Datasets:

polymathic-ai
/

planetswe

language:
  - en
license: cc-by-4.0
tags:
  - physics
task_categories:
  - time-series-forecasting
  - other
task_ids:
  - multivariate-time-series-forecasting

This Dataset is part of The Well Collection.

How To Load from HuggingFace Hub

Be sure to have the_well installed (pip install the_well)
Use the WellDataModule to retrieve data as follows:

from the_well.data import WellDataModule

# The following line may take a couple of minutes to instantiate the datamodule
datamodule = WellDataModule(
    "hf://datasets/polymathic-ai/",
    "planetswe",
)
train_dataloader = datamodule.train_dataloader()

for batch in dataloader:
    # Process training batch
    ...

PlanetSWE

One line description of the data: Forced hyperviscous rotating shallow water on a sphere with earth-like topography and daily/annual periodic forcings.

Longer description of the data: The shallow water equations are fundamentally a 2D approximation of a 3D flow in the case where horizontal length scales are significantly longer than vertical length scales. They are derived from depth-integrating the incompressible Navier-Stokes equations. The integrated dimension then only remains in the equation as a variable describing the height of the pressure surface above the flow. These equations have long been used as a simpler approximation of the primitive equations in atmospheric modeling of a single pressure level, most famously in the Williamson test problems. This scenario can be seen as similar to Williamson Problem 7 as we derive initial conditions from the hPa 500 pressure level in ERA5. These are then simulated with realistic topography and two levels of periodicity.

Associated paper: Paper.

Domain expert: Michael McCabe, Polymathic AI.

Code or software used to generate the data: Dedalus, adapted from this example.

Equation:

$\begin{align*} \frac{ \partial \vec{u}}{\partial t} &= - \vec{u} \cdot \nabla u - g \nabla h - \nu \nabla^4 \vec{u} - 2\Omega \times \vec{u} \\ \frac{ \partial h }{\partial t} &= -H \nabla \cdot \vec{u} - \nabla \cdot (h\vec{u}) - \nu \nabla^4h + F \end{align*}$

with $h$ the deviation of pressure surface height from the mean, $H$ the mean height, $\vec{u}$ the 2D velocity, $\Omega$ the Coriolis parameter, and F the forcing which is defined:

def find_center(t):
    time_of_day = t / day
    time_of_year = t / year
    max_declination = .4 # Truncated from estimate of earth's solar decline
    lon_center = time_of_day*2*np.pi # Rescale sin to 0-1 then scale to np.pi
    lat_center = np.sin(time_of_year*2*np.pi)*max_declination
    lon_anti = np.pi + lon_center  #2*np.((np.sin(-time_of_day*2*np.pi)+1) / 2)*pi
    return lon_center, lat_center, lon_anti, lat_center

def season_day_forcing(phi, theta, t, h_f0):
    phi_c, theta_c, phi_a, theta_a = find_center(t)
    sigma = np.pi/2
    coefficients = np.cos(phi - phi_c) * np.exp(-(theta-theta_c)**2 / sigma**2)
    forcing = h_f0 * coefficients
    return forcing

Visualization:

Dataset	FNO	TFNO	Unet	CNextU-net
`planetswe`	0.1727	$\mathbf{0.0853}$	0.3620	0.3724

Table: VRMSE metrics on test sets (lower is better). Best results are shown in bold. VRMSE is scaled such that predicting the mean value of the target field results in a score of 1.

About the data

Dimension of discretized data: 3024 timesteps of 256x512 images with "day" defined as 24 steps and "year" defined as 1008 in model time.

Fields available in the data: height (scalar field), velocity (vector field).

Number of trajectories: 40 trajectories of 3 model years.

Estimated size of the ensemble of all simulations: 185.8 GB.

Grid type: Equiangular grid, polar coordinates.

Initial conditions: Sampled from hPa 500 level of ERA5, filtered for stable initialization and burned-in for half a simulation year.

Boundary conditions: Spherical.

Simulation time-step ( $\Delta t$ ): CFL-based step size with safety factor of 0.4.

Data are stored separated by ( $\delta t$ ): 1 hour in simulation time units.

Total time range ( $t_{min}$ to $t_{max}$ ): $t_{min} = 0$ , $t_{max} = 3024$ .

Spatial domain size: $\phi \in [0, 2 \pi]$ , $\theta \in [0, \pi]$ .

Set of coefficients or non-dimensional parameters evaluated: $ν \nu$ normalized to mode 224.

Approximate time to generate the data: 45 minutes using 64 icelake cores for one simulation.

Hardware used to generate the data: 64 Icelake CPU cores.

What is interesting and challenging about the data:

Spherical geometry and planet-like topography and forcing make for a proxy for real-world atmospheric dynamics where true dynamics are known. The dataset has annual and daily periodicity forcing models to either process a sufficient context length to learn these patterns or to be explicitly time aware. Furthermore, the system becomes stable making this a good system for exploring long run stability of models.

Please cite the associated paper if you use this data in your research:

@article{mccabe2023towards,
  title={Towards stability of autoregressive neural operators},
  author={McCabe, Michael and Harrington, Peter and Subramanian, Shashank and Brown, Jed},
  journal={arXiv preprint arXiv:2306.10619},
  year={2023}
}