SVDROM Grid Search

This project performs grid search experiments for training and evaluating Optimized Dynamic Mode Decomposition (OptDMD) models on climate data (the publicly available ERA5 dataset). The SVD-ROM package is employed for performing OptDMD. The main workflow is controlled by train.py, with configuration specified in params.yaml.

Main Files

train.py

• Trains OptDMD models for all combinations of parameters defined in params.yaml.
• Computes reconstruction and forecast RMSE for each model.
• Saves metrics and (optionally) models to disk.
• Uses Dask for parallel computation.
• Requires input SVD and scaler (containing the time average) files tracked by DVC.

params.yaml

• Configuration file for grid search parameters.
• Key sections:
  • dask: Number of threads for parallel processing.
  • train: List of mode counts, Hankel preprocessing options, trial settings for parallel bagging.
  • reconstruct/forecast: Time ranges for evaluation.
  • ins: Paths to input data and models.
  • outs: Output directories for models and metrics.
  • misc: Additional settings (e.g., pressure level of the dataset).

Usage

1. Ensure all input files (SVD, scaler, groundtruth) are available and tracked by DVC.

2. Adjust parameters in params.yaml as needed.

3. Run the training script from the command line:

   python train.py

4. Results (RMSE metrics and optionally models) are saved in the specified output directories.

Results

Results are presented in the notebooks evaluation_winter_2020.ipynb and evaluation_summer_2020.ipynb, where an evaluation of 45 day DMD forecasts of winter and summer 2020 are presented.

Requirements

• Python 3.10+
• Dask
• xarray
• numpy
• ruamel.yaml
• python-box
• svdrom
• DVC (for data tracking)

Install dependencies:

pip install -r requirements.txt

Folder Structure

• input_data/: Input datasets(tracked by DVC).
• metrics/: Output RMSE metrics.
• models/: Saved DMD models (if enabled).

Notes

• Pressure level consistency is checked between params.yaml and DVC metadata.
• If the Dask dashboard link printed at runtime doesn't work, try replacing the hostname with localhost in the URL.

---

For more details, see train.py and params.yaml.