Physics-constrained inverse neural estimation (PINE) of daily NOx emissions from TROPOMI NO2 columns over the North China Plain — data archive (2023)

Data and model-product archive accompanying the manuscript "PINE: physics-constrained inverse neural estimation of daily NOx emissions from TROPOMI NO2 columns over the North China Plain" (submitted to Atmospheric Chemistry and Physics). PINE-NOx is a physics-constrained autoencoder (CBAM-UNet neural encoder + a fixed, differentiable advection-decay transport-chemistry decoder) trained label-free against Sentinel-5P/TROPOMI tropospheric NO2 columns with an EDGAR v8.1 (FT2022) reference prior. Domain: North China Plain (36-41N, 113-120E), 0.1deg; year 2023 (364 days; 2023-10-15 absent due to insufficient TROPOMI coverage).Contents (data only): (01) CF-1.10 daily inventory NetCDF; (02) 364 per-day E_final flux maps + CV uncertainty fields; (03) seasonal four-fold cross-validation model weights; (04) WRF-Chem July-2023 forward-validation fields + validation statistics; (05) figure intermediate data (four-backbone ablation, 80-day spatial-correlation validation, Fig.1 basemaps, EDGAR v8.1 prior on the 0.1deg grid, gridded daily TROPOMI columns). A self-contained verify.py reproduces all headline numbers from the files.Verified key numbers: annual total = EDGAR prior = 3.14 Tg (inventory 3.13; c=0.1777); log-space spatial correlation vs TROPOMI on 80 season-balanced days 0.395 (EDGAR) -> 0.837 (PINE-NOx; dr=+0.442, p<0.001), robust across four encoder backbones (0.823-0.837) and four seasons; annual-mean grid log-correlation r=0.941; emission winter/summer ratio ~1.4; four-fold CV mean r=0.918 (uncertainty ~12-15%).Third-party input data are cited at source and not redistributed here: TROPOMI/S5P (Copernicus; only processed 0.1deg gridded intermediates are included), EDGAR v8.1 (EU JRC), ERA5 (ECMWF/C3S), ground NO2 (CNEMC), and the MEIC inventory (meicmodel.org.cn; comparison only, not redistributed). The PINE-NOx source code and figure scripts are released separately.