. As transportation network analysis expands beyond the hillside area, GPU resources on DesignSafe will support scalable network-level hazard assessment.
Assistant Professor University of Kentucky
A Physics-Informed DeepONet Framework for Rainfall Infiltration Forecasting in Unsaturated Slopes
Co-Author: Shian Cao (University of Kentucky)
Abstract: Forecasting rainfall-induced landslides constitutes a critical challenge in geological hazard assessment. The forecasting of such events is complicated by the highly nonlinear Richards equation, which governs rainfall infiltration in unsaturated slopes. Although existing theoretical models describe these processes, the substantial nonlinearity of hydraulic conductivity makes calculation computationally expensive and its determination, ranging from on-site sampling to laboratory testing, is a labor-intensive and time-consuming process that hinders rapid landslide hazard forecasting. This study proposes an integrated forward-inverse physics-informed DeepONet framework for pressure head forecasting and hydraulic conductivity identification. The proposed framework inversely estimates the hydraulic conductivity and is validated through an unsaturated soil slope case study. Results demonstrate that the proposed framework is capable of predicting both pressure head evolution and hydraulic conductivity with sparse observation data. Compared to inverse physics-informed neural networks (PINNs), the proposed framework exhibits superior robustness against data bias due to sensor perturbation and achieves higher accuracy in inverse identification. Crucially, the trained model can instantly predict pressure head evolution under varying stochastic rainfall scenarios without the need for retraining. This capability significantly reduces the computational cost required for rapid landslide forecasting, broadening the potential applications of physics-informed machine learning in geological hazards.
PhD Student UC Berkeley
Regional Simulations of Rainfall-Induced Landslides: The example of Hurricane Maria in Puerto Rico
Co-Author: Dimitrios Zekkos (UC Berkeley)
Abstract: Rainfall-induced landslides remain a significant global hazard. Although numerous predictive models have been developed, most struggle to balance mechanistic rigor, accuracy, and computational efficiency at regional scales. Even more rigorous mechanistic models are often limited by high computational costs, sensitivity to input quality, and poor scalability. Our research leverages advances in monitoring, sensing technologies, big data analytics, and high-performance computing through platforms like DesignSafe to enable scalable and robust hazard predictions. We present a new model for Coupled Regional Rainfall-Induced and Seismic Slope Instability Simulations (CRISIS). The model integrates a pseudo-three-dimensional slope stability approach with the Parallel Integrated Hydrologic Model (ParFlow) and a seismic triggering model. The framework operates in two modes: back-analysis of mapped landslides and forward predictive modeling. Both failed and stable slopes are back-analyzed to generate high-resolution geospatial maps of shear strength variability. Unlike models that assume uniform strengths, this approach captures spatial heterogeneity and iteratively refines after each event, enhancing predictive accuracy. The derived strengths, along with other properties, are then used in forward simulations to predict the location, size, depth, and timing of landslides triggered by future storms or earthquakes. We demonstrate the model using the 2017 Hurricane Maria event in a 3.25 km2 watershed in Utuado, Puerto Rico. The model accurately reproduced slope failures observed in the area during Hurricane Maria when geospatially back-calculated strengths were applied. Moreover, it was able to capture varying failure mechanisms across the watershed, reflecting the influence of topographic, hydraulic, and hydrological factors.
