The reality of AI in weather and climate science

🔑 Enhanced Key Takeaways

•AI models significantly reduce the computational resources and time required for global weather forecasts, enabling predictions in seconds or minutes on single GPUs, compared to hours on supercomputers for traditional numerical weather prediction (NWP).
•For many standard atmospheric variables and lead times up to 10 days, leading AI models like GraphCast and Pangu-Weather have demonstrated accuracy competitive with or superior to the best traditional numerical weather prediction systems for general conditions.
•Despite advancements in global forecasting, AI models currently underperform traditional high-resolution NWP models in forecasting intense, localized precipitation events and fine-scale atmospheric phenomena, often underestimating their intensity and frequency.
•Most AI weather models are trained on reanalysis datasets (e.g., ERA5), which are themselves partly derived from traditional NWP, raising questions about true independent skill and the 'black box' nature of AI models limits scientific understanding and interpretability.
•Recent developments include the emergence of open-source AI weather model frameworks (e.g., NVIDIA's Earth-2 Atlas) and the deployment of AI-native satellite constellations (e.g., Tomorrow.io's DeepSky) designed to provide the high-frequency observational data needed for advanced AI forecasting.

📊 Competitor Analysis▸ Show

Feature / Model	Google DeepMind GraphCast	Huawei Pangu-Weather	NVIDIA FourCastNet3	ECMWF AIFS	Traditional NWP (e.g., ECMWF HRES)
Architecture	Graph Neural Network (GNN)	3D Earth System Transformer	Spherical Neural Operator (Convolutional NN)	Graph-based architecture	Physics-based equations (fluid dynamics, thermodynamics)
Training Data	39-40 years of ERA5 reanalysis data	43 years of ERA5 reanalysis data	ERA5 reanalysis data	ERA5 reanalysis data	Observational data + data assimilation
Speed (Global 10-day forecast)	Under 1 minute on a single machine/TPU	10 seconds on a single GPU server	Under 4 minutes for 60-day rollout on single H100 GPU	Operational in 2024 (faster than NWP)	~6 hours on supercomputers
Accuracy (Medium-Range)	Outperforms HRES on 90% of metrics	Outperforms NWP methods in accuracy	Matches leading ML models, exceeds IFS-ENS	Competitive with traditional NWP	Gold standard for many years
Extreme Event Prediction	Struggles with unprecedented events	Struggles with unprecedented events	Struggles with unprecedented events	Struggles with unprecedented events	Generally outperforms AI for record-breaking events
Operational Status	Used by Google DeepMind	Publicly accessible on ECMWF website since late July 2023	NVIDIA Earth-2 platform	Operational since 2024	Fully operational for decades (e.g., ECMWF IFS)

🛠️ Technical Deep Dive

Google DeepMind GraphCast: Employs a Graph Neural Network (GNN) architecture. It was trained on 39-40 years of ERA5 reanalysis data (1979–2017) at a 0.25-degree horizontal resolution across 37 pressure levels.
Huawei Pangu-Weather: Utilizes a 3D Earth System Transformer (3DEST) architecture. It was trained on 43 years of ERA5 data (1979–2021), using 13 pressure levels with 5 variables (temperature, humidity, geopotential, u/v wind components) and 4 surface variables (2-meter temperature, u/v 10-meter wind components, and mean sea level pressure).
NVIDIA FourCastNet3 (FCN3): Employs a fully convolutional, spherical neural operator architecture, based on local and spectral spherical convolutions parameterized by Morlet wavelets. It is designed for spherical geometry and trained on ERA5 reanalysis data, capable of large-scale training on up to 1,024 GPUs.
ECMWF Artificial Intelligence Forecasting System (AIFS): Uses a graph-based architecture and was trained on ERA5 reanalysis data.
Google DeepMind GenCast: Uses a diffusion model architecture to produce probabilistic ensemble forecasts, generating multiple plausible future atmospheric states from a single learned distribution.

🔮 Future ImplicationsAI analysis grounded in cited sources

Hybrid AI-physics models will become the dominant paradigm in operational weather forecasting.

This approach leverages AI's speed and pattern recognition with physics models' interpretability and ability to handle unprecedented events, as evidenced by major meteorological agencies adopting hybrid systems.

Development of AI models will increasingly focus on improving the prediction of 'gray swan' extreme weather events.

Current AI models struggle with out-of-distribution events, and climate change is increasing the frequency and intensity of such events, necessitating new approaches like physics-informed learning or synthetic data generation.

The role of human meteorologists will evolve towards interpreting AI outputs, communicating uncertainty, and specializing in complex, localized phenomena.

AI handles routine forecasting rapidly, freeing meteorologists to focus on high-impact communication, technology curation, and refining forecasts for specific, challenging scenarios.

⏳ Timeline

1922

Lewis Fry Richardson publishes 'Weather Prediction by Numerical Process,' conceptualizing numerical weather prediction (NWP).

1950

Jule Charney and colleagues run the first digital weather forecast on the ENIAC computer.

1980s

AI techniques, including expert systems and early neural networks, begin to be explored in weather forecasting.

2023-07

Huawei's Pangu-Weather model is published in Nature, demonstrating superior accuracy and speed over traditional NWP for medium-range forecasts.

2023-11

Google DeepMind's GraphCast model is published in Science, showing it outperforms ECMWF's HRES on many meteorological metrics.

2024

The European Centre for Medium-Range Weather Forecasts (ECMWF) makes its Artificial Intelligence Forecasting System (AIFS) operational.

The reality of AI in weather and climate science

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (32)

👉Related Updates