CoastalBench: A Decade-Long High-Resolution Dataset to Emulate Complex Coastal Processes

We thank the reviewer for the thoughtful feedback and hope our responses clarify the concerns.

Related Models: Thank you for raising this important point. We agree that baselines are important. To our knowledge, no existing deep learning method is specifically designed for complex regional coastal processes, so we plan to include a 3D U-Net as a baseline in the final version. Due to the large scale of our dataset, we were unable to complete this comparison for the rebuttal. If there are specific baselines you recommend, we welcome your suggestions and will aim to include them.

Related Datasets: We include a comparison table of representative works on regional coastal ocean modeling, highlighting that existing datasets are generally smaller in scale and focus on simpler processes with limited variable coverage.

Kumar, P., & Leonardi, N. (2023). A novel framework for the evaluation of coastal protection schemes through integration of numerical modelling and artificial intelligence into the Sand Engine App.

Ishida, K., et al. (2020). Hourly-scale coastal sea level modeling in a changing climate using long short-term memory neural network.

Melo, C. B., et al. (2023). Coastal morphodynamic emulator for early warning short-term forecasts.

Wei, Z., et al. (2022). A convolutional neural network based model to predict nearshore waves and hydrodynamics.

O'Donncha, F., et al. (2019). Ensemble model aggregation using a computationally lightweight machine-learning model to forecast ocean waves.

Thank you for your comments and thoughtful questions. Below are our responses:

Comparison with existing datasets: We include a comparison table of representative works on regional coastal ocean modeling, highlighting that existing datasets are generally smaller in scale and focus on simpler processes with limited variable coverage.

Kumar, P., & Leonardi, N. (2023). A novel framework for the evaluation of coastal protection schemes through integration of numerical modelling and artificial intelligence into the Sand Engine App.

Ishida, K., et al. (2020). Hourly-scale coastal sea level modeling in a changing climate using long short-term memory neural network.

Melo, C. B., et al. (2023). Coastal morphodynamic emulator for early warning short-term forecasts.

Wei, Z., et al. (2022). A convolutional neural network based model to predict nearshore waves and hydrodynamics.

O'Donncha, F., et al. (2019). Ensemble model aggregation using a computationally lightweight machine-learning model to forecast ocean waves.

Baselines: Thank you for raising this important point. We agree that baselines are important. To our knowledge, no existing deep learning method is specifically designed for complex regional coastal processes, so we plan to include a 3D U-Net as a baseline in the final version. Due to the large scale of our dataset, we were unable to complete this comparison for the rebuttal. If there are specific baselines you recommend, we welcome your suggestions and will aim to include them.

Physics-aware positional embedding: We have provided the ablation study to show the effectiveness of incorporating the physics information into positional embedding, as shown in Figure 5(b).

Dataset summary table: Thank you for the suggestion. We added the following dataset summary table to clearly describe key properties of the dataset:

Other comments or suggestions: Thanks for your suggestions. We have corrected the typo on Page 4, clarified the notations and descriptions following Equation (5), and added the missing citations in the “Dataset Construction” section on Page 3.

Thank you very much for your valuable review; it is crucial for improving the quality of our manuscript.

Motivation and applications: Our problem setting is motivated by the need to efficiently emulate complex coastal processes for practical applications. High-resolution numerical models such as ROMS are computationally expensive, while deep learning offers a fast and scalable alternative. The dataset includes variables critical for key downstream tasks: for example, storm surge and coastal flood forecasting benefit from predictions of free surface elevation; water quality and stratification modeling rely on temperature and salinity; and sediment transport analysis depends on vertical diffusivity.

Why not use ROMS: The goal is to emulate ROMS, with the primary motivation being computational efficiency—numerical models like ROMS are significantly slower than deep learning. For example, our experiments show that the proposed model reduces the runtime of ROMS for a 72-hour forecast from 2,477 seconds (using 512 CPU cores) to 34.14 seconds on a single A100 GPU, achieving over a 70× speedup. We have added a detailed runtime benchmark to our manuscript.

Use of observational data: We recognize the value of observational data, but they are often limited in spatial and temporal coverage. For instance, RECON provides measurements at sparse, fixed locations and irregular intervals. The ROMS model used in this study was previously validated against RECON observations (Hewageegana et al., 2023), showing strong agreement in key physical processes such as water level variability, currents, salinity, and temperature. This prior validation supports the use of ROMS outputs for training the deep learning model. While our focus is on emulating ROMS rather than evaluating it against observations, we agree that the reliance on ROMS simulations could be considered a limitation. We will clarify this in the main text.

Hewageegana, V. H., Olabarrieta, M., & Gonzalez-Ondina, J. M. (2023). Main Physical Processes Affecting the Residence Times of a Micro-Tidal Estuary. Journal of Marine Science and Engineering, 11(7), 1333.

Expand data to other regions: Expanding the dataset to other regions is non-trivial. First, it requires region-specific input data such as boundary conditions and external forcings. Second, running high-resolution ROMS simulations is computationally intensive. Third, manual calibration by oceanographers and domain experts is essential to ensure simulation quality.

Evaluation procedures: There is no universally accepted standard for determining whether a coastal ocean model is "sufficiently good" for practical use, as acceptable error thresholds depend on the specific downstream application (e.g., storm surge vs. long-term climatology). In our case, the goal is only to approximate ROMS outputs as closely as possible.

Uncertainty: This is a great point. While forecasting inevitably introduces uncertainty, our dataset is based on a ROMS hindcast calibrated using real-world observations. However, we do not currently provide ensemble simulations or uncertainty quantification. Probabilistic models or metrics are not suitable for this scenario.

Future inputs: This is a fair concern. In our setup, the inputs at the target time are assumed to be available from external forecast systems (e.g., global atmospheric forecasts) at coarse resolution. This mirrors real-world practice, where future forcings and boundary conditions are obtained from other (global) forecast models, and regional numerical models take them as input. Making predictions using only initial conditions is not reliable for regional systems due to their sensitivity to future external forcings and boundary conditions. This is actually one of the key motivations for proposing this regional dataset, which differs from global forecasting datasets.

Dataset split: Our dataset is indeed split chronologically. Specifically, the first 8 years of data are used for training, while the 9th and 10th years are used for validation and testing.

Practical information: Thanks for the suggestion. We have added practical dataset details in the paper. The dataset will be hosted on Hugging Face and provided in standard NetCDF (.nc) format. We also include a compressed version in PyTorch (.pth) format (converted from float64 to float16) for efficient use in ML pipelines. The full dataset is approximately 18 TB. To improve usability, we provide a lightweight subset and a base training script to help users quickly run models.

Q1: The current dataset uses manually calibrated ROMS simulations based on observations, which helps ensure general adherence to realism. However, we do not apply data assimilation, and the outputs may not perfectly align with real-world observations.

Q2: Apologies for the mistake — this is not absolute error, but the difference computed as $Label - Prediction$. We have corrected this in the manuscript.