💻 computer science

Geospatial AI for Liquefaction Hazard and Impact Forecasting: A Demonstrative Study in the U.S. Pacific Northwest

This study presents a high-resolution, machine learning-driven geospatial model that predicts soil liquefaction hazards across Washington and Oregon for 85 earthquake scenarios, demonstrating its utility for regional infrastructure vulnerability assessments and disaster planning while offering improved performance over prior approaches by integrating mechanical principles with extensive geospatial data.

Original authors: Morgan D. Sanger, Brett W. Maurer

Published 2026-03-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Morgan D. Sanger, Brett W. Maurer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Pacific Northwest (Washington and Oregon) as a giant, beautiful house sitting on a shaky foundation. We know that at some point, the ground will shake hard enough to cause trouble. One of the biggest worries isn't just the shaking itself, but liquefaction.

Think of liquefaction like a bowl of Jell-O. When you shake a bowl of Jell-O, the solid gelatin turns into a wobbly liquid. Similarly, when the ground shakes hard, wet, sandy soil can lose its strength and turn into a liquid soup. This causes buildings to sink, roads to crack, and bridges to collapse.

For a long time, predicting exactly where this Jell-O effect would happen was like trying to guess the weather in every single backyard without any thermometers. You had to dig deep holes (called Cone Penetration Tests, or CPTs) to measure the soil, but you couldn't dig holes everywhere. So, big maps were often just rough guesses.

The New "Crystal Ball" (Geospatial AI)

This paper introduces a new, high-tech "crystal ball" called Geospatial AI. Instead of needing a hole dug in every single backyard, this system uses a super-smart computer brain (Machine Learning) to look at the neighborhood from above.

Here is how it works, using a simple analogy:

The "Chef" and the "App": Imagine a master chef (the traditional geotechnical model) who knows exactly how to cook a perfect meal (predict soil behavior) but only has ingredients from one specific kitchen (the data from the few holes that were dug).
The "App" Learns: The new AI system is like a student who watches the master chef cook thousands of times. It learns the rules of cooking (the physics of soil) without needing to be in the kitchen.
The "App" Cooks Everywhere: Once the student learns the rules, it can go to any kitchen in the world (anywhere on a map) and predict what the meal would taste like, even if it has never been there before. It looks at the "ingredients" available from space—like how steep the hill is, how close the river is, and what the soil looks like from satellite images.
The "Local Taste-Test": If there is a hole dug nearby, the AI checks its prediction against the real data and adjusts its answer to be even more accurate.

What Did They Do?

The researchers took this "App" and ran it through 85 different earthquake scenarios for Washington and Oregon. They didn't wait for an earthquake to happen; they simulated them.

The Scenarios: They imagined everything from a medium-sized earthquake on a local fault (like a crack in the sidewalk) to a massive, catastrophic earthquake on the Cascadia Subduction Zone (the big one that could shake the whole region).
The Output: They produced detailed, color-coded maps. These maps show the "Probability of Ground Failure" (PGF).
- Green areas: "Safe, the soil is solid like concrete."
- Red areas: "Danger, the soil might turn into Jell-O."

Why Does This Matter? (The Real-World Impact)

These maps aren't just for scientists; they are tools for saving lives and money. Here is how different people can use them:

The City Planner: "Oh, I see that the new highway I'm planning goes right through a red zone. Let's move the road or reinforce the ground before we build."
The Emergency Manager: "If a tsunami comes, people need to run to high ground. But if the roads near the river turn to Jell-O, the roads will break. Let's plan evacuation routes that avoid the red zones."
The Insurance Company: "We need to know which neighborhoods are at highest risk to set fair prices and prepare for potential claims."
The Bridge Owner: "I have 500 bridges. Which ones are sitting on 'Jell-O' soil? I'll fix those first."

The Big Takeaway

The study found that while the massive "Cascadia" earthquake gets all the news attention, local earthquakes (like one hitting right under Seattle or Portland) might actually cause more liquefaction damage in those specific cities because the shaking is more intense right there.

