1st Associate Fellow at the KIT Depart- ment of Economics and Management

In 2021, the KIT Associate Fellow concept had its 10th anniversary and, fittingly, the 10th of the eleven KIT departments assigned the status for the first time. The idea behind the concept is to allow scientifically independent, not yet habilitated, young group leaders to supervise their doctoral students and participate in their examination procedures.

Sebastian Lerch started his independent group with funding from the Vector Stiftung. After having been successful within their "MINT für die Umwelt" scheme, in 2021, he also became KIT Associate Fellow. The Vector Stiftung had specifically asked for Lerch to get the right to take part in the doctoral proceedings of his group members. Encouragingly, more and more funding agencies take proactive measures to secure the independence of the research groups they invest in.

Lerch works on improving probabilistic weather forecasts using artificial intelligence. So, how is he the first Associate Fellow at the KIT Department of Economics and Management? While he initially worked in the math department at the Institute for Stochastics, he is now to be found at the Institute of Economics. “At KIT, as it is common in Germany, research in statistics happens at multiple departments,” Lerch explains. “More theoretical approaches are often closer to mathematics; more applied research might be associated with economics.” As his research combines both sides in about equal measures, Lerch feels at home at both KIT departments. His research lives and strives through interdisciplinarity. Besides, research at the intersection of mathematics and statistics, meteorology, physics, and computer science plays a crucial role on the way to a next generation of probabilistic weather forecasts.

Have you ever consulted 5 weather apps and gotten 5 different forecasts? What makes the weather so hard to predict, is the chaotic system we call atmosphere: physical variables like temperature, humidity, and cloud coverage are constantly changing and interacting with each other. Nonetheless, physical models attempt to replicate and predict what is happening in the atmosphere. Nowadays, forecasts are based on repeated model runs generating a collection of different forecast scenarios. These ensembles provide a likely description of future weather events and have become widely used in practice. However, they are often characterized by systematic errors that require correction via statistical post-processing methods.

The work of Sebastian Lerch’s group starts, where established statistical methods reach their limits in the face of large data sets and complex interrelationships. His team uses neural networks to develop novel post-processing methods. The advantage: neural networks can find non-linear relationships in complex datasets.

“We don’t want to replace physical models as the core element for weather prediction”, Lerch continues. “Instead, we incorporate neural networks as tools to provide improved predictions, for example, by correcting systematic errors due to the coarse spatial resolution of the physical models. So far, we have been successful at significantly improving over established techniques from statistics.” With his group, he now plans to develop novel approaches to incorporate new sources of information into models and to address challenges in the use in operations.

Many of the underlying methodological questions and developments are not only relevant for probabilistic weather prediction but will also provide insights into the use of neural networks in other areas of application such as energy forecasting, economics, and epidemiology. This is where the strong interdisciplinary research environment at the Institute of Economics and KIT in general offers fruitful opportunities for collaboration.