Hamburg VISTA - Virtual Initiative for Science & Technology in AI

Thursday 27 Feb 2025, 10:00 → 13:00 Europe/Berlin

ESA W-221 (Universität Hamburg)

Gregor Kasieczka (Universität Hamburg)

Welcome to the second event of Hamburg VISTA - Virtual initiative for Science & Technology in AI!

 

Prof. Luisa Lucie-Smith of UHH and Dr. Sebastian Götschel of TUHH will offer exciting insights into their research on ML in physics. Following the talks, participants will have an opportunity to meet and network over lunch.

 

Participation (including lunch) is free, but registration is required. Registration deadline for in-person attendance: 20.02.2025. Remote (zoom-only) registration will be possible until the end of the event.

 

---

 

The event will take place at the UHH main campus, please see the following map and photos for directions:

 

 

10:00 Arrival and registration

With coffee and cookies

1 Welcome and introduction to Hamburg VISTA

Speaker: Gregor Kasieczka (Universität Hamburg)

VISTA_Welcome_2025.pdf

Invited talks

2 Cosmology in the deep learning era

Machine learning has significantly improved the way cosmologists model and interpret cosmological data; yet, its "black box" nature often limits our ability to trust and understand its results. In this talk, I will discuss how we can apply deep learning techniques, including explainable AI methods, to tackle complex problems in cosmology. I will show how we can gain new insights into the formation of cosmological structures on small scales such as dark matter halos and galaxies. I will also show how we can retrieve the independent degrees of freedom to which cosmological data is sensitive, thus potentially highlighting the origin of current cosmological tensions.

Speaker: Prof. Luisa Lucie-Smith (UHH)

3 Data Meets Physics: Physics-Informed Machine Learning and Learning-Informed Physics

Almost all areas in the physical or engineering sciences rely on computational models. These models can be based on fundamental physical principles, typically formulated as a set of differential equations. Alternatively, machine learning can be used to infer input-output relations from vast datasets. Both approaches come with different strengths and weaknesses, thus, in recent years, there has been a growing interest in leveraging synergies between these paradigms. Examples include integrating machine learning as part of a numerical algorithm for solving the differential equations of a physics-based model, or incorporating physically motivated constraints to improve machine learning models. This talk will explore techniques to combine physics-based and data driven models, with applications spanning medical imaging and computational fluid dynamics