High-resolution weather forecasts are vital for UK weather prediction.
The UK has very varied terrain, land-surface types and a complex coastline. This means that weather models must run at a resolution that’s high enough to capture these features. Otherwise, the weather phenomena they induce will be smeared out.
Global models, such as the GFS and ECMWF, run at coarser resolutions of 9-25km, or about 0.1 to 0.25 degrees. At this scale, models are unable to resolve fine meteorological texture that is driven by the UK’s varied and complex landscape. Higher resolution models are therefore required.
Operationally, models at a high sub-2km resolution are very computationally demanding, requiring substantial computing infrastructure and energy to run. This is a known downside of normal NWP (Numerical Weather Prediction).
NVIDIA’s Earth-2 provides a set of models capable of learning how to take low-resolution, global weather data and output high-resolution predictions that encapsulate the UK’s complex terrain and land state. This allows for much faster forecast turnarounds and lower operational costs compared to traditional methods.
VWHR-AI (VeryWeather High-Resolution by Artificial Intelligence)
VeryWeather has trained a custom-built model (VWHR) using Earth-2 CorrDiff architecture to predict temperature, wind speed, fog, and precipitation at a high resolution of 1.5km.
It ingests 6.3km ICON-EU data for the next 120 hours, and downscales meteorological variables to a high 1.5km resolution for the UK domain, covering all of England, Wales, Scotland, Northern Ireland and Ireland. That means the grid density of VWHR is 17x that of its parent model (ICON-EU).
The model uses surface states such as terrain, urban density and land use, as well as observation-driven variables such as snow cover and soil temperature at 1.5km resolution. It uses these variables to predict important phenomena such as the urban heat island, terrain-induced wind gusts and orographic precipitation.
The observed computational speed-ups of our CorrDiff implementation compared with traditional NWP methods are significant. VWHR is able to output an hour of weather prediction for the British Isles at 1.5km resolution in 0.5 seconds for the regression stage, and 10 seconds for subsequent diffusion using H100 GPUs (Graphics Processing Units). By comparison, using traditional methods that employ AMD EPYC servers, the same prediction takes about 20-30 minutes to complete. This represents a speed-up of at least 2400x for regression, and 120x for diffusion.
VWHR-R (VWHR Regression): Output from the initial regression stage, covering the British Isles at 1.5km resolution.
VWHR-NW-D (VWHR North West Diffusion): Output from the secondary diffusion stage, covering the North West Midlands domain at 1.5km resolution.
Once trained, AI-based forecasting systems can substantially reduce the computational resources required to generate high-resolution predictions, lowering inference-time infrastructure and energy requirements.
What’s Next
VeryWeather distributes high-resolution predictions from the HARMONIE DMI 2.5km weather model on our snow and wind trackers. Running in-house AI predictions will enable us to fine-tune model parameters for UK severe weather prediction. We intend to use this intelligence to power our trackers with a phased transition over the next 2 years. The model will undergo continuous verification, involving comparison against operational HARMONIE DMI 2.5km and AROME 1.3km to improve model performance. Until this verification is completed, operational use of VWHR remains restricted to non-commercial use only.
Read our technical blog here.