Ammonia inversions no longer need 60 chemistry runs: EnsAI builds the ensemble 3300 times faster
Synopsis
Ammonia emissions inversions rely on ensembles to carry meteorology-dependent error covariances, but each ensemble means repeatedly running the GEM-MACH chemistry model. EnsAI replaces that with a U-Net that maps ammonia emissions perturbations plus temperature and wind fields directly to surface concentration perturbations, producing a week of simulated data per member in 7 seconds against roughly 6.5 hours for GEM-MACH on 720 nodes. Inversions using the EnsAI ensemble gave annual total emissions 2.6% lower than the GEM-MACH ensemble inversions, with monthly totals within 10% for every month.
Interpretation
Ensemble generation for surface ammonia moves from repeated runs of a chemical transport model to a single neural-network forward pass, so the flow-dependent error covariances an inversion needs can be obtained quickly on a GPU. Such inversions could previously be run only occasionally because generating the ensemble with GEM-MACH was expensive, and static background error covariances are not uncommon in chemical data assimilation systems; at Environment and Climate Change Canada, both operational systems that produce chemical analyses use static background error covariances. For a week of simulated data per ensemble member, GEM-MACH took about 6.5 hours of wall time on a CPU cluster using 720 computing nodes, while EnsAI took 7 seconds on a single GPU, about 3300 times faster.
The network learns only a diagnostic map from emissions perturbations to surface ammonia concentration perturbations, taking the emissions perturbation together with temperature and horizontal wind at the same time, with no fields from earlier time steps. This exploits the short atmospheric lifetime of ammonia, whose perturbations are local and short-lived, so no latent space and denoising step of the kind used in diffusion generative models is needed. Training used the GEM-MACH ensemble for January to December 2016 with 60 members at hourly resolution, 527040 samples; test and validation used the 2015 ensemble and July 2014.
Substituting the ensemble inside the emissions inversion gives increments close to those obtained with the GEM-MACH ensemble while using far less computation. Compared with inversions using static covariances, the EnsAI inversions stay closer to the GEM-MACH ensemble inversions. Monthly total emissions from EnsAI inversions were within 10% of the GEM-MACH ensemble inversions for all months, and the annual total was 2.6% lower; the static covariance inversions were 6.9% lower annually and exceeded 10% monthly differences from March to May and August to November.
Perspective
The demonstrated setting is the North American domain and a single variable, surface ammonia concentration, with training from the 2016 GEM-MACH 60-member ensemble and evaluation on the 2015 and July 2014 ensembles plus AMoN and NAPS surface observations. The author states that until further tests in different regions, EnsAI in its present form should be considered restricted to its current spatial domain.
The author notes that training data come from one year over North America and may not include drastically different meteorological conditions or chemical regimes, so continued training or fine-tuning would be needed if future conditions change significantly. EnsAI outputs a single surface concentration field, whereas GEM-MACH produces hundreds of chemical fields on 85 levels; extending it to longer-lived species would require fields from previous time steps and a prognostic relationship. Against surface observations, 38 stations had higher RMSE for EnsAI than for the GEM-MACH ensemble inversion, with 0.16 ppbv the largest increase.