Efficient Analog-Initialized Latent Transport for Spatially Coherent Probabilistic Downscaling from Global to Kilometer Scales

We introduce an efficient probabilistic downscaler that generates 20-member ensembles of 81 atmospheric fields on a million-point, 2.5-km grid. Historical residual fields from similar global weather states initialize each member, and a conditional flow adjusts them in a compressed spatial representation before full-domain decoding. On 96 held-out Nordic cases from 2025, ensemble continuous ranked probability score (CRPS) improves over deterministic mean absolute error for all 81 fields. A compute-matched 32-case atmospheric source comparison using separately trained transports shows lower pointwise and spatial errors with matched analogs than with Gaussian initialization. On the 50-case 2023 validation sample, member-wise 2-m temperature and 10-m wind-component RMSE is about 12–16% lower than for NVIDIA's released CorrDiff checkpoint. Training the complete selected system, including its deterministic predictor and decoding stage, requires about 12 hours on one H200 GPU. The same frozen networks generate all 81 fields over France using new coordinates, static data and an IFS-derived historical bank.