There is a silent revolution going on in the weather world powered by AI. European Centre for Medium-Range Weather Forecasts has introduced the Artificial Intelligence Forecasting System (AIFS), a data-driven ensemble weather forecasting model, similar to those developed by Google, Huawei & NVIDIA. What's the big deal?👇
First some context.
Weather forecasting models can be classified into two: deterministic and ensemble.
Deterministic models provide a single forecast based on a given set of initial conditions of the atmosphere, for a specific location and time - essentially the 'best forecast' within the laws of physics. These are great for short-to-medium-term forecasts (7 days).
While deterministic models can be really accurate due to the potential for high-resolution forecasts in the best-case scenario, they can also be way off especially due to the range of possible states of the atmosphere and the uncertainties associated with 'guessing' them.
Enter ensemble models.
Ensembles run many simulations accounting for all the uncertainty in the initial conditions of the atmosphere and, hence provide several possible results. These are great for medium-to-long-term forecasts (up to 2 weeks) esp. by comparing the results.
Ensemble models (like AIFS) are inherently probabilistic - if 90% of the results predict heavy rain, there is high confidence in the forecast. Similarly, even if only 10% of the results predict a storm, it can facilitate early warning systems and support disaster management.
Back to the ECMWF news, why is this a big deal?
First, seeing a public-sector organisation keep up with the private sector advancements and continuously innovate is great. This is fundamentally important as I believe weather is a public good (free/open) and should always be.
Second, it is amazing that the model and its source code will be open-source. This means that anyone in the weather community can use the AIFS model and identify strengths and weaknesses so that the forecasts become better and better.
Finally, AI and weather forecasting are truly a match made in the atmosphere. With a given set of data and boundary conditions, the AI model is capable of generating hundreds to thousands of simulations which can help us forecast that disaster, that we may have missed before.
This is important especially for climate adaptation plans, given the rise in frequency and/or intensity of natural disasters. I am excited about the possibility of having better early warning systems in the global south, which suffers more due to the lack of Wx infrastructure.
Check out the announcement from ECMWF and the accompanying paper in the comments.
For more demystification of Earth observation and its applications like weather, follow me and subscribe to the TerraWatch Space newsletter.
PS. I had to simplify (and sometimes oversimplify) the concepts to make the narrative simple enough to be digestible. Feel free to add more information to complement this.