Platform Science Use Cases Team Contact
Scientific Methodology

A New Scientific Standard for
Liquefaction Prediction.

At the core of GeoLiquefy is a domain-aware AI system trained to understand geotechnical behavior, seismic stresses, and soil dynamics with scientific precision.

AI + Geotechnics: A Hybrid Scientific Approach

We combine the rigor of traditional geotechnical engineering with the power of modern AI to create a system that respects physical laws while capturing complex non-linearities.

Traditional Geotechnical Engineering

  • Empirical correlations (Seed & Idriss, Boulanger & Idriss)
  • Simplified boundary value problems
  • Factor of safety calculations
  • Decades of field validation

Probabilistic Seismic Hazard Analysis

  • Ground motion prediction equations
  • Site response analysis
  • Uncertainty quantification
  • Risk-based decision frameworks

Physics-Informed AI

  • Domain constraints in loss functions
  • Multi-scale feature learning
  • Interpretable reasoning chains
  • Continuous model refinement

Soil Stratification & Seismic Response

Ground Surface Layer 1: Sandy Silt (SPT=8, Vs=150m/s) Layer 2: Loose Sand (SPT=12, Vs=180m/s, fc=15%) Layer 3: Dense Sand (SPT=25, Vs=280m/s) Water Table Seismic Wave Propagation

Our AI Architecture

A multi-layered system that combines domain expertise with deep learning to deliver accurate, interpretable predictions.

01

Input Parsing Layer

SPT, CPT, Vs, groundwater depth, fines content, seismic parameters

02

Geotechnical Reasoning Layer

Domain-aware fine-tuned LLM incorporating geotechnical equations, soil mechanics, and empirical relationships

03

Physics & Code-Based Constraints

Safety factors, liquefaction triggering formulations, Seed & Idriss logic validation

04

AI Prediction Engine

Multi-model ensemble, confidence quantification, uncertainty bounds

05

Interpretability & Explanation Layer

Step-by-step reasoning, parameter sensitivity analysis, feature attribution

06

Output Layer

Final liquefaction probability, risk classification, engineering explanation

Training the Model

Our models are trained on the most comprehensive liquefaction dataset ever assembled, with rigorous quality control and domain-balanced sampling.

📊

Dataset Sources

  • 50,000+ global CPT/SPT case histories
  • Peer-reviewed earthquake studies
  • Post-earthquake field reconnaissance data
  • Curated laboratory test results
  • Historical seismic records
⚖️

Domain-Balanced Sampling

  • Geographic diversity (Japan, NZ, USA, Taiwan)
  • Soil type stratification
  • Earthquake magnitude distribution
  • Liquefaction vs non-liquefaction balance
  • Uncertainty-weighted sampling
🔬

Physics-Constrained Training

  • Loss function with domain penalties
  • Monotonicity constraints on key parameters
  • Energy conservation checks
  • Boundary condition validation
  • Stress-strain consistency
🎯

Fine-Tuning Strategy

  • Pre-training on general geotechnical corpus
  • Task-specific fine-tuning on liquefaction
  • Reasoning chain optimization
  • Adversarial robustness testing
  • Continuous model updates
🧪

Synthetic Scenario Generation

  • Physics-based data augmentation
  • Rare event simulation
  • Edge case coverage
  • Interpolation validation
  • Extrapolation boundary testing
🌍

Cross-Earthquake Generalization

  • Leave-one-earthquake-out validation
  • Regional transfer learning
  • Temporal consistency checks
  • Multi-site performance tracking
  • Bias detection and mitigation

Model Performance Convergence

Training Epochs Accuracy (%) 60 70 80 90 95

Scientific Validation & Benchmarks

Our models are rigorously evaluated against classical methods and real-world earthquake data to ensure reliability and accuracy.

