The variables needed for machine learning are different than a standard air quality model. It's actually a much shorter list because you don't have to numerically model atmospheric motion/chemistry/deposition.
Most AQMs have a spatial domain that has horizontal and vertical depth. However, the machine learning forecasts I've seen are only for point locations where meteorological and air quality variables are measured. That is to say, the machine learning model is trained using variables from a ground station, and is only valid for that location. You can train multiple locations using similar variables, but each station would have it's own set of data to train with.
I've seen machine learning forecasts from Washington State University for ozone and PM2.5, and the list of important meteorological variables are basically the same.
For ozone they say:
Hourly data for six meteorological variables (temperature and relative
humidity at 2m, wind components u & v, planetary boundary layer height, and sea level pressure) from 4km WRF archives, time information (month, weekday and hour), and the previous day’s observed moving 8-hour averaged O3 mixing ratios are used to train the RF classifier models.
For more info you can see the latest presentation given.