Wednesday, October 15, 2025, 10:30 - 14:45

SR 02-016/18 Geb. 101

This workshop explores how large language models (LLMs) can be applied to automatic control and engineering in general. It is particularly aimed at researchers from the DFG Priority Programme SPP 2364 - Autonomous processes in particle technology. As a complementary topic, we provide an introduction to normalizing flows and diffusion models in the context of control in the early morning. After a break for late arrivals (in particular the two speakers from TU Braunschweig), the more generic part of the workshop starts at 11:30, and ends after a short concluding discussion at 14:45.

The event will be streamed online via zoom in the channel "syscop public":

https://uni-freiburg.zoom-x.de/j/62791737415?pwd=UDJnbkZlS3NkVm1TSVZLSWxHSktZZz09

Meeting ID: 627 9173 7415
Passcode: syscop2021

Program and Schedule

10:30 An MPC Prior for Diffusion (Jasper Hoffmann, abstract below)

11:15 Break (and welcome of late arrivals)

11:30 Potentials of AI in Process Engineering (Prof. Dr.-Ing. Carsten Schilde und Christoph Thon, TU Braunschweig)

12:30 Lunch Break

13:30 Beyond Narrow AI: Transformer Models and LLMs in Process Engineering (Christoph Thon,TU Braunschweig)

14:15 Concluding Discussion (chaired by Moritz Diehl)

14:45 End

Location

We meet at the technical faculty in Geb. 101, which is at the front of the campus:

Abstracts

An MPC prior for Diffusion (Jasper Hoffmann)

This talk introduces diffusion models through the lens of flow matching, highlighting their advantages as simulation-free alternatives to continuous normalizing flows. After building the mathematical foundation of flows and conditional vector fields, we discuss how diffusion models can be applied to generative modeling tasks and extended toward control problems. In particular, we explore how trajectory distributions can be learned with diffusion and conditioned on states, goals, and constraints. We further discuss applications that use flow matching to imitate MPC planners and conclude by identifying open challenges and avenues for future research at the intersection of diffusion and control.