Metric Traditional Methods GeoLiquefy AI Improvement
Overall Accuracy 78% 94% +16%
False Negative Rate 18% 4% -14%
False Positive Rate 24% 8% -16%
Confidence Calibration N/A 0.92 New
Processing Time ~30 min <1 sec 1800x faster

Cross-Event Validation

2011 Tōhoku Earthquake

Japan • Mw 9.1

Accuracy 96%

Validated against 2,400+ field observations

2010-2011 Christchurch

New Zealand • Mw 6.3

Accuracy 93%

Extensive CPT database validation

1999 Chi-Chi Earthquake

Taiwan • Mw 7.6

Accuracy 91%

Cross-validated with SPT records

1964 Niigata Earthquake

Japan • Mw 7.5

Accuracy 89%

Historical case study validation

"Our models have been reviewed and validated by leading geotechnical experts including Prof. Ashok Kumar Gupta, Prof. Robb Moss, and Prof. Saurabh Rawat, ensuring scientific rigor and practical applicability."

Interpretability & Reasoning

GeoLiquefy is not a black box. Every prediction comes with step-by-step reasoning, parameter sensitivity analysis, and engineering-readable explanations.

Step-Wise Explanation Generation

Our models generate human-readable reasoning chains that show exactly how input parameters lead to final predictions.

Parameter Sensitivity Maps

Understand which soil and seismic parameters have the greatest influence on liquefaction risk for any given site.

Probability Decomposition

Break down the final probability into contributing factors: soil strength, seismic demand, and site conditions.

Engineering-Readable Rationale

Every output includes plain-language explanations that engineers can include directly in reports.

Sample Reasoning Trace

Input Parameters: SPT = 12, Vs = 180 m/s, fc = 15%, σ'v = 85 kPa, Mw = 7.0, PGA = 0.35g
Step 1: Normalized SPT (N1)60 = 14.2 → Indicates loose to medium-dense sand
Step 2: Cyclic stress ratio (CSR) = 0.28 → High seismic demand
Step 3: Cyclic resistance ratio (CRR) = 0.18 → Low soil resistance
Step 4: Factor of Safety = CRR/CSR = 0.64 < 1.0 → Liquefaction likely
Step 5: Fines content correction applied → Adjusted CRR = 0.21
Final Prediction: Liquefaction Probability = 87% (High Risk)
Confidence: 94% (based on training data density in this parameter space)

Safety, Reliability, and Responsible Deployment

Engineering ethics and safety are at the core of our platform. We maintain transparent assumptions, clear error boundaries, and human-in-the-loop validation.

🛡️

Transparent Assumptions

All model assumptions, limitations, and applicability ranges are clearly documented and communicated to users.

📏

Error Boundaries

Confidence intervals and uncertainty bounds are provided for every prediction to support risk-informed decision making.

⚠️

Suitable Data Regimes

The system alerts users when input parameters fall outside validated ranges or when predictions may be unreliable.

👥

Human-in-the-Loop

Critical decisions require professional engineering review. Our platform augments, not replaces, human expertise.

Reliability Thresholds

Minimum confidence thresholds ensure predictions meet engineering standards before being presented to users.

🔍

Continuous Monitoring

Model performance is continuously tracked against new earthquake data to detect and correct any drift or degradation.

Technical White Paper

A comprehensive deep dive into our model architecture, training methodology, scientific validation, and performance benchmarks.

Download the White Paper →

What's Next: The Future of GeoHazard Intelligence

We're pushing the boundaries of what's possible in geohazard prediction, working toward a future where every city has real-time, AI-powered resilience intelligence.

Regionalized Models

Japan-specific, California-specific, and New Zealand-specific models that capture local geological characteristics and seismic patterns.

City-Scale Susceptibility Maps

High-resolution liquefaction risk maps for entire metropolitan areas, updated in real-time as new data becomes available.

Multi-Hazard Fusion

Integrated prediction of liquefaction, lateral spreading, ground deformation, and settlement in a unified framework.

Multi-Modal Inputs

Incorporating borehole logs, GIS layers, satellite imagery, and historical seismic records for richer context.

Digital Twin Integrations

Real-time integration with city digital twins for scenario planning and infrastructure resilience assessment.

Long-Term Resilience Forecasting

Probabilistic forecasting of seismic risk over decades, accounting for climate change and urban development patterns.