Station-trained air-quality forecasts without regional emission inventories
Abstract
Air-quality forecasts rely on regional chemical transport models (CTMs) and emission inventories, which are uncertain in many regions and costly to update. Here we present SOLACE, whose inner forecast is trained on monitoring-station observations alone and evolves gridded concentration fields through explicit transport, layer exchange, sources and sinks. With 140,000 trained parameters and no regional emission inventory, it forecasts six pollutants for 48 h on a 4-km grid; one forecast takes 4 s on a CPU. Trained independently in five regions on three continents and evaluated in 2024 at stations withheld from training and initialization, it had smaller errors than an observation-corrected global CTM forecast (persistence for PM10) in 57 of 60 region-species-window comparisons, with day-1 PM2.5 errors 32-77% smaller. With operational forecast meteorology in Korea, the same weights lost 2-12% and kept their lead; SOLACE now issues daily forecasts in all five regions. Exchanging learned processes between regions showed that swapping emission and net change together reproduces 73-97% of the loss from foreign weights, whereas learned transport details mattered little; models retrained without them kept skill in two regions. The regional information a forecast needs resides largely in the effective source-sink budget, and station observations suffice to learn it.
Related
- Learning joint probabilistic weather forecasts from station observations alone
- Improving global precipitation forecasts with an AI weather model trained on satellite observations
- Evaluating global emission inventories of biogenic bromocarbons
- Regional economic impacts and emission responses under solar radiation modification
- Reconciliation Through Integration: A Source-Level Bayesian Framework for Developing Measurement-Integrated Methane Inventories
BibTeX
@misc{lee2026stationtrained,
title = {{Station-trained air-quality forecasts without regional emission inventories}},
author = {TK Lee},
howpublished = {EarthArXiv},
year = {2026},
month = oct,
doi = {10.31223/x54z4t},
url = {https://eartharxiv.org/repository/view/15887/},
note = {CC BY-NC 4.0},
}
RIS
TY - UNPB
TI - Station-trained air-quality forecasts without regional emission inventories
AU - TK Lee
PY - 2026
DA - 2026/10/31
PB - EarthArXiv
DO - 10.31223/x54z4t
UR - https://eartharxiv.org/repository/view/15887/
AB - Air-quality forecasts rely on regional chemical transport models (CTMs) and emission inventories, which are uncertain in many regions and costly to update. Here we present SOLACE, whose inner forecast is trained on monitoring-station observations alone and evolves gridded concentration fields through explicit transport, layer exchange, sources and sinks. With 140,000 trained parameters and no regional emission inventory, it forecasts six pollutants for 48 h on a 4-km grid; one forecast takes 4 s on a CPU. Trained independently in five regions on three continents and evaluated in 2024 at stations withheld from training and initialization, it had smaller errors than an observation-corrected global CTM forecast (persistence for PM10) in 57 of 60 region-species-window comparisons, with day-1 PM2.5 errors 32-77% smaller. With operational forecast meteorology in Korea, the same weights lost 2-12% and kept their lead; SOLACE now issues daily forecasts in all five regions. Exchanging learned processes between regions showed that swapping emission and net change together reproduces 73-97% of the loss from foreign weights, whereas learned transport details mattered little; models retrained without them kept skill in two regions. The regional information a forecast needs resides largely in the effective source-sink budget, and station observations suffice to learn it.
ER -