‘One major achievement that could be realized by EXPECT is to better characterize our uncertainty in future climate.’ After studying physical geography in Brussels and doing a PhD in glaciology, Vincent Verjans joined the Barcelona Supercomputer Center as Recognised Researcher to work on EXPECT. Before that, he lived in South Korea, working on how wildfires affect climate.
At BSC, Vincent is now part of the Climate Variability and Change group, where he spends his time writing computer code to analyse climate data. His work involves extracting patterns and statistics from large datasets, which are then used to make predictions about how the climate might change in the future.
Challenges of predicting the future
How can we test predictions about a future that hasn’t happened yet?
‘We are pretty good at predicting future trends, but we are still struggling to predict specific events that occur on short time scales or small spatial scales. The main goal of EXPECT is to really improve the understanding of which features of the climate system will influence future extreme events, future regional trends’, or future local changes,’ says Vincent.
A novel approach to climate modelling
‘I’m kind of taking a different approach of what has been done so far,’ says Vincent. ‘I want to compare how a statistical technique can complement predictions from climate models.’ Unlike traditional climate models, which simulate the full climate system based on physics, his approach relies on identifying patterns in data, working on a statistical model to predict sea surface temperatures, an important driver of weather patterns like rainfall and heatwaves.
Combining physical laws and data
Climate models are built on the laws of physics, like how air moves through the atmosphere or how ocean currents flow. Even with that solid foundation, scientists don’t fully understand every part of the climate system. Some processes are difficult to describe with equations alone, which is why large datasets and real-world observations are essential. ‘There are still many aspects that are very difficult to constrain with physical equations. Our models typically have a resolution of 20-50 kilometres when we run them on the global scale and we can get very good simulations of atmospheric circulation of these scales,’ says Vincent. However, at such resolution, models miss smaller, more localised interactions, like how wind interacts with forests or how soil moisture moves between soil and air. These small-scale processes are crucial because they affect cloud formation and rainfall, but they’re too fine-grained for the models to simulate directly.
Instead, researchers use data to understand how these small-scale interactions typically behave under certain larger-scale conditions. This way, they can include those effects in the models, even if they can’t represent them in full detail.
Cross-validation: testing model accuracy
‘Since we can’t compare predictions with the actual future, scientists use a method called cross-validation to test how accurate their models are,’ says Vincent.
This involves removing a portion of existing data, building the model without it, and then using the model to predict that missing part. These predictions are called hindcasts, in analogy to forecasts for the past.
For instance, if temperature records span from 1950 to 2024, researchers might leave out the last five years, build a model using data only up to 2019, and then see how well the model predicts the 2019–2024 temperatures. Then scientists compare these predictions to the actual recorded temperatures from those years.
This process is repeated with different parts of the data removed each time, helping scientists assess the model’s reliability and improve its accuracy. Cross-validation provides a fair test of how well the model performs, almost like simulating a real prediction into the future.
Want to learn more about regional climate changes and the work done in EXPECT?
We’re on LinkedIn, X, and Bluesky. Follow us there if you haven’t already!