Satellite inversion of North China Plain NOx emissions now runs in milliseconds: PINE-NOx lifts observation-reconstruction correlation from 0.332 to 0.856
Satellite nitrogen dioxide columns constrain where nitrogen oxides are emitted, but conventional inversions are costly or depend on existing emission products. PINE-NOx maps Sentinel-5P/TROPOMI nitrogen dioxide columns to emissions with a neural encoder, then renders them back through a fixed, differentiable two-dimensional transport-chemistry decoder, trained without any emission labels. On 54 validation days used only to select the training epoch, emissions from the inversion raise the log-space spatial correlation between simulated and observed columns from 0.332 for the EDGAR inventory to 0.856; this is a reconstruction of the observations, not emission accuracy, and the annual total is normalized to EDGAR while the observations adjust the daily relative pattern.