Research map / Probabilistic and scientific modeling / Differential equations and PDEs
CoastalBench: A Decade-Long High-Resolution Dataset to Emulate Complex Coastal Processes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces CoastalBench, a decade-long (2008–2017) high-resolution ROMS coastal circulation dataset for Charlotte Harbor, Florida, and benchmarks a customized ViT emulator on it.
Problem
Coastal numerical models (ROMS) are slow and expensive, and deep learning progress for coastal process emulation is hindered by the lack of large-scale, high-resolution simulation datasets.
Approach
A ViT-based emulator that tokenizes each ocean and meteorological variable separately with cross-attention aggregation, injects static bathymetry/grid features into positional embeddings, and conditions attention on lead time, boundary conditions, and river inflows, trained with delta-prediction and autoregressive rollouts.
Key results
- ViT model predicts 72-hour forecasts; reported runtime drops from 2,477 s on 512 CPU cores (ROMS) to 34.14 s on a single A100 GPU, a over 70x speedup.
- Ablation shows removing boundary conditions sharply increases velocity RMSE, and removing river inflow most strongly affects salinity.
- Scaling test shows ViT-B outperforms ViT-T and ViT-S, with the gap widening at 48-72 hour lead times.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the CoastalBench ViT instead of ROMS for a 72-hour forecast | CoastalBench ViT · 72-hour forecast | runtime: 2,477 seconds to 34.14 seconds; over a 70× speedup | Helped |
| Remove meteorological forcings (ablation) | CoastalBench ViT · 72-hour forecasting; salinity and temperature | RMSE | Hurt |
| Remove static physical features (ablation) | CoastalBench ViT · 72-hour forecasting; temperature and salinity | RMSE | Hurt |
| Remove boundary conditions (ablation) | CoastalBench ViT · 72-hour forecasting; velocity | RMSE | Hurt |
| Remove river inflow (ablation) | CoastalBench ViT · 72-hour forecasting; salinity | RMSE | Hurt |
| Scale the model up from ViT-Tiny and ViT-Small to ViT-Base (ablation) | CoastalBench ViT · 72-hour forecasting; key ocean variables | RMSE | Helped |
Surprising findings
- Vertical velocity (w) and temperature show anomalous RMSE/MAE or unstable correlation trends, attributed to sparse nonzero w and low variance in temp.
Limitations
- Error accumulates over long autoregressive forecasts, especially in estuarine zones and coastal boundaries.
- Only ViT-Base-scale models tested; larger models and longer autoregressive fine-tuning not explored due to compute.
- Purely data-driven model without physical constraints may yield physically inconsistent predictions.
Experimental setup
Datasets: CoastalBench (ROMS simulations, 2008-2017) HYCOM + NCODA global analysis (boundary forcings) NARR (atmospheric forcings) USGS river flow data NOAA CUDEM bathymetry USGS 2018 Lidar DEM RECON observations (calibration)
Benchmarks: CoastalBench
Metrics: RMSE MAE Pearson correlation
Baselines: ROMS numerical simulation ViT-S ViT-T
Builds on: Vision Transformer Pangu-Weather ROMS adaLN
Closest papers by meaning
- Early Detection of Extreme Storm Tide Events Using Multimodal Data Processing (AAAI 2024 · similarity 0.85)
- MFogHub: Bridging Multi-Regional and Multi-Satellite Data for Global Marine Fog Detection and Forecasting (CVPR 2025 · similarity 0.85)
- SeafloorAI: A Large-scale Vision-Language Dataset for Seafloor Geological Survey (NeurIPS 2024 · similarity 0.85)
- FIDLAR: Forecast-Informed Deep Learning Architecture for Flood Mitigation (AAAI 2025 · similarity 0.85)
- FuXi-Ocean: A Global Ocean Forecasting System with Sub-Daily Resolution (NeurIPS 2025 · similarity 0.84)
- Off to new Shores: A Dataset & Benchmark for (near-)coastal Flood Inundation Forecasting (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.