Clarification on input data needed for Aurora Air Pollution

[ehojeily]

Hi there,

First, thank you for the excellent work behind Aurora!

I'm interested in running Aurora Air Pollution and have a question regarding constructing batches. In the tutorial for Aurora Air Pollution (https://microsoft.github.io/aurora/example_cams.html) the previous 12-hour analysis dataset is included in data array used in the batch. The Bodnar et al. (2025) paper states that for some pollutants, Aurora predicts a 24-hour difference (Section B7 in the supporting information). If I'm interested in generating surface PM2.5 forecasts, should the data array used in the batch include the previous 24-hours in order to initialize with the 24-hour difference?

Additionally, if I'm only interested in PM2.5, does Aurora require trace gas (NO, O3, CO, etc.) input from CAMS in order to work properly?

Thanks so much!

[wesselb]

Hey ehojeily (@ehojeily)! Thanks for opening an issue, and thank you for your kind words :). Sorry for the delayed response!

If I'm interested in generating surface PM2.5 forecasts, should the data array used in the batch include the previous 24-hours in order to initialize with the 24-hour difference?

For AuroraAirPollution, the batch should be [t - 24h, t - 12h] (precisely this order) concatenated along the history dimension (see here).

Additionally, if I'm only interested in PM2.5, does Aurora require trace gas (NO, O3, CO, etc.) input from CAMS in order to work properly?

Yes, the model (unfortunately) requires all variables, even if you're interested in just one or two variables.

[ehojeily]

Hi there! No worries, I appreciate the reply. I did experiment with removing some of the trace gases and found some interesting results., but can definitely concur that they are all needed.

Just to clarify, if a forecast is initialized at say, October 10 at 12 UTC, the batch should be [October 9 at 12 UTC, October 9 at 0 UTC]?

Thanks for the help!

[wesselb]

If you want to make a prediction for t+12h, the batch needs to be [t-12h, t]. Thus, if you want to forecast UTC 12, the batch needs to be [UTC 12 previous day, UTC 00].