Skills & Technologies

Machine Learning / Deep Learning

• Generative Adversarial Networks (GANs) — Conditional WGAN-GP architecture with gradient penalty for stable training
• Physics-Informed Machine Learning — Custom monotonic distance-attenuation penalty encoding seismological prior knowledge into the loss function
• PyTorch — Model definition (nn.Module), custom training loop, autograd for gradient penalty computation, GPU acceleration
• Residual Networks — Pre-activation residual blocks with LayerNorm for both Generator and Critic
• Learned Embeddings — Shared period embedding MLP mapping continuous spectral period to a higher-dimensional representation

Data Engineering

• Pandas — Loading, cleaning, and reshaping (~10K records x 48 columns) from wide to long format (255K samples)
• Feature Engineering — Log transforms (Rrup, Vs30, Period, SA), PGA replacement for log-domain compatibility
• scikit-learn — StandardScaler for feature normalization, train/test splitting, regression metrics (RMSE, MAE)
• Serialization — Model checkpointing with

  torch.save

  , scaler persistence with

  joblib

Domain Knowledge

• Earthquake Engineering — Ground Motion Models, Spectral Acceleration, NGA-Subduction database
• Seismological Parameters — Moment magnitude (Mw), rupture distance (Rrup), depth to top of rupture (Ztor), site shear-wave velocity (Vs30)
• Physical Constraints — Distance-attenuation relationship (SA decreases with increasing Rrup)
• Response Spectra Analysis — Per-event evaluation across 25 spectral periods (PGA to T=10s)

Evaluation & Visualization

• Matplotlib — Loss curves, real-vs-predicted scatter plots, residual analysis, per-event response spectra
• Diagnostic Plots — Residuals vs spectral period for period-dependent bias detection
• Regression Metrics — RMSE and MAE on held-out test set in log(SA) space

Development Environment

• Google Colab — Primary development environment with CUDA GPU
• Jupyter Notebooks — Iterative experimentation with inline visualization
• Git / GitHub — Version control and project hosting