How can chaos be predicted?
Early in the morning on the 26th of December 1999, the people of France were shocked by two fatal storms. How did these massive cyclones take the meteorologists by surprise? And can machine learning aid weather forecasts of the future?
Main content
Natacha Galmiche was only four years old when the cyclone hit her country. She doesn’t remember it happening, but she remembers the aftermath.
“For years afterwards, we would be driving in our car, and I would see large swaths of knocked over trees.”
She is now doing her PhD in informatics but is still interested in how nature works.
“I was always interested in things like applied maths and physics, but it seemed like everyone doing that wanted to work with finance, which I wasn’t interested in.”
As a part of her master’s degree, she ended up going to the Nansen Center for two internships, hoping to do something with an environmental focus.
“We were assessing plankton in the sea and using the data to study the overall health of the sea. It was so exciting. I got to use my computer science skills on something I cared about. It made me certain that I had chosen the right field of study.”
Data in the clouds
While Natacha didn’t return to the Nansen center, she did return to Bergen, doing her PhD at both the Department of Informatics and the Geophysical Institute at UiB.
To explain what Natacha is doing, you will need a basic understanding of how weather forecasts are made…
Data about the weather is constantly collected from weather stations, satellites and even weather balloons. This data is used to create 50 simulated scenarios. These scenarios are combined to make the forecast presented to the public.
If all the data was a hundred percent perfect, you would – in theory – only need one simulation.
“The atmosphere is both stable and chaotic”, Natacha explains.
“If you drop a ball on top of a mountain, it can roll down either side of the mountain and end up in completely different places.”
In other words, a tiny change in the starting perimeters can cause a completely different outcome.
“A cloud approaching Bergen from the coast can either end up hitting the mountains and causing a downpour in the city, or the whole cloud might be swept away, leaving a beautiful cloud-free day in the city.”
So why did the meteorologists get it so wrong in 1999?
When you are making a weather forecast, it would be impossible to see every scenario for every specific location. Instead, meteorologists largely have to rely on averages.
So in 1999, 42 hours before cyclone Martin hit France, the forecast for atmospheric pressure looked like this:
Simplified map of France showing the expected atmospheric pressure over France.
Maybe a bit windy in the north, but all in all there was no cause for alarm. But this was just the mean.
When looking at all 50 models of the potential weather, you can see how the ball might roll down the other side of the mountain:
These maps show the 50 potential weather outcomes as predicted 42 hours before the storm.
So why don’t the meteorologists just look at all the different models?
“It’s a matter of time. You can look at a single location and study every scenario over time. But if you’re making a weather forecast for a whole country, you don’t have the time to do that.”
Natacha is working on creating machine learning models that can spot these concerning outliers, so called “clusters”, so they can be brought to the attention of the weather forecasters; and to get them to trust the result.
“Meteorology has been around for thousands of years and is very well studied and understood. The machine learning model must be interactive, so meteorologists can tweak it based on their instinct or experience”, explains Natacha.
“This is also important so they can see how the model interprets the data – so they know it’s not just making something up.”
This technology is likely to become more important in the future, as extreme weather is becoming more common and harder to predict.
Weather forecasts are based on historical data about the climate, but as the climate is changing, forecasting is becoming more chaotic.
“I hope this can be a tool for spotting unprecedented variations in forecasting. The final decision about extreme weather warnings is always going to be made by a meteorologist, but I hope my machine learning model can assist as their job gets harder.”
