Physics-Informed CWGAN-GP for Earthquake Ground Motion Prediction

A Conditional Wasserstein GAN with Gradient Penalty (CWGAN-GP) that learns to predict earthquake Spectral Acceleration (SA) from seismological parameters, augmented with a physics-informed monotonic distance-attenuation penalty.

This is Phase 1 of a multi-phase research project developed in collaboration with IIT Mandi.

Overview

Ground Motion Models (GMMs) are essential in seismic hazard analysis for predicting the intensity of earthquake shaking at a site. This project replaces traditional empirical GMMs with a generative model that captures the full conditional distribution of Spectral Acceleration.

Generator input: earthquake metadata (Mw, Rrup, Ztor, Vs30) + spectral period T + latent noise z Generator output: predicted log(SA) (natural log of Spectral Acceleration)

Key Design Choices

• WGAN-GP training with gradient penalty (lambda_GP = 10) and 5:1 critic-to-generator update ratio for stable training
• Physics-informed monotonic penalty (lambda_mono = 10) enforces the physical constraint that SA must decrease with increasing rupture distance (Rrup)
• Period Embedding MLP — a shared 2-layer network maps the 1D log(Period) to a 16-dimensional representation used by both Generator and Critic
• Residual blocks with LayerNorm — both Generator and Critic use pre-activation residual blocks (2 x ResBlock(128))

Architecture

| Component | Parameters | |------------|-----------| | Generator | 74,545 | | Critic | 70,577 |

Dataset

NGA-Subduction (NGA-Sub) — a widely used strong-motion database in earthquake engineering.

• Source: nga_subduction.xlsx (10,239 records x 48 columns)
• Conditioning variables: Mw (magnitude), Rrup (rupture distance), Ztor (depth to top of rupture), Vs30 (site shear-wave velocity)
• Target: Spectral Acceleration at 25 periods from PGA (T=0) to T=10s
• Working dataset: 255,975 samples after melting to long format (one row per record-period pair)
• Train/Test split: 80/20 (204,780 / 51,195 samples)

Results

Training for 100 epochs with batch size 2048, Adam optimizer (lr=1e-4, betas=(0.5, 0.9)).

| Loss Curves | Real vs Predicted | |:-----------:|:-----------------:| | | |

| Residuals vs Period | Response Spectra (Single Event) | |:-------------------:|:-------------------------------:| | | |

The monotonicity penalty converges to near zero (~0.0001), confirming the physics constraint is well-satisfied.

Project Structure

.
├── Phase1_CWGAN_GP.ipynb        # Main notebook (current)
├── Phase1_CWGAN_GP_old.ipynb    # Earlier version (no train/test split)
├── Phase1_Architecture.png      # Model architecture diagram
├── nga_subduction.xlsx           # NGA-Sub earthquake dataset
├── global_G.pth                  # Trained Generator weights
├── global_D.pth                  # Trained Critic weights
├── condition_scaler.pkl          # Fitted StandardScaler for conditioning features
├── loss_curves.png               # Training loss plots
├── real_vs_pred.png              # Test-set scatter plot
├── residuals_vs_period.png       # Residuals vs spectral period
├── response_spectra_event.png    # Per-event response spectra comparison
└── old/                          # Artifacts from a previous run

Getting Started

Prerequisites

torch
numpy
pandas
matplotlib
scikit-learn
joblib
openpyxl

Running on Google Colab (recommended)

1. Upload the project folder to Google Drive under MyDrive/WGAN_IITM/
2. Open Phase1_CWGAN_GP.ipynb in Colab
3. The notebook auto-detects Colab and mounts Google Drive
4. Run all cells sequentially

Running Locally

pip install torch numpy pandas matplotlib scikit-learn joblib openpyxl
jupyter notebook Phase1_CWGAN_GP.ipynb

A CUDA GPU is recommended but not required — the models are small (~75K and ~71K parameters).

Saved Artifacts

After training completes, the following artifacts are saved for downstream use (Phase 2):

| File | Description | |------|-------------| | global_G.pth | Generator state dict | | global_D.pth | Critic state dict | | condition_scaler.pkl | Fitted StandardScaler for the 5 conditioning features |

License

This project is part of ongoing research. Please contact the authors before using the code or data for publications.