How is AI changing weather forecasting?

The European Geosciences Union (EGU) is the leading organization for Earth, planetary and space science research in Europe and around 15,000 scientists attend its annual conference. Professor Christensen shared a summary of her talk with us.

As Hannah explains, in the last three – four years we’ve seen big changes in the field of weather forecasting in the form of new AI weather forecast models. Traditionally, weather forecasts have been made by taking physical understanding and encoding it into a computer model. Effectively, the laws of Physics are deep in this computer model, which is used to predict how the state of the weather will change over the coming hours. Change came in late 2022, when several purely statistical weather forecasting models were published. These come from machine learning algorithms that take in the current state of the atmosphere and tell you what the weather will be like over the coming days and weeks. They do this remarkably well.

This innovation has caused huge shockwaves in the community, raising the question of how do we deal with these models? They don’t explicitly encode any Physics, and instead learn atmospheric behaviour from vast amounts of historical data. This can make people a bit nervous but, for Hannah, it’s also very exciting because they are much quicker and cheaper than the physics-based models. For example, a model called Pangu is around 10,000 times cheaper and quicker to run than a numerical model.

Given these high-quality weather forecasts, the next question is, what about climate? Making a climate prediction uses similar physics ideas to making a weather forecast. However, instead of predicting changing weather patterns for a week or two, when Hannah and her team are predicting the climate, they run their model for 10 years, or for 100 years, with changing levels of carbon dioxide and other greenhouse gases in the atmosphere. The team is less interested in exactly what the weather will be like in 100 years’ time. Instead, they are focusing on how the statistics of weather change.

People have started building machine-learnt atmospheric models that can be used for climate timescales as well. This is an area where it’s hard to predict how the field will evolve. As with the machine-learnt weather forecasting models, there can be quite a lot of hype, so it is important to test these models extensively to see exactly what they can and can’t do.

In her talk at EGU26, Hannah explained how her team broke new ground by taking a machine-learnt atmosphere model and coupling it with a physical ocean model. This was technically challenging, but it meant that they could see how this atmosphere and ocean evolve in time together, which is essential for realistic climate simulations. An example of this is El Nino. This natural pattern of ocean temperature in the Pacific Ocean affects how much snowfall you get over North America in winter as well as how much rainfall you get in other places around the world. So, it is important to take El Nino into account when making a prediction about the seasonal weather outlook. For the first time, Hannah and her colleagues coupled an atmospheric AI model with an ocean model to see if these real-world phenomena can be predicted.

Hannah’s research straddles the two. She believes that there is a lot of merit in the physics-based models and they are not going to go away. One reasons for this is that we’re asking the AI-based climate models to do something that we know AI is not good at, which is to generalise. Or, in other words, you train your machine-learnt model based on the current climate and we’re asking it to extrapolate some way into a future climate. It is important to test these things, but physics-based models give greater confidence when exploring changing climate conditions, as we know that the laws of Physics are the same now as in the future.

AI models produce predictions quite quickly and accurately. We asked Hannah, does their accuracy tail off after a certain time period? She explained that visually, if you look at the output from a ML (machine-learning) model around 10 days in the future, you can see it is a very smooth field. This reflects the way many machine-learning models account for uncertainty. As a result, you end up with something that will do well according to error metrics, but which is less physical than the original model you had. There is another generation of ML models that tries to combat this but ultimately, if you have just one prediction from an ML model, it will have some form of smoothing in it, which you wouldn't get from a physical model. A single physics-based forecast, in contrast, represents one possible physically-consistent future.

Yes. For example, a physical model will have sharper weather fronts, and it will maintain these sharp fronts as it forecasts further into the future, even if their exact position is slightly wrong. This can be valuable for human forecasters interpreting the model output: they can say, OK, it looks quite certain that there’s a frontal system coming in, but we know that there's going to be an error in the position or the timing. However, in an ML model, this uncertainty is reflected as a much smoother field, making it hard to discern fronts. The user’s role would be interpret this, and so to work with AI and the old and new models to achieve the best result going forward.

AI is already transforming weather forecasting, making high-quality forecasts much faster and less computationally expensive than before. Rather than replacing physics-based models, Hannah believes the greatest advances will come from combining the strengths of both approaches. By integrating AI with our understanding of the physical climate system, Hannah and her team hope to build faster, more capable models for predicting weather, seasonal extremes, and future climate change.

Hannah’s article on an ML atmosphere coupled to a physical ocean is under review at Journal of Advances in Modeling Earth Systems, and is available as a preprint here: https://arxiv.org/html/2603.28704v1