CGM-GM

Learning Physics for Unveiling Hidden Earthquake Ground Motions via Conditional Generative Modeling

Paper link: arXiv

---

Overview

This work addresses a critical problem in seismic hazard assessment and infrastructure resilience: predicting high-fidelity ground motions for future earthquakes. We propose a new AI simulator, Conditional Generative Modeling for Ground Motion (CGM-GM), to synthesize high-frequency and spatially continuous earthquake ground motion waveforms. CGM-GM uses earthquake magnitudes and geographic coordinates of earthquakes and sensors as inputs.

Highlights

• Learning physics: CGM-GM can capture the underlying spatial heterogeneity and physical characteristics.

• Comprehensive evaluations: The framework provides comprehensive evaluations in time and frequency domains.

• Great potential: CGM-GM demonstrates a strong potential for outperforming a state-of-the-art non-ergodic empirical ground motion model and shows great promise in seismology and beyond.

Below is an example of generated FAS maps in the San Francisco Bay Area.

---

System requirements and installation

Hardware requirements

All the experiments are performed on an NVIDIA Titan RTX Graphics Card.

Python requirements

Install the required dependencies:

conda create -n cgm_gm python=3.9.17
conda activate cgm_gm
pip install -r requirements.txt

---

Usage

Dataset

The earthquake dataset in the SFBA was originally downloaded from NCEDC. The training and testing dataset in this study is preprocessed and can be found in a data report, which will be made public soon.

Implementations

1. Train the CGM-GM with Hyperopt for hyper-parameters optimization:

   python train_hyperopt.py
   
   # if training CGM-baseline (with rupture distances)
   python train_hyperopt.py --tcondvar 4

2. Evaluate the best model:

   python test_best_model.py

3. Generate waveforms from a 100x100 grid:

   python generate_points.py

The implementations of ergodic and non-ergodic GMM for the SFBA can be found in this paper.

The evaluations include the comparisons of waveform shapes, P and S arrival time, amplitude spectra, and Fourier amplitude spectra maps. We use the PhaseNet to pick the arrival time of P and S waves.

License

This project is released under the GNU General Public License v3.0.