Can AI models be trusted to predict heavy rainfall?

"My project is about making AI weather predictions more trustworthy, especially in the case of heavy rainfall predictions", Robin Guillaume-Castel says.

Just over a year ago he arrived in Bergen as part of the first group of postdoctoral fellows in the LEAD AI programme. His project is already yielding promising results.

AI in weather predictions

Today, neural networks, a type of artificial intelligence (AI), are increasingly used for weather predictions. According to Guillaume-Castel these can be very accurate, but they mostly behave like black boxes: they give a prediction without clearly showing why.

"My work focuses on checking whether these AI models base their decisions on real physical weather processes, rather than on accidental patterns in the data. This is an important and topical topic today because neural networks are starting to be used operationally in weather forecasting", Guillaume-Castel says, adding:

"For example, Bris is a forecasting model based on neural networks that should be added soon to Yr weather predictions by the Norwegian Meteorological Institute. If we want to rely on them, for example for extreme weather alerts, we need to be confident that they are making physically meaningful decisions".

From large scale to local events

Guillaume-Castel's previous work focused on the causes of global climate change, which is an important but very large-scale and somewhat abstract topic.

"In this project, I enjoy working with individual, concrete and local weather events such as heavy rainfall that people directly experience. I am also motivated by the challenge of working with neural networks, which is a new field for me. I find it especially important to explain how these neural networks work. As they are becoming increasingly more common in weather and climate science, I think that understanding and interpreting these models is becoming just as crucial as improving their accuracy".

Promising early results

One year into his project, the results are already promising:

"I started with a case study of heavy rainfall in Western Norway, a region strongly affected by North Atlantic storms. I wanted to test if a neural network was able to learn about the physical processes governing heavy rainfall here. I found that a relatively simple kind of neural network was able to correctly identify North Atlantic storms and use this information to predict heavy rainfall in Western Norway several days in advance", he says.  

"This is promising, as it shows that neural networks can make predictions for the right physical reasons, and that we can verify and interpret their behaviour".

Providing valuable insights

Guillaume-Castel believes his project could contribute to important knowledge in two main ways:

"First, by helping to establish benchmarks that show when and how AI models can be trusted in weather and climate applications. This would add to the usual benchmarks assessing the pure predictive skill of these models", he says; "and second, by using similar AI techniques as tools for scientific discovery. By studying why neural networks make their decisions, we could uncover new insights about the physical processes behind extreme weather, ultimately improving research and knowledge".

A community of AI researchers

For the climate scientist, one of the main challenges so far has been moving into a new research field.

"More specifically, I did not have a strong background in AI and neural networks at the beginning of the project, but using these methods was crucial for my research questions", he explains.

"To address this, I spent time learning the fundamentals and discussing with the growing group of machine learning scientists at the Geophysical Institute and the Bjerknes Centre for Climate Research. Additionally, the LEAD AI cohort has been especially helpful, in the sense that it provides a community of researchers with other AI interests and expertise. This environment has been a valuable source of feedback, support, and potential collaboration".