PhD Student University of Arkansas
A Computational Framework for Separating Shallow and Deep Resonance in HVSR Analysis
Co-Authors: Clinton M. Wood (University of Arkansas), Mersad Fathizadeh (University of Arkansas), and Mohammadyar Rahimi (Keller North America)
Abstract: Horizontal-to-Vertical Spectral Ratio (HVSR) measurements in deep sedimentary basins frequently exhibit multiple peaks. While the low-frequency peak (f₀) typically originates from the deep soil-bedrock interface, higher-frequency peaks (f₁) may arise from shallow impedance contrasts within the sediment column. However, determining which layer interface in the Vs profile is associated with which peak is not always straightforward. This study presents hvstrip-progressive, an open-source Python package for forward HVSR modeling using the diffuse-wavefield assumption and systematic progressive layer stripping. The progressive layer stripping methodology iteratively removes the deepest finite layer from a velocity model, promotes the next layer to the half-space, and tracks how the fundamental frequency and amplitude evolve with each step. The package features adaptive frequency scanning to capture resonance peaks across diverse conditions, rigorous model validation ensuring physical consistency, and comprehensive reporting capabilities. The methodology is demonstrated using synthetic multi-layered soil profiles representative of sedimentary basins, showing how the technique attributes observed HVSR peaks to specific subsurface interfaces. The transparent workflow, reproducible outputs, and publication-quality visualizations make hvstrip-progressive a practical computational tool for seismic site characterization, providing quantitative constraints for interpreting multi-peak HVSR curves that contribute to site response analyses.
Professor University of California, Davis
Seismic Shielding for Soil-Structure Systems
Abstract: Earthquake mechanical waves carry seismic energy and excite soil-structure systems (buildings, bridges, tunnels, dams, power plants, etc.). The Earthquake-Soil-Structure-Interaction (ESSI), the propagation of seismic waves, seismic energy in time and space, through soil-structure system, determines the extent of damage. Controlling, directing propagation of seismic energy through the soil-structure system can be used to improve safety and economy of infrastructure objects. If seismic energy can be deflected from and/or dissipated outside of structure or dissipated within structures using designated dissipation devices, earthquake damage can be reduced and even completely alleviated.
Presented will be analysis methodology, modeling and simulation tools, new understanding of seismic energy propagation and practical design recommendations to control and direct propagation of seismic energy within soil-structure systems. Proposed methodology to control and direct propagation of seismic energy encompasses use of:
- Inelastic soil adjacent to, beneath the structure, and the soil-foundation interface zone,
- Energy dissipators, energy sinks, within structure,
- Viscous dampers and viscous coupling between fluid and solid/structure,
- External trenches surrounding the structure,
- Meta-materials/meta-devices, resonant unit cells, negative stiffness meta-materials...
High fidelity models of soil-structure building systems, ASCE-7-22 standard buildings are used to investigate and asses seismic energy control approaches, as noted above. Developed methodology and tools are used to improve safety and economy of new objects, as well as improving safety and economy of existing objects through upgrades.
PhD Student University of Washington
Fast Efficient Metamodeling for OpenSees-based Resilience Analysis (Femora)
Co-Author: Pedro Arduino (University of Washington)
Abstract: High-fidelity assessment of seismic infrastructure performance requires a rigorous understanding of the interaction between regional wave propagation, local site effects, and structural response. Traditional analysis methods often decouple these phenomena, failing to capture the intricate physics of soil-structure interaction (SSI), particularly for deep foundations and nonlinear soil behavior.
This talk introduces Femora (Fast Efficient Metamodeling for OpenSees-based Resilience Analysis), a high-performance Python-based library designed to bridge the gap between regional-scale seismology and local-scale structural engineering. Femora provides a holistic computational framework by automating the generation of analysis-ready finite element models for OpenSees. It integrates state-of-the-art numerical features, including the Domain Reduction Method (DRM) for physically consistent seismic input from 3D wavefields, Perfectly Matched Layers (PML) for robust boundary absorption without spurious reflections, and mortar-based embedded element formulations for efficient soil-pile coupling without the constraints of conformal meshing.
To make these advanced tools accessible and scalable, Femora has been integrated into the NHERI SimCenter’s EE-UQ application. This integration enables the research community to leverage high-performance computing (HPC) resources for rigorous uncertainty quantification, reliability analysis, and large-scale parallel simulations of complex soil-foundation-structure systems. Femora thus offers a scalable solution for next-generation performance-based earthquake engineering, moving from individual structural models to high-fidelity, site-specific regional assessments.