Air quality is not thought enough of in terms of localized data. Some modeling works at larger scales, but if you want to know the forecast for other pollutants (especially ones that disperse or transition) or the source of methane leaks, you need very localized data from many modalities (wind direction, temperature, topographic, seasonal, traffic, etc).

I worked at https://aclima.io on air quality for 6.5 years. My role was managing backend data pipelines, but I worked with scientists and data scientists who were pushing the boundaries of models' capabilities. Models are complex and expensive - any advancement here, like Graphcast, is very important. [1] One job our team was responsible for is to reduce the cost of high quality data, so we drove vehicles around to collect very localized data, which ended up being temporally sparse. Modeling can fill gaps to some measurable level of certainty.

It should also be said that policy is far behind the science, but the burden will remain on science and data to continue make conclusions irrefutable.

[1] - https://deepmind.google/discover/blog/graphcast-ai-model-for...