An open smelter-resolved spatial prior for CF4 (PFC-14) emissions

An open, reproducible spatial prior for CF4 (PFC-14) emissions, gridded by primary-aluminium smelter location as a physically-grounded input for atmospheric inversions. It targets the gap that EDGAR and GAINS grid F-gas emissions by a population / built-up proxy, which misplaces a gas actually emitted at a finite, mappable set of smelters (anode-effect PFCs). The smelter registry also serves as a documented aluminium-location layer for sectoral attribution cross-checks (e.g. it carries no primary smelters in South Korea or Taiwan). Method: (1) a registry of operating (2020 status) primary-aluminium smelters with location and nameplate capacity, from public sources (GEM gem.wiki, IAI, USGS MCS, Wikipedia "List of aluminium smelters"): 94 smelters, Europe-complete; China is covered through its 12 largest smelter clusters out of ~120 plants (disclosed); (2) three weighting variants: capacity-weighted, presence-only (robustness), and production-rescaled (country totals matched to USGS 2020 national primary-aluminium production; the recommended variant for global use, since the raw registry under-weights China at ~28% of weight vs ~57% of world production); (3) rasterized as a relative (normalized) spatial weight, to be rescaled to the inversion total before use. The values must never be read as absolute emissions. The field covers the aluminium sector alone (~60–80% of global CF4); combine with separate fields for electronics/other sources rather than using it as a total-CF4 prior. Validation (Europe, ICOS PARIS CF4 posterior, 2020, 6-member flat-prior ensemble, with block-bootstrap and spatial-permutation significance tests): the smelter prior beats EDGAR's population/built-up proxy where the inversion resolves point sources, as in Iceland (0.25 vs ~0 in spatial correlation; statistically significant; RHIME members only), and in the pooled smelter-country aggregate (0.057 vs 0.015), where the direction is consistent across all 6 ensemble members but not significant per member against spatial autocorrelation. The observation-driven posterior places the 22 in-domain smelter cells at a mean 60th percentile while EDGAR places them at the 26th (zero CF4 at 14 of 22 smelter cells). At the 1-degree grid of the distributed files: pooled 0.14 vs 0.07; Iceland 0.55 vs ~0. The win is relative and concentrated in well-constrained theatres; absolute fine-scale skill is low for all priors, and EDGAR's smooth field tracks the low-emission background better (rank correlation). Direction is consistent with Kim et al. (2021). Contents: gridded prior (NetCDF; capacity, presence-only and production-rescaled variants; European 0.1° and global 1° extents, plus CF-convention drop-in files on the egusphere-2025-5656 deposited 1° grid), the open smelter registry (CSV with per-row provenance), a README file manifest, source documentation, and a one-page method note. Code: https://github.com/zachdissington/climate-solutions-research-kit