Earthquake Ground Motion Prediction — NGA-West2

Predicting structural displacement demand SD(1.0) from the NGA-West2 ground motion database. Compares classical GMPE regression against modern ML models (Random Forest, XGBoost, Physics-Informed Neural Networks) with rigorous earthquake-group cross-validation.

See REPORT.md for the full project writeup.

---

Setup

Requires Python 3.10+.

python3 -m venv .venv
source .venv/bin/activate
pip install numpy pandas matplotlib seaborn scikit-learn xgboost torch scipy openpyxl requests

Reproduce from scratch

Step 1 — Download data and build database (~1.5 GB download, requires PEER account):

python download.py

This downloads the NGA-West2 flatfiles and supporting data from PEER and loads everything into earthquake.db.

Step 2 — Exploratory data analysis:

python eda.py           # → eda_figures/

Step 3 — Spectral interpolation pipeline (Models A–H):

python pipeline.py      # → model_figures/, model_comparison.csv

Step 4 — GMPE-style pipeline (source/site features only):

python pipeline_gmpe.py # → gmpe_figures/

Step 5 — Neural network tuning:

python tune_nn.py       # → nn_figures/, model_comparison_nn.csv

Step 6 — Physics-informed neural network:

python physics_nn.py    # → physics_nn_figures/

Step 7 — Diagnostic comparison (GMPE vs XGBoost):

python diagnostics.py   # → diagnostics_figures/

---

Key Results

All results use an earthquake-group train/test split — each earthquake appears entirely in train or entirely in test, preventing leakage across the 36 stations that typically record each event.

Spectral interpolation (recorded SA neighbors as features)

Model	RMSE (log₁₀)	R²
OLS — PGA only	0.572	0.822
OLS — spectral shape	0.086	0.996
Ridge / Lasso / XGBoost	~0.086–0.092	~0.996

Once neighboring spectral values are available, prediction is near-deterministic — all models saturate at R²=0.996.

GMPE-style (source/site only — the real engineering problem)

Model	RMSE (log₁₀)	R²
OLS — M + log(R) + log(Vs30)	0.440	0.883
Classic GMPE OLS (full form)	0.376	0.914
Physics-Informed NN	0.370	0.917
XGBoost	0.340	0.937

Sigma decomposition

Model	τ (inter-event)	φ (intra-event)	σ total
Classic GMPE (OLS)	0.285	0.285	0.404
XGBoost	0.253	0.260	0.363
Aleatory floor	~0.13–0.17	~0.20–0.24	~0.25–0.30

Both models are approaching the aleatory floor — the irreducible scatter in ground motion that cannot be reduced without path-specific observables.

---

Project Structure

earthquake/
├── download.py          # Data download and SQLite ingestion
├── eda.py               # Exploratory data analysis
├── pipeline.py          # Spectral interpolation pipeline (Models A–H)
├── pipeline_gmpe.py     # GMPE-style pipeline (source/site only)
├── tune_nn.py           # PyTorch hyperparameter search
├── physics_nn.py        # Physics-informed neural network
├── diagnostics.py       # GMPE vs XGBoost sigma decomposition & bias analysis
├── SPEC.md              # Original project specification
├── REPORT.md            # Full project report (~8 pages)
├── model_comparison.csv         # Spectral pipeline results table
├── model_comparison_nn.csv      # NN tuning results table
├── eda_figures/         # EDA plots
├── model_figures/       # Spectral pipeline plots
├── gmpe_figures/        # GMPE pipeline plots
├── nn_figures/          # NN tuning plots
├── physics_nn_figures/  # Physics-informed NN plots
└── diagnostics_figures/ # Diagnostic comparison plots

Data

NGA-West2 is available at PEER Ground Motion Database. A free account is required. download.py automates the download and ingestion.

The database sentinel value −999 is treated as NULL throughout.