How accurate is AI weather forecasting compared to traditional methods?

AI weather models like GraphCast have demonstrated skill that matches or exceeds the European Centre for Medium-Range Weather Forecasts (ECMWF) HRES model on over 90% of standard verification targets. For 10-day forecasts, GraphCast shows particular strength in predicting large-scale atmospheric patterns, though traditional NWP models can still outperform AI on certain regional details and rare extreme events where training data is sparse.

What is ERA5 and why is it important for AI weather models?

ERA5 is a global atmospheric reanalysis dataset produced by ECMWF that covers the period from 1940 to the present. It provides hourly estimates of atmospheric variables on a 0.25-degree grid (roughly 31 km resolution) at 37 pressure levels. ERA5 is critical because it provides the consistent, gridded training data that AI weather models need. GraphCast was trained on 39 years of ERA5 data, learning the physical relationships between atmospheric states at consecutive time steps.

Can AI weather models predict extreme events like hurricanes?

AI models are increasingly capable of predicting extreme events. GraphCast demonstrated strong skill in tracking tropical cyclone trajectories, and in several cases outperformed traditional models on hurricane track predictions. However, AI models currently struggle with predicting the intensity of extreme events, particularly rapid intensification of hurricanes, because these events are rare in training data. This is an active area of research, with specialized models being developed for extreme weather prediction.

How fast are AI weather forecasts compared to traditional NWP?

This is where AI models truly shine. A traditional NWP model like ECMWF HRES requires roughly one hour of computation on a supercomputer with thousands of processors to generate a 10-day forecast. GraphCast generates the same 10-day forecast in under 60 seconds on a single TPU or GPU. This speed advantage enables ensemble forecasting (running hundreds of slightly varied predictions to quantify uncertainty) that would be prohibitively expensive with traditional methods.

Will AI replace traditional weather forecasting entirely?

Not in the near term. AI and traditional NWP models are complementary. AI models are excellent for medium-range forecasting (3-10 days) and ensemble generation but currently rely on traditional models for initialization data. Operational weather centers like ECMWF are integrating AI models into their workflows alongside physics-based models, using AI to augment rather than replace existing systems. The future is likely a hybrid approach where AI handles rapid ensemble generation while physics-based models provide the initial conditions and handle scenarios outside the training distribution.

AI Weather Forecasting: GraphCast Beats Traditional Models in 90% of Tests

The Problem with Traditional Weather Prediction

Weather forecasting is one of the most computationally expensive scientific endeavors on Earth. Every day, national meteorological agencies around the world run numerical weather prediction (NWP) models that divide the atmosphere into millions of grid cells, then solve the Navier-Stokes equations of fluid dynamics at each cell for every time step. The European Centre for Medium-Range Weather Forecasts (ECMWF) runs its flagship HRES model on a supercomputer with tens of thousands of processors, and a single 10-day global forecast takes roughly one hour to compute.

This approach has been the backbone of operational weather prediction since the 1950s. It works remarkably well — modern 5-day forecasts are as accurate as 3-day forecasts were 20 years ago. But NWP faces fundamental limitations. The computational cost scales exponentially with resolution, making it difficult to increase grid detail without massive hardware investments. Ensemble forecasting, where you run dozens of slightly varied predictions to quantify uncertainty, is constrained by compute budgets. And the physics-based approach requires hand-tuned parameterizations for processes like cloud formation and convection that occur at scales smaller than the grid.

Note

ECMWF spends over 100 million euros annually on supercomputing infrastructure for weather prediction. The computational demands of traditional NWP are a primary bottleneck for improving forecast accuracy and resolution.

Enter AI Weather Models

In 2022 and 2023, a wave of AI weather models demonstrated that machine learning could match or exceed traditional NWP accuracy while running orders of magnitude faster. Three models stand out: Google DeepMind's GraphCast, Huawei's Pangu-Weather, and NVIDIA's FourCastNet. Each takes a fundamentally different approach to the same problem, but they share a common strategy: learn the dynamics of the atmosphere directly from decades of observational data.

GraphCast

GraphCast uses a graph neural network (GNN) architecture that represents the Earth's atmosphere as a mesh of interconnected nodes. Given two consecutive atmospheric states (6 hours apart), GraphCast predicts the next state by passing messages between nodes on a multi-scale icosahedral mesh. It was trained on 39 years of ERA5 reanalysis data and operates on a 0.25-degree grid with 37 pressure levels — the same resolution as ECMWF HRES. For a deeper dive, see our GraphCast explainer.

Pangu-Weather

Huawei's Pangu-Weather uses a 3D vision transformer architecture rather than a graph network. It treats the atmospheric state as a 3D volume (latitude x longitude x pressure levels) and applies a modified Swin Transformer to learn spatial relationships. Pangu-Weather trains separate models for 1-hour, 3-hour, 6-hour, and 24-hour prediction horizons, then chains them together for longer forecasts. This multi-timescale approach helps it maintain accuracy over extended forecast periods.

FourCastNet

NVIDIA's FourCastNet takes yet another approach, using Adaptive Fourier Neural Operators (AFNOs) that perform learning in the frequency domain. By operating on Fourier transforms of the atmospheric data rather than the raw spatial data, FourCastNet can efficiently capture global-scale patterns while maintaining computational efficiency. It was one of the first AI weather models to demonstrate NWP-competitive accuracy.

