AI in Weather Forecasting: Google’s GraphCast

Weather forecasting has relied on massive supercomputers and complex physics equations for decades. Now, Google DeepMind is shifting that paradigm with GraphCast. This artificial intelligence model has demonstrated the ability to predict global weather patterns more accurately and significantly faster than the world’s leading traditional systems.

The Shift from Supercomputers to Neural Networks

Traditional weather forecasting is known as Numerical Weather Prediction (NWP). It works by feeding current atmospheric conditions into massive computer simulations that solve physics equations regarding fluid dynamics and thermodynamics. While effective, this method is incredibly computationally expensive. It requires hundreds or thousands of CPUs running for hours to generate a 10-day forecast.

GraphCast takes a different approach. It is a machine learning model based on graph neural networks (GNNs). Instead of solving physics equations, it learns from patterns in historical data. Specifically, Google DeepMind trained GraphCast on four decades of historical weather data from the European Centre for Medium-Range Weather Forecasts (ECMWF), known as the ERA5 dataset.

By recognizing patterns in how weather systems evolve over time, the AI can predict future states without needing to “understand” the underlying physics in the traditional sense.

Speed and Efficiency Breakdown

The most striking advantage of GraphCast is its speed. The efficiency metrics released by Google DeepMind and published in the journal Science highlight a massive gap between AI and traditional computation:

• Processing Time: A traditional 10-day forecast using the gold-standard HRES (High Resolution Forecast) system can take hours on a supercomputer cluster. GraphCast can generate the same 10-day forecast in under 60 seconds.
• Hardware Requirements: While NWP requires a supercomputer, GraphCast runs on a single Google Cloud TPU v4 device.
• Energy Consumption: Because it runs on a single chip for only a minute, the energy required for a GraphCast prediction is roughly 1,000 times less than traditional methods.

Accuracy Comparison: GraphCast vs. HRES

Speed is useless without accuracy. To validate the model, researchers compared GraphCast directly against the HRES system operated by the ECMWF. The ECMWF is widely regarded as producing the most accurate global weather forecasts in the world.

The results showed that GraphCast is not just a faster alternative, but a more precise one.

• Global Variables: GraphCast outperformed HRES on more than 90% of the 1,380 test variables tracked. These variables include temperature, pressure, wind speed, and humidity at various altitudes.
• Troposphere Predictions: When focusing specifically on the troposphere (the lowest layer of the atmosphere where most weather occurs), GraphCast outperformed HRES on 99.7% of the test variables.
• Resolution: The model operates at a high resolution of 0.25 degrees longitude/latitude, which corresponds to roughly 28km x 28km at the equator. It tracks 37 different vertical levels of the atmosphere.

Real-World Case Study: Hurricane Lee

Theoretical metrics are important, but performance during extreme weather events is vital for public safety. Google DeepMind highlighted the performance of GraphCast during Hurricane Lee in September 2023.

Predicting the landfall of a cyclone is notoriously difficult. Nine days before Hurricane Lee made landfall in Nova Scotia, GraphCast correctly predicted that it would hit the region. Traditional forecasts at that same 9-day mark were inconclusive regarding the storm’s path and timing.

This ability to identify severe weather events earlier gives governments and emergency services more time to prepare evacuations and resource allocation. GraphCast also shows high proficiency in predicting:

• Atmospheric Rivers: Narrow corridors of concentrated moisture that can cause severe flooding.
• Extreme Temperatures: Identifying heat waves or cold snaps well in advance.

Limitations and the Hybrid Future

Despite the success of GraphCast, meteorologists are not retiring their supercomputers yet. There are specific limitations to AI-based forecasting that keep traditional physics-based models relevant.

First, AI models are bound by their training data. If the climate shifts drastically into territory not seen in the historical ERA5 dataset (such as unprecedented temperature extremes caused by accelerating climate change), the AI might struggle to predict the outcome accurately. Physics-based models, however, are grounded in the fundamental laws of nature, which remain constant regardless of historical precedent.

Second, GraphCast is currently a deterministic model. It provides a single best-guess forecast. Meteorologists often rely on “ensemble” forecasts, where they run a model 50 times with slight variations to see a range of probabilities (e.g., a 30% chance of rain vs. a 70% chance). While AI can be adapted for ensemble forecasting, this is an area where traditional methods currently have established workflows.

The immediate future of meteorology is likely a hybrid approach. Agencies like the ECMWF are already experimenting with their own AI models (like the AIFS). We will likely see AI used to generate rapid, low-energy “first look” forecasts, while supercomputers continue to run physics simulations to verify severe weather threats and provide detailed ensemble data.

Open Source and Collaboration

In a move to accelerate development in this field, Google DeepMind has open-sourced the code for GraphCast. This allows weather agencies, researchers, and other scientists to inspect the model, run it on their own hardware, and retrain it with new data.

By making the code available, Google is allowing the global scientific community to audit the tool and integrate it into existing weather service pipelines. This transparency is crucial for building trust among meteorologists who are accustomed to understanding exactly why a model produces a specific prediction.

Frequently Asked Questions

Is GraphCast available to the public? The code for GraphCast is open source and available on GitHub. However, running it requires access to specific historical data (ERA5) and hardware capable of running the model (like a GPU or TPU). It is not currently a consumer-facing app like a weather widget on your phone.

Does GraphCast replace meteorologists? No. GraphCast is a tool that assists meteorologists. It provides faster data and potentially earlier warnings, but human experts are still needed to interpret the data, communicate risks to the public, and make decisions based on the forecasts.

How far in advance can GraphCast predict weather? GraphCast is designed for medium-range forecasting. It specializes in predicting weather conditions up to 10 days in advance.

Who developed the data GraphCast uses? GraphCast was trained on the ERA5 dataset, which is produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). This dataset is a reanalysis of global climate data spanning over 40 years.