arXiv Science⌕ Search
Home/Archive/arXiv

arXiv · 2610.03736

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

Ophélia Miralles
View original source ↗Read source document ↗Export metadata ↓

Abstract

Probabilistic downscaling must represent kilometer-scale structure left unresolved by a deterministic regional prediction. We introduce an analog-initialized latent transport method that uses historical regional residuals as a meteorologically informed empirical source. A frozen graph neural network predicts the deterministic state from the Integrated Forecasting System (IFS). For each input, residuals from 20 training dates with similar IFS conditions initialize the ensemble. A conditional flow-matching model transports these fields in a carefully designed 65-factor, 100,000-node latent space, and a grid decoder reconstructs 81 atmospheric channels including precipitation at 2.5km spacing. On a season- and cycle-balanced 2025 test, ensemble CRPS improves over deterministic mean absolute error for all 81 variables. Matched analogs improve point and spatial errors over Gaussian initialization and reduce spectral error compared with covariance matching. Comparisons with released CorrDiff, a local CorrDiff-mini model and full grid EDM provide context on the 2023 validation sample. Total training of the downscaling model takes approximately 12h on one NVIDIA H200, and inference takes about 12s per 81-field member. The frozen networks also generate fields over France when supplied with local static data and a high-resolution IFS-derived analog bank. The method combines empirical initialization with whole-domain latent transport for computationally practical regional ensembles.

Explore related subjects

physics.ao-phcs.AIcs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Ophélia Miralles. 2026-09-15. Efficient Analog-Initialized Latent Transport for Spatially Coherent Probabilistic Downscaling from Global to Kilometer Scales. https://arxiv.org/abs/2610.03736

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Graph-Learned Spring SST Evolution and Land-Ocean Synergy Improve Yangtze Summer Rainfall Prediction

Seasonal prediction of Yangtze summer rainfall is challenged by nonstationary oceanic precursors. Here we demonstrate that accounting for the evolution of spring SST and state-dependent land surface information improves predictability. Graph total variation (GTV) is applied to March-May SST fields over the western Pacific, Niño3.4, and tropical Indian Ocean to identify spatially coherent, month-specific SST coefficients for each grid cell and month to form a scalar GTV index. The learned coefficient patterns capture April-to-May ENSO decay and an opposing Indian Ocean adjustment, consistent with coupled Indo-Pacific variability. During 2008-2022, the GTV index prediction outperforms Niño3.4-based and sparse-regression benchmarks. Further improvement is achieved by including interactions of May Indo-China Peninsula soil moisture with Indian Ocean SST, importantly also correcting several underestimated wet summers. Circulation diagnostics associate the learned index with a summer geopotential-height dipole and Yangtze moisture convergence, while dry-Indochina/warm-Indian-Ocean conditions coincide with stronger westward extension of the western Pacific subtropical high.

physics.ao-ph↗
arXiv2026

Atmospheric Chemistry-Climate Modeling with an AI Foundation Model

Atmospheric composition and chemistry-climate interactions are central to predicting future climate change. Yet simulating chemical and aerosol processes is typically the most expensive component of an atmospheric model and is often highly simplified or prescribed entirely. AI weather forecasting foundation models have shown skill rivalling process-based models and are increasingly fine-tuned for new atmospheric tasks. Here, we present the first foundation model to capture interactions between emissions, atmospheric composition, and radiation. We fine-tune the Aurora Earth system foundation model on simulations from the chemistry-climate model UKCA. Our model reproduces daily meteorology, composition, aerosol, and radiation fields of the coupled system with 23 of 27 predicted fields achieving $R^2$ above 0.85 in both the troposphere and stratosphere. It reproduces ozone, CO, and temperature with annual mean biases below 4% and the downwelling shortwave and longwave fluxes to within 6%. Our model remains stable in autoregressive simulation over 50 years while running ~$100\times$ faster than UKCA. By training on paired emissions scenarios, our model learns responses to perturbations by capturing tropospheric ozone's response to changes in NO and CO emissions and surface radiation's response to aerosol. We further show that incorporating MLS satellite ozone retrievals into training corrects biases in the underlying UKCA fields, reducing zonal mean ozone error against collocated observations by 23% between 250 and 20 hPa. Our approach does not suggest process-based modeling can be replaced but demonstrates that a foundation model can be adapted to respond to forcings and learn from observations at a relatively small computational cost.

physics.ao-ph↗
arXiv2026

High-level, efficient ocean modeling at all scales with Oceananigans

We describe the user interface, governing equations, and numerical methods underpinning the ocean modeling software called "Oceananigans". Oceananigans development has been led by the Climate Modeling Alliance as part of a larger effort to build a trainable climate model with quantifiable uncertainty. Notably, Oceananigans is written in the Julia programming language, is optimized for GPUs, and can efficiently simulate all scales of ocean motion ranging from millimeter-scale turbulence in a small box to planetary-scale ocean circulation. We argue that this combination of capabilities is enabled by a design that combines a basic, easy-to-optimize, GPU-friendly structured finite volume algorithm exposed behind a high-level, programmable user interface. This design negotiates a dual mandate for highest-possible performance (to support state-of-the-art applications) and enhanced accessibility (to facilitate adoption and development). The dual mandate aims ultimately to accelerate the progress of Earth system science.

physics.ao-ph↗