How They're Trained: The ERA5 Foundation

All three models rely on ERA5, the fifth-generation atmospheric reanalysis produced by ECMWF. ERA5 combines historical weather observations (from satellites, weather stations, radiosondes, aircraft, and ships) with a physics-based model to produce a globally consistent dataset of atmospheric conditions from 1940 to the present. It provides hourly data at 0.25-degree resolution across 37 pressure levels, covering variables like temperature, wind, humidity, and geopotential height.

Training an AI weather model on ERA5 is conceptually straightforward: given the atmospheric state at time T, predict the state at time T+6h (or T+1h, depending on the model). The model learns the mapping between consecutive states by seeing millions of examples from decades of weather history. What makes this work is that the atmosphere, despite its complexity, follows consistent physical patterns that neural networks can learn to approximate.

Query a weather forecast via SciRouter (coming soon)

import requests

API_KEY = "sk-sci-your-api-key"
BASE = "https://api.scirouter.ai/v1"
HEADERS = {"Authorization": f"Bearer {API_KEY}"}

# Get a 7-day forecast for a location
response = requests.post(f"{BASE}/climate/forecast",
    headers=HEADERS,
    json={
        "latitude": 40.7128,
        "longitude": -74.0060,
        "forecast_days": 7,
        "variables": ["temperature_2m", "precipitation",
                      "wind_speed_10m", "humidity"]
    })

forecast = response.json()
for day in forecast["daily"]:
    print(f"{day['date']}: {day['temp_max']:.1f}C / "
          f"{day['temp_min']:.1f}C, "
          f"Precip: {day['precipitation_mm']:.1f}mm")

Why AI Is Faster: Seconds vs Hours

Traditional NWP models must numerically integrate differential equations forward in time, step by step. Each time step requires solving equations at every grid cell, and the time step size is constrained by numerical stability (the Courant-Friedrichs-Lewy condition). This means a 10-day forecast requires thousands of sequential computation steps.

AI weather models sidestep this entirely. GraphCast generates a 6-hour forecast in a single forward pass through the network. A 10-day forecast requires 40 autoregressive steps (10 days divided by 6-hour increments), and each step takes less than a second on a modern GPU. The total time is under 60 seconds, compared to roughly one hour on a supercomputer for ECMWF HRES.

This speed advantage has profound implications. It makes it practical to run hundreds of ensemble members instead of the 50 that ECMWF typically runs, providing much better uncertainty quantification. It enables real-time reforecasting as new observations come in. And it democratizes weather prediction — a researcher with a single GPU can generate forecasts that previously required supercomputer access.

Tip

The speed of AI weather models does not mean they can replace all of NWP. They still require NWP-generated initial conditions (analysis fields) and cannot yet match NWP for all variables and scenarios. Think of AI models as a powerful complement that excels at rapid ensemble generation and medium-range prediction.

Accuracy Benchmarks

In head-to-head comparisons, GraphCast outperformed ECMWF HRES on over 90% of the 1,380 verification targets in the WeatherBench2 benchmark. These targets cover different atmospheric variables (temperature, wind, humidity, geopotential) at different pressure levels and forecast lead times. The improvements are particularly notable at 5-10 day lead times, where GraphCast's predictions degrade more slowly than HRES.

Pangu-Weather showed similar competitiveness, with particular strength in upper-atmospheric variables. FourCastNet, while slightly less accurate than GraphCast and Pangu-Weather on average, demonstrated impressive efficiency and was one of the first to show that AI could genuinely compete with operational NWP.

GraphCast: Beats ECMWF HRES on 90%+ of WeatherBench2 targets, strongest at 5-10 day range
Pangu-Weather: Competitive with HRES, particularly strong for upper-level geopotential and wind fields
FourCastNet: Slightly lower accuracy but extremely fast, pioneered the Fourier operator approach
GenCast: Google DeepMind's diffusion-based model generates calibrated probabilistic forecasts, outperforming ENS

The Future of Weather Forecasting

Weather agencies are rapidly adopting AI models into operational workflows. ECMWF launched its AI Forecasting Engine (AIFE) to integrate machine learning alongside traditional NWP. National weather services in the US, UK, Japan, and China are all experimenting with AI forecast products. The trend is toward hybrid systems that combine the physical consistency of NWP with the speed and pattern recognition of AI.

Key areas of active development include:

Probabilistic forecasting: Models like GenCast use diffusion architectures to generate full probability distributions rather than single deterministic predictions
Higher resolution: Current AI models operate at 25 km resolution but regional models at 1-5 km are under development
Extreme event prediction: Specialized models for hurricanes, severe convection, and atmospheric rivers where standard training data is insufficient
Longer horizons: Extending skillful forecasts from 10 days toward subseasonal-to-seasonal timescales (2-6 weeks)

Accessing AI Weather Forecasts via API

SciRouter's climate endpoints will bring AI-powered weather forecasting to developers through a simple REST API. Instead of running your own model or parsing complex GRIB files from ECMWF, you can query forecasts programmatically with latitude, longitude, and desired variables. The API handles model inference, data formatting, and caching.

To explore what AI weather forecasting looks like in practice:

GraphCast Explained – deep technical dive into the GNN architecture
Historical Climate Data API – access ERA5, NASA POWER, and NOAA datasets
Climate Change and Your City – use historical data to measure local climate trends

Ready to see AI weather prediction in action? Open the Weather Forecaster Studio or get a free API key to start building with climate data.