Combining Direct Simulation and Surrogate Modelling for Simulation-Consistent Probabilistic Seismic Hazard Analysis

This repository contains the code for reproducing the results from the paper.

Note: It is advised to use a Linux machine. All instructions below assume a Linux OS. If you encounter any problems, please raise an issue on the GitHub repository.

Installation

Prerequisites

The following libraries must be available:

• GMT version 6.6

Setup

1. Clone the repository:

   git clone <repository-url>

2. Install the package:

   pip install ./post_event_gm_gnn

Data Setup

Environment Configuration

Most data loading/saving is done relative to the wdata environment variable.

1. Create a data directory for this project (e.g., base_data_dir)
2. Add the following to your .bashrc (or equivalent):

   export wdata=/path/to/base_data_dir

Download Data

Download the data file from here and unzip it.

This will create the following directory structure:

base_data_dir
    |--- 20251222_CS200_DSSlab_rotd50_empdb.duckdb
    |--- 20251222_CS200_DSSlab_rotd50_imdb.duckdb
    |--- test_events.npy
    ├── cs200_hazard
    ├── other
    │   ├── dupuis_moderate_mag
    │   ├── dupuis_small_mag
    │   └── lee_nz_validation_2022
    ├── results
       ├── 0613_1752_cv_baseV5p1_25Epochs_6E6Folds
       │   ├── event_mera_site_term
       │   └── mera_site_term
       ├── 0613_2113_full_baseV5p1_10Epochs
       │   ├── ds_hazard
       │   │   ├── ref_sites
       │   │   └── uniform_grid
       │   └── test_mera
       ├── 0613_2142_cv_baseV5p1_locAdjV4_25Epochs_6E6Folds
       │   ├── event_mera_site_term
       │   └── mera_site_term
       ├── 0614_0214_full_baseV5p1_locAdjV4_10Epochs
       │   ├── ds_hazard
       │   │   ├── ref_sites
       │   │   └── uniform_grid
       │   └── test_mera
       ├── emp_ds_hazard
       │   ├── ref_sites
       │   └── uniform_grid
       └── loc_models
           └── 0929_1130_optimized_siteCond_25Epochs

Data Overview

Databases

• 20251222_CS200_DSSlab_rotd50_imdb.duckdb: Primary data resource. Contains all rupture, source, and site information, in addition to the simulation IM values.
• 20251222_CS200_DSSlab_rotd50_empdb.duckdb: Contains IM distribution parameters for the empirical GM logic tree

Results

The results folder contains the following

• 0613_1752_cv_baseV5p1_25Epochs_6E6Folds - Cross-validation (CV) results for the base surrogate model
• 0613_2142_cv_baseV5p1_locAdjV4_25Epochs_6E6Folds - Cross-validation results for the location-specific surrogate model
  • Each of these contains:
    • One directory per cross-validation run
    • A mera_site_term folder, which contains the mixed-effects regression results
    • A event_mera_site_term folder, which contains the mixed-effects regression results at the event level
    • A val_results.parquet file, which is a dataframe of the combined validation results
• 0613_2113_full_baseV5p1_10Epochs - Full base surrogate model
• 0614_0214_full_baseV5p1_locAdjV4_10Epochs - Full location-specific surrogate model
  • Each of these contains:
    • test_mera folder, which contains the test dataset mixed-effects regression results
    • test_results.parquet file, which is a dataframe with test dataset results
    • ds_hazard folder, which contains the distributed seismicity hazard results
      • ref_sites distributed seismicity hazard at a small set of reference sites (aligned with Cybershake NZ 200m paper)
      • unfiorm_grid distributed seismicity hazard on the uniform grid
• emp_ds_hazard - Empirical GM logic tree distributed seismicity hazard results
• loc_models/0929_1130_optimized_siteCond_25Epochs - Pre-trained location-embedding model

Other

The other folder contains the mixed-effects regression results for the Lee et al. (2022) and Dupuis et al. (2025) simulation validation studies.

The cs200_hazard folder contains the fault and distributed seismicity ERF for the 2010 NZ NHM.

Reading Data Files:

• .pickle files: Use pd.read_pickle() from pandas
• .parquet files: Use pd.read_parquet() from pandas
• .npy files: Use np.load() from numpy
• .pt files: Use torch.load(filepath, weights_only=False) from PyTorch

Model Training

Use the ./train_models.sh script for training the models. Note, that the base CV & full models have to be trained first, as these are needed to subsequently train the location-specific models. A GPU is highly recommended, it may be possible without, but not tested and most likely extremely slow. Modify the number of processes to suit the system used, more processes results in higher memory usage (both main and GPU).

Post-processing

Use the run_cv_post.sh script to compute mixed-effects regression results for the CV runs. For the full models, use run_full_post.sh to compute distributed seismicity hazard results. Both are slow and require large amounts of memory, run times were ~12 hours and ~12-24 hours, respectively, on a system with 16 CPU cores and 64 GB of memory. Number of processes is configured for this system, adjust as needed.

Generating Paper Figure

Use the gen_paper_figures.sh script to compute the paper figures.