Naturecode/coastal-dynamics

Naturecode Coastal Dynamics

State-of-the-Art AI for Coastal Wave Dynamics and Ocean Modeling

Initial Release.0 - Foundation Release

Designed to augment/replace traditional numerical models like MIKE 21

---

Overview

Naturecode Coastal Dynamics is a cutting-edge deep learning system for predicting coastal wave dynamics, sediment transport, and ocean conditions. This foundation release establishes the core architecture incorporating the latest advances from 2025-2026 oceanographic AI research.

---

Architecture Features

Feature	Source	Description
Fourier Neural Operator (FNO)	Li et al. 2021	Spectral convolutions for PDE solving
Mixture-of-Time (MoT)	FuXi-Ocean NeurIPS 2025	Adaptive temporal fusion for multi-scale forecasting
Adaptive Layer Normalization (AdaLN)	FuXi-Ocean NeurIPS 2025	Context-aware normalization
Earthformer Cuboid Attention	NeurIPS 2022 + Science Advances 2024	Remote swell detection from distant storms
Mamba Neural Operator	J. Comp Physics Dec 2025	90% error reduction over Transformers
MC Dropout	XWaveNet	Uncertainty quantification via Monte Carlo dropout
Energy Conservation Loss	OceanCastNet	Physical constraint for long-term stability
Diffusion Refinement	OmniCast/GenCast	Probabilistic ensemble forecasting
VAE Latent Compression	OmniCast NeurIPS 2025	Efficient latent-space diffusion
Extreme Event Detection	XWaveNet	Multi-threshold wave height exceedance prediction

Model Statistics

Specification	Value
Parameters	14,027,854 (14M)
Architecture	Hybrid FNO + Mamba + Swin Transformer
Input Channels	8
Output Channels	5
Training Epochs	500
Training Hardware	8x NVIDIA H100 GPUs
Training Data	18.4M real ocean observations

---

Input/Output Specification

Input Channels (8)

Channel	Description	Units
0	Bathymetry	meters (negative = depth)
1	Wind U-component	m/s
2	Wind V-component	m/s
3	Previous wave height	meters
4	Previous U-velocity	m/s
5	Previous V-velocity	m/s
6	Previous surface elevation	meters
7	Time encoding	normalized [0, 1]

Output Channels (5)

Channel	Description	Units
0	Significant wave height	meters
1	U-velocity	m/s
2	V-velocity	m/s
3	Surface elevation (eta)	meters
4	Sediment transport	kg/m2/s

---

Intended Use

Primary Use Cases

• Coastal Engineering: Wave prediction for harbor design, breakwater planning
• Climate Adaptation: Storm surge and extreme event forecasting
• Environmental Monitoring: Sediment transport and coastal erosion prediction
• Marine Operations: Sea state forecasting for shipping and offshore operations
• Research: Accelerating ocean/coastal simulations (1000x faster than MIKE 21)

Out-of-Scope Uses

• Real-time tsunami warning (requires specialized systems)
• Operational weather forecasting without domain validation
• Areas without adequate bathymetric data

---

Training Data

Data Sources

1. Synthetic Physics-Based Data: Generated using simplified shallow water equations
2. NOAA NDBC Buoy Data: Real ocean observations from 60 buoys (2015-2025)
   • Records: 18.4 million timestamped observations
   • Coverage: Pacific, Atlantic, Gulf of Mexico, Hawaii
   • Variables: Wave height, period, direction, wind, SST, pressure

---

How to Use

Installation

pip install torch numpy

Basic Inference

import torch
from model import CoastalDynamicsModel

# Load model
model = CoastalDynamicsModel(embed_dim=128, dropout=0.1)
checkpoint = torch.load('pytorch_model.pt', map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Prepare input (B, 8, H, W)
inputs = torch.randn(1, 8, 128, 128)

# Forward pass
with torch.no_grad():
    outputs = model(inputs, return_uncertainty=True, return_extreme_probs=True)

# Access outputs
wave_height = outputs['mean'][:, 0]      # Significant wave height
uncertainty = outputs['std'][:, 0]        # Prediction uncertainty
extreme_probs = outputs['extreme_probs']  # P(wave > 2m, 4m, 6m, 8m)

Uncertainty Quantification

# Monte Carlo Dropout + Diffusion ensemble
results = model.predict_with_uncertainty(
    inputs, 
    num_mc_samples=20,      # Epistemic uncertainty
    num_diffusion_samples=10 # Aleatoric uncertainty
)

print(f"Mean prediction: {results['mean'].shape}")
print(f"MC uncertainty: {results['mc_std'].shape}")
print(f"Diffusion uncertainty: {results['diffusion_std'].shape}")
print(f"Extreme event probs: {results['extreme_probs'].shape}")

---

Performance

Metric	Value	Description
Charbonnier Loss	0.035	Robust L1-like loss
Physics Loss	0.067	Physical consistency
NLL (Diffusion)	-3.5	Log-likelihood
Energy Conservation	0.000	Perfect conservation
Best Total Loss	-1.01	Combined metric

---

Limitations

1. Spatial Resolution: Optimized for 128x128 grids
2. Temporal Resolution: Best for 6-hourly predictions
3. Geographic Bias: Training data primarily from US coastal waters
4. Extreme Events: Rare events (>99th percentile) have inherent prediction challenges
5. Bathymetry Dependency: Requires accurate bathymetric input

---

Environmental Impact

Metric	Value
Training Hardware	8x NVIDIA H100 GPUs
Training Time	~4 hours
Estimated CO2	~15 kg CO2eq
Cloud Provider	Google Cloud (renewable mix)

---

Citation

@misc{naturecode_coastal_dynamics_2026,
  title={Naturecode Coastal Dynamics: Physics-Informed Deep Learning for Ocean Wave Prediction},
  author={Naturecode Team},
  year={2026},
  version={1.0},
  publisher={Hugging Face},
  url={https://huggingface.co/Naturecode/coastal-dynamics}
}

---

License

This model is released under the Apache 2.0 License.

---

Acknowledgments

This model builds upon research from:

• FuXi-Ocean (NeurIPS 2025)
• OmniCast (NeurIPS 2025)
• OceanCastNet
• XWaveNet
• Earthformer (NeurIPS 2022)
• Mamba Neural Operator (J. Comp Physics 2025)
• NOAA National Data Buoy Center

---

Contact

For questions, collaborations, or access requests:

• Organization: Naturecode

---

Built by Naturecode - Advancing coastal science through AI