Vertical line in prediction outputs

[thisistaimur]

Upon inspection, it seems like the vertical line in the predictions appears always at the Prime Meridian, i.e. 0 degrees longitude. See figure below for reference.

In geospatial or climate models like Aurora over the entire globe:

• Longitude wraps around: 179.9° → -180° is just one step
• But Convolutional layers treat input as rectangular grids with edges → introducing artificial edge effects at 0°/360° unless we explicitly fix this

This means:

• our prediction tensor is still spatially continuous but incorrectly wraps at 0°.

• Likely → model outputs garbage at the “roll boundary” (because input lon was rolled, but model’s convolutional logic isn’t aware of it).

Core Issue:

Aurora’s UNet-like models (especially Conv layers) assume:

• cyclic or periodic behavior across longitude.

If you roll your input data to align longitudes, but do not:

• apply circular padding in the model
• or train the model on rolled data properly
  we get seams at roll edges (like 0° longitude line).

Debug Pipeline:

1. Roll lon grid to 0-360
2. Roll inputs (species_distribution)
3. Forward pass
4. Roll prediction back

[thisistaimur]

Aurora expects 0-360 degree longitudes, and 90 to -90 degree latitudes. Latitudes do have edges so no problem, but longitudes don't have edges but convolutions are treating them as rectangular tiles without padding and padding_made ="circular".

This info can be found in the batch.metadata documentation: https://microsoft.github.io/aurora/batch.html#batch-metadata

See screenshot: