For best experience please turn on javascript and use a modern browser!
You are using a browser that is no longer supported by Microsoft. Please upgrade your browser. The site may not present itself correctly if you continue browsing.
What role does technology play in the work of scientists at the University of Amsterdam? In this series, we will discuss this with researchers from the Faculty of Science. This time, we talked to Erwin Luesink, a postdoc in the Stochastics research group. Among other things, he is looking for ways to calculate the probability that a weather forecast will actually be correct.
Erwin Luesink
Erwin Luesink

“They predicted sunshine, but now it’s pouring! How is that possible?” Even with the most advanced technologies, the weather is not always predicted correctly. Why is predicting the weather so difficult?

‘Weather forecasts use one of the most complex computer models imaginable,’ says Erwin Luesink, a postdoc at the UvA Korteweg-de Vries Institute for Mathematics. This complex model has a complicated mathematical equation for which there is not yet an exact solution, and processes such as cloud formation and rain are added to this.

Luesink: ‘With a weather forecast, the model is already flawed, because you cannot represent it neatly and miss the details.’ This means that weather forecasts can't always be correct. But what are the chances that, for example, your weather app will give the correct prediction?

The chance that it’s correct

‘In a forecast, you want to add the chance that your prediction is correct, such as “tomorrow there is a 70% chance of rain”,’ explains Luesink. There are various methods to determine such a chance.’ Meteorologists often look at data from the past and compare it with the prediction of their model. They can already derive a chance from this, but it remains a challenge to determine this accurately.

This is precisely one of the applications that Luesink is working on. He is trying to calculate the reliability of prediction models more accurately by combining two disciplines, stochastic analysis and differential geometry. Luesink: ‘I use a specific mathematical technique that can be applied to many different predictions.’

Adding noise

This technique involves adding noise (or small changes) to the data of the prediction model. Then, he lets the model make a prediction based on the data with the noise. By repeating this several times, he gets a kind of “spaghetti” of different predictions, or possible outcomes. The real outcome lies somewhere in that spaghetti.

By looking at how close those outcomes are to the truth, Luesink can assess how reliable the prediction was. Luesink: ‘If all the outcomes are close to each other and to the truth, the prediction is accurate. But if the outcomes differ widely, the prediction is less reliable.’

Graph showing the results of various weather forecasts with noise.
“Spaghetti” of different outcomes of a weather model with noise. The colored bundles become wider over time (t), which means that the uncertainty is increasing. Image: Erwin Luesink

Luesink mainly investigates how noise can be added to the physics models. ‘The model is based on laws of physics, while noise is not physics. How do you make sure that you create a physics model that can deal with that noise, without suddenly losing all the nice bits of physics? And how do you do this effectively on a computer? That is what we are currently investigating.’

No more pen and paper

Technology plays a major role in Luesink’s research. ‘Many people see mathematicians as people who do calculations with pen and paper in their room. But nowadays, mathematics is a lot of computer work.’ By working on a computer, mathematicians can immediately implement their theories. This is important, because an approach can sometimes be more suitable for the theory, and sometimes for the practicality.

For example, Luesink is looking at different programming languages for his model, such as the relatively new language Julia, which is very convenient to use for certain mathematical calculations.

He’s also investigating which computers are most suitable for executing the model: can it be done on a laptop, or is a supercomputer required? Luesink: ‘We want to keep our methods up-to-date with the developments of supercomputers, because ultimately supercomputers are indispensable for weather forecasting.’

Heat map of wind on a rotating planet
Example of a simulation performed with a supercomputer: wind on a rotating planet. Image: Erwin Luesink.

Sustainable choices

Weather forecasting requires so much computing power and data that they can't be performed on smaller computers. But when developing new models, there are often multiple ways to perform calculations, with one requiring more energy than the other. According to Luesink, mathematicians are also increasingly aware of this.  

Luesink: ‘The danger of using machine learning, for example, is that every calculation requires enormous amounts of energy. Given the climate problems, we are becoming increasingly aware of this. Suppose you can solve a problem very efficiently with mathematical techniques, but not completely perfectly. If you can solve it perfectly with machine learning, but that costs the energy of half a village, then maybe you should not use it.’