However, the Cascadia earthquake is scary because it could turn the soil into Jell-O in hundreds of cities at the same time, affecting millions of people simultaneously.

In short: This paper gives us a high-resolution, super-smart map that tells us exactly where the ground might turn to liquid during an earthquake. It turns a scary, unknown risk into a manageable problem that cities can plan for, helping us build a safer, more resilient Pacific Northwest.

1. Problem Statement

Recent major earthquakes globally (e.g., Turkey 2023, New Zealand 2010–2011) have highlighted the catastrophic consequences of soil liquefaction, including loss of life, infrastructure collapse, and service disruption. The U.S. Pacific Northwest (Washington and Oregon) is uniquely vulnerable due to its potential for crustal, intraslab, and subduction zone earthquakes.

However, current methods for regional-scale liquefaction forecasting face significant limitations:

Data Dependency: State-of-practice geotechnical models rely on in-situ measurements (e.g., Cone Penetration Testing, CPT), which are sparse and cannot be deployed continuously over large regions.
Existing Geospatial Model Flaws: Current geospatial models often:
- Predict liquefaction manifestation directly without grounding in established mechanics.
- Exclude available subsurface geotechnical data.
- Utilize limited geospatial variables.
- Rely on traditional statistical methods (e.g., logistic regression) that fail to capture complex non-linear relationships.
- Lack packaged, user-friendly software for rapid deployment.

There is a critical need for a model that combines the physical rigor of geotechnical engineering with the scalability of machine learning (ML) to produce high-resolution, probabilistic hazard maps for disaster planning and impact assessment.

2. Methodology

The study utilizes and demonstrates a novel framework developed by Sanger et al. (2025), termed Geospatial Liquefaction Models (GLMs).

A. The Surrogate Modeling Approach

Instead of predicting liquefaction directly from surface features, the GLMs act as surrogates for established CPT-based geotechnical models.

Target Models: The ML models are trained to mimic three industry-standard geotechnical indices:
1. Liquefaction Potential Index (LPI)
2. Modified LPI (LPIISH)
3. Liquefaction Severity Number (LSN)
Training Data: The models were trained on a global dataset of ~37,000 CPT profiles sourced from regional, national, and global repositories.
Feature Engineering: 37 globally available geospatial variables were used as inputs (features), including:
- Topographic variables (slope, curvature, ruggedness).
- Hydrologic data (depth to groundwater, distance to rivers).
- Soil properties (sand/silt/clay fractions, bulk density).
- Geologic proxies (depth to bedrock, Vs30).
Machine Learning Architecture: A bagged decision tree ensemble was selected after iterative evaluation to minimize Mean Squared Error (MSE). The model learns the relationship between geospatial features and the parameters ( $A$ and $B$ ) required to fit the geotechnical indices to Peak Ground Acceleration (PGA).

B. Geostatistical Updating

To address the "data-sparse" problem, the model employs regression kriging:

Local Correction: In areas near known CPT sites, predictions are geostatistically updated using spatial interpolation of residuals. This ensures predictions proximal to measured data are driven by actual subsurface conditions.
Global Prediction: In data-sparse areas, predictions rely on the ML surrogate model informed by global geospatial variables.

C. Scenario Implementation

The study applied the GLMs to 85 scenario earthquakes in Washington and Oregon, derived from three USGS catalogs:

BSSC14: 54 deterministic ruptures affecting Oregon.
WA22: 26 scenarios involving active Quaternary faults in Washington.
CSZM9: 5 ensemble scenarios (2nd to 98th percentile) of a magnitude 9.0 Cascadia Subduction Zone (CSZ) rupture.

For each scenario, the workflow involved:

Ingesting USGS ShakeMap PGA data.
Converting PGA to magnitude-scaled PGA ( $PGA_M$ ).
Generating spatial rasters for LPI, LPIISH, and LSN.
Calculating the Probability of Ground Failure (PGF) using fragility functions (Geyin and Maurer, 2020).
Creating an ensemble PGF by averaging the results of the three geotechnical models.

3. Key Contributions

Mechanistic ML Integration: The study bridges the gap between data-driven ML and physics-based geotechnical engineering. The models are "mechanically informed" because they surrogate established geotechnical equations rather than learning purely from surface observations.
Hybrid Data Utilization: The framework uniquely combines global geospatial data for broad coverage with local CPT data for high-precision geostatistical updating.
Operational Utility: The models are packaged as a reproducible, GIS-ready workflow (Python/Matlab scripts) that allows users to rapidly generate hazard maps for any scenario without requiring ML expertise or high-performance computing.
Comprehensive Regional Assessment: The study provides the first high-resolution, regional-scale liquefaction hazard forecasts for 85 distinct earthquake scenarios across the Pacific Northwest.

4. Results

A. Model Performance

Benchmarking: The GLMs were tested against the state-of-practice model (Rashidian and Baise, 2020; "RB20") using the 2001 Nisqually earthquake case histories and global "unseen" datasets (Puerto Rico, Ridgecrest, Turkey).
Metrics: Using the Brier Score (BS), the GLMs significantly outperformed RB20.
- In global unseen tests, the GLM ensemble improved the BS by approximately 1.5 standard deviations (Cohen's $d \approx -1.59$ ).
- In the Pacific Northwest (Nisqually), the GLMs showed similar trends of improvement, though statistical significance was limited by the small sample size of case histories.
Interpretation: Lower BS values indicate the GLMs are better calibrated and more accurate in predicting the probability of liquefaction manifestation.

B. Hazard Forecasts

Spatial Patterns: High PGF values consistently correlated with low-lying, saturated alluvial deposits and river confluences (e.g., Portland's waterfront, Seattle's Duwamish River corridor, Snohomish River delta).
Crustal vs. Subduction Events:
- Crustal Faults (e.g., Seattle Fault, Portland Hills Fault): Produced higher localized PGF values (up to 25% higher in urban cores) due to proximity to the rupture.
- Cascadia Subduction Zone (CSZ): While PGF values were slightly lower in specific urban cores due to distance, the spatial extent of the hazard was immense, affecting hundreds of square kilometers and multiple regional hubs simultaneously.
Impact Analysis:
- Infrastructure: High PGF zones overlapped significantly with critical infrastructure, including Interstate-5, port facilities, and bridge networks.
- Population: In a median M9.0 CSZ scenario, >4 million people were exposed to PGF >50%. Areas with population density >1,000 persons/km² showed a median PGF of nearly 40%.
- Assets: The study identified thousands of earth dams and arterial bridges in western Washington susceptible to liquefaction, particularly in estuarine environments.

5. Significance and Limitations

Significance

Decision Support: The outputs serve as actionable tools for land-use planning, evacuation route optimization, infrastructure vulnerability screening, and insurance loss modeling.
Scalability: The approach demonstrates that high-fidelity liquefaction forecasting is possible even in data-sparse regions by leveraging global datasets and ML surrogates.
Open Science: All 85 scenario results are published in a public, GIS-ready repository (DesignSafe), democratizing access to advanced hazard modeling.

Limitations and Uncertainties

Site-Specific Data: The models are surrogates; site-specific geotechnical analysis and measured data should always supersede GLM predictions at specific locations.
Input Uncertainties: Uncertainties in ground motion simulations (source, path, site effects) and static assumptions for dynamic variables (e.g., groundwater depth) propagate through the model.
Asset Data Quality: Impact forecasts rely on the accuracy and completeness of infrastructure inventories (e.g., bridge and dam databases), which can be sparse.
Model Constraints: The models may be less accurate in highly interbedded soil profiles or for extreme magnitude subduction events where training data is limited.

Conclusion

This study validates the Geospatial Liquefaction Models (GLMs) as a superior alternative to existing regional mapping approaches. By integrating machine learning with mechanistic geotechnical principles and geostatistical updating, the GLMs provide a robust, scalable, and physically grounded framework for forecasting liquefaction hazards and their societal impacts, significantly advancing earthquake preparedness in the Pacific Northwest and beyond.