Deep Learning for Skillful Coastal Wave Downscaling: Remote Wind as Boundary Conditions

Deep learning implementation for coastal wave parameter downscaling using multi-scale wind data. Employs 3D U-Net architecture for significant wave height (hs), wave direction (dir), and mean wave period (t02) prediction.

Architecture

Wendy3d: Multi-encoder 3D U-Net with temporal-spatial processing

• Large-scale wind encoder: 3D U-Net (10-day temporal window)
• Local wind encoder: 3D U-Net (3-day temporal window)
• Fusion decoder: 2D U-Net for final prediction

Installation

pip install -r requirements.txt

Data Structure

processed_data_dir/
├── CFSR/input/wind_input_{au_large,au_local}/
└── WW3/output/{hs,dir,t02}/au/

Usage

Training:

python run_all_params.py    # All parameters
python train.py             # Single model

Prediction:

python predict_metrics.py

Configuration:

config = Config()
config.y_data_desc = "hs"   # Target: hs, dir, t02
config.model_filename = "Wendy3d"

Model Specifications

• Input: Large wind (3×61×174), Local wind (2×61×121)
• Temporal Windows: 80 steps (large), 24 steps (local)
• Loss: Masked MSE (hs, t02), Masked Angle Cosine (dir)
• Optimizer: Lion with weight decay
• Training: Mixed precision, early stopping (patience=7)