Artificial intelligence (AI) is revolutionising weather forecasting by matching the speed and accuracy of traditional models. This breakthrough development has the potential to transform how we predict weather patterns.
Advancements in weather forecasting have allowed us to anticipate severe storms and heatwaves, but they come at a high cost. Supercomputers running constantly result in billions of dollars in expenses.
Now, AI is leading a new wave of numerical weather prediction. Trained AI systems can generate 10-day forecasts that are just as accurate, if not better, than traditional models. The European Centre for Medium-Range Weather Forecasts (ECMWF) is embracing this technology and conducting its own experimental AI forecasts. This exciting development allows for more frequent predictions and frees up computing resources for other important tasks.
Tech giants like Google DeepMind and Huawei are competing to create the most precise AI weather models. Google's GraphCast model and Huawei's Pangu-Weather model are at the forefront of this race. Google has even developed an AI model that surpasses most traditional weather agencies in making 24-hour predictions. The progress made in such a short time is remarkable.
Traditional weather models rely on complex equations and immense computational power, while AI models use "deep learning" techniques. By analyzing 40 years of ECMWF data, AI models can identify patterns in the atmosphere's natural evolution. Google's GraphCast model has proven its superiority by outperforming ECMWF's forecasts up to 10 days in advance, even accurately predicting hurricane paths and extreme temperatures.
These promising results have been tested in realistic settings. ECMWF researchers evaluated Huawei's Pangu model using limited observations similar to their operational weather model. Surprisingly, Pangu demonstrated comparable forecasting skills, with slight differences in predicting rainfall and fine-scale features. This showcases the potential of AI models to enhance weather predictions even further.
The next challenge for AI researchers is to adapt generative AI techniques for ensemble forecasting. This technique captures uncertainty by running a model multiple times to generate a range of possible outcomes. By applying these techniques, AI models can improve their ability to predict extreme events like intense hurricanes, which are currently underestimated in intensity.
To achieve further advancements, AI models could be trained directly on raw observation data held by weather agencies. Google's short-term weather model already utilises data from weather stations, radar, and satellites for training.
AI models also have the potential to assist in developing high-resolution climate models that run on ultrafast computers. Once these models generate sufficient output for AI training, they can accelerate climate simulations by emulating them 100 times faster.
Although traditional forecasts will still be used for now, AI is becoming a valuable complement. Concerns about the black-box nature of AI, where researchers struggle to explain how these systems reach their conclusions, may slow down adoption. However, traditional models are already complex and opaque in nature.
Ultimately, the future of weather forecasting will depend on user preferences. Whether it's a more accurate forecast or one derived from physical equations, the choice will depend on the needs and preferences of various stakeholders, including farmers in the field.