Physics-constrained inverse neural estimation (PINE) of NOx emission spatial patterns from TROPOMI NO2 columns over the North China Plain: data, model and figure scripts

Data, trained model, inference code, figure data and figure scripts for Wang, Pan and Lyu, Physics-constrained inverse neural estimation (PINE) of NOx emission spatial patterns from TROPOMI NO2 columns over the North China Plain, Atmospheric Chemistry and Physics (preprint egusphere-2026-3233). PINE-NOx is a physics-constrained autoencoder: a neural encoder (U-Net with a CBAM bottleneck) maps daily Sentinel-5P/TROPOMI tropospheric NO2 columns, ERA5 10 m wind and the EDGAR v8.1 NOx prior to emissions, and a fixed, differentiable transport and chemical-decay operator maps the emissions back to columns. The network is trained without emission labels. The observations constrain only the spatial pattern; the domain annual total is set equal to EDGAR v8.1 by one constant, so the total is a design constraint and not an independent estimate. Domain: 36-41 N, 113-120 E, 0.1 deg; year 2023. Emission fields (NetCDF, kg m-2 s-1, NOx as NO2): 350 daily fields (296 training and 54 validation days; 15 days with fewer than 50 valid TROPOMI cells are absent), which are effective emissions at the TROPOMI overpass time and not 24 h means; 12 monthly fields (monthly PINE-NOx pattern scaled to the EDGAR v8.1 monthly domain totals); one annual field (mean of the 350 daily fields, 0.025 % below the EDGAR annual total). Monthly and annual files include the 1-sigma uncertainty of ln(emission share), method-and-data term only; the prior-inventory term of the paper's budget is not included because it is derived from MEIC. Also included: trained network weights with inference code (bit-identical reproduction of the daily fields on the tested GPU, within 1.6e-5 relative on CPU); the gridded inputs (TROPOMI NO2 at 0.1 deg with qa_value at least 0.75 and cloud radiance fraction below 0.5, ERA5 daytime 10 m wind, EDGAR v8.1 regridded area-conservatively); the data and scripts for all figures and Tables 2 and S2 of the paper. Figures 8 and 10 additionally need MEIC 2023, which is not redistributed. Training code, model-variant runs, WRF-Chem inputs and raw output, raw TROPOMI and ERA5 files, MEIC fields and raw CNEMC data are not included. Version 3 (2026): TROPOMI inputs re-gridded with corrected quality screening; all emission fields, model weights and figure data replaced; this version supersedes versions 1 and 2. Contains modified Copernicus Sentinel data (2023). Generated using Copernicus Climate Change Service information (2023). EDGAR v8.1 (European Commission JRC, CC BY 4.0), WRI Global Power Plant Database (CC BY 4.0), NOAA ETOPO and Natural Earth data are used as described in README.txt.