Search arXiv⌕ Search

arXiv · 2611.09573

Half-precision inference of AI weather prediction models: negligible loss of skill at a fraction of the cost

Abstract

AI weather prediction models are released with 32-bit (fp32) weights and run in fp32 by default, although current GPUs compute much faster in 16-bit formats. We run five publicly available models -- Aurora, Pangu-Weather, FourCastNet, SFNO and FastNet -- on a single 16 GB GPU from 52 initial dates in 2024, in fp32 and with their networks in half precision (fp16), and verify the forecasts against ERA5. The physical state must stay in fp32: converting the whole model rounds it in physical units, which in fp16 overflows and in bf16 discards meaningful information. With the state, its normalisation and denormalisation in fp32, running the network in fp16 changes the global RMSE by at most 1.4%, and by about 0.5% or less for four of the five models, equivalent to less than one hour of lead time up to day 8 and at most two hours at day 10. The differences show no systematic pattern by region, scale or for tropical and extratropical storms, apart from a small tropical penalty in FourCastNet, and fp16 leaves the CRPS, spread and calibration of Aurora's ensemble almost unchanged. No value overflowed fp16's numeric range (+/-65,504); the largest stay 38 to 544 times below it. A 10-day forecast in fp16 is 1.3-3.5 times faster than in fp32 and uses 1.4-3.7 times less energy. Sensitivity experiments show that rules of thumb from other applications of reduced precision carry over only partly: the first and last layers dominate for some models but not for others, and no type of layer is sensitive on its own. For running AI weather models on a single GPU, fp16 with the state in fp32 is practically lossless at a fraction of the cost. bf16, often preferred for its wider range, is as fast but less accurate than fp16 for every model, and simple 8-bit quantisation costs far more skill.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Stefan Siegert. 2026-11-06. Half-precision inference of AI weather prediction models: negligible loss of skill at a fraction of the cost. https://arxiv.org/abs/2611.09573

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

KEEP EXPLORING

Related papers

Reconciling advective upwelling with diapycnal downwelling: A transient kinematic perspective on Munk's conundrum

Bottom-intensified turbulent fluxes observed near the seafloor drive abyssal densification and diapycnal downwelling through their vertical divergence. This appears paradoxical under steady stratification because dense bottom waters continuously enter from Antarctic sources. Motivated by evolving Indo-Pacific stratification, we use idealized simulations to examine abyssal overturning under transient and steady-state conditions. Using a kinematic rather than thermodynamic water-mass-transformation approach, we diagnose advective overturning, isopycnal volume tendency, and diapycnal transport in two cooling experiments with different mixing intensities, one warming experiment, and one quasi-equilibrium experiment. Advective upwelling occurs in all four, whereas diagnosed diapycnal transport varies in direction. In Type-I cooling, the isopycnal volume tendency exceeds advective overturning in magnitude, yielding downward diapycnal transport; in Type-II cooling, advective overturning exceeds the isopycnal volume tendency, yielding upward diapycnal transport. The warming and quasi-equilibrium experiments also exhibit upward diapycnal transport. These relationships persist qualitatively in potential-temperature coordinates. Type-I cooling also spontaneously produces an observationally documented bottom-intensified buoyancy-flux profile with an e-folding scale comparable to previous estimates. Thus, cooling alone does not determine diapycnal transport direction; it depends kinematically on the relative strengths of advective overturning and isopycnal volume tendency. Within this idealized framework, the results provide a transient kinematic interpretation of the interior-downwelling conundrum: abyssal waters can rise through the basin interior while undergoing diapycnal downwelling, without requiring an upslope return pathway within the bottom boundary layer.

physics.ao-ph↗

RainFlow: Flow Matching for Infrared-Only Quantitative Precipitation Estimation

Geostationary infrared imagers observe precipitating systems every 10 to 15 minutes over large areas, but they sense cloud tops rather than the rain beneath. Their brightness temperatures constrain surface rain only weakly, so retrievals trained by regression return a smooth estimate that misses heavy rain and carries no measure of uncertainty. We treat infrared-only retrieval as conditional generation and introduce RainFlow, a flow-matching model that generates ensembles of rain fields from MSG/SEVIRI infrared observations, in a direct and a residual variant. Evaluated against radar over metropolitan France for a full year, the generated members are realistic rain fields, with the observed dry areas, heavy rain and spatial structure, and the ensemble mean is as accurate as a deterministic baseline. Used as probabilities, the ensembles locate heavy rain at scales where the baseline has no skill, and their spread captures most of the retrieval error, with only slight underdispersion. Trained with a time-sampling density that favours the noise end of the generative trajectory, the two variants reach similar skill. The direct variant depends strongly on this choice of density, whereas the residual variant is nearly insensitive to it. Evaluating the generated rain as a function of brightness temperature shows that the ensembles reproduce how rain depends on it over most of the range. It also isolates the remaining gap, an underestimate of both the amount and the variability of rain under the coldest cloud tops, which scores computed over all pixels do not reveal.

physics.ao-ph↗

New Astronomical, Meteorological and Geological Study of the "Torre Spaccata District" - VII district of the Municipality of Rome

This study, part of the NGICS (New Italian City Geological Study) project, analyzes 25 years of multi-disciplinary data to assess long-term environmental, climatic, and geological trends in Torrespaccata, an urban district within Rome's Seventh Municipality (VII Administrative District). Specifically, it investigates the temporal evolution of Heating Degree Days (HDD) by comparing the historical stationary climate period (1961-1989) with the recent multidecadal interval (1990-2014). Statistical analysis reveals a clear climatic transition: while the pre-1989 period showed a stationary trend (m = -1.14 HDD/year), the modern era exhibits a sharp decline in heating demand, characterized by a negative trend of -11.52 HDD/year (R2 = 0.511). This significant reduction of 80.14 HDD between the two periods is robustly validated via Welch's t-test (t = 3.045, p = 0.0039). Concurrently, interannual variability increased by 29.8%, driven by the synergistic feedback between global warming and the Urban Heat Island (UHI) effect. Furthermore, the correlation between climatic anomalies and local geological events is evaluated using robust statistical imputation to mitigate uncertainties caused by data gaps. These findings provide critical insights into long-term anthropic and natural changes within the urban ecosystem.

physics.ao-ph↗