Examples of Aardvark forecasts for the U10 wind component (10-meter eastward wind speed) from the forthcoming paper.
Examples of Aardvark forecasts for the U10 wind component (10-meter eastward wind speed) from the paper.
A deep learning system known as Aardvark Weather offers accurate weather forecasts that are orders of magnitude quicker to generate than existing systems. Described in a Nature article (currently posted as a preprint), the system can generate predictions on four NVIDIA A100 GPUs that would otherwise take roughly 1,000 node-hours on a traditional supercomputer system like the High-Resolution Forecast system (HRES) just to perform the core data assimilation and forecasting steps. For context, the A100 is a high-performance accelerator designed for data centers that launched in 2020, while HRES is the flagship forecasting product of the European Centre for Medium-Range Weather Forecasts (ECMWF), an intergovernmental organization backed by more than 30 European countries.
Breaking down the math in the headline, the Aardvark uses dramatically less processing time:
HRES: 1,000 node-hours = 3,600,000 node-seconds (1,000 hours × 3,600 seconds)
Aardvark: 1 second × 4 GPUs = 4 GPU-seconds
The advance would enable a meteorologist with a PC and cloud-based GPU access to offer weather forecasts that rival those of national weather services. According to the paper, Aardvark Weather is “the first end-to-end data-driven weather forecasting system capable of generating predictions with no input from conventional NWP by instead learning a mapping from raw input observations to output forecasts.” This is significant because it replaces the entire complex chain of traditional forecasting models—data assimilation, numerical forecasting, and post-processing—with a single integrated system.
An order of magnitude fewer observations
The researchers demonstrate that using “an order of magnitude fewer observations than those available to operational baselines and orders of magnitude less compute, Aardvark is capable of producing forecasts on a global 1.50° grid that achieve lower root mean squared error than operational NWP systems across multiple variables and lead times.” This lower error rate—measured through the widely used root mean squared error (RMSE) metric—indicates Aardvark’s predictions more closely match actual weather outcomes than those from traditional systems for several important weather variables.
Aardvark “provides local forecasts that achieve lower errors than post-processed NWP and a full end-to-end operational forecasting system for multiple lead times, and can be optimised end-to-end to maximise performance over variables and regions of interest.” This customization potential could revamp the creation of specialized weather forecasts for different industries and regions. In other words, it could make sophisticated forecasting accessible to areas that currently lack the resources to run conventional systems.
Potential to democratize weather forecasting
Traditional weather forecasting relies on a complex, multi-stage pipeline built over decades. Such methods start with gathering observations from myriad sources—satellites, ground stations, balloons, ships, and planes. These inputs must then be processed and assimilated, blending them with previous forecasts to create an initial atmospheric snapshot. Only then can fluid-dynamic–grounded numerical models simulate how the atmosphere might evolve. Finally, specialists post-process these complex simulations to generate useful local forecasts. Each step demands significant expertise and immense computing power.
Aardvark promises to disrupt this chain, replacing the entire pipeline with just three interconnected neural network modules. Its encoder directly ingests raw observations, skipping the assimilation step and the need for prior forecasts, to generate a gridded initial state of the atmosphere. From this state, the processor module generates global weather predictions, stepping them forward in time. Finally, task-specific decoder modules translate these global predictions into practical local forecasts for specific geographic locations.
Researchers can first pre-train the system on decades of high-quality historical reanalysis data, then fine-tune it using scarcer real-world observations. This tackles the challenge of limited observational records while completely bypassing the need for traditional NWP systems when generating live forecasts.
In their article abstract, the researchers note that an Aardvark could enable the following:
The local station forecasts are skillful up to ten days lead time, competing with a post-processed global NWP baseline and a state-of-the-art end-to-end forecasting system with input from human forecasters.
Despite its performance and efficiency gains, Aardvark Weather, as presented, is a foundational step rather than a finished replacement for operational systems like HRES or GFS. The researchers acknowledge several limitations in the paper: “As with all current AI-NWP systems, Aardvark does not yet run at the resolution of IFS. Further work is required both to increase grid resolution and to produce forecast ensembles through, e.g. diffusion.” They also note challenges with observation data: “Other limitations centre around the use of observations… It is also important to consider how data from new instruments for which there are no training data available can be usefully integrated into the system.” The researchers suggest such shortcomings could be addressed by “training on simulated data” and propose regular fine-tuning with recent data to adapt to instrument changes over time. Despite these challenges, the authors conclude that “Aardvark Weather will be the first of a new generation of end-to-end weather forecasting systems” tackling a diverse range of meteorological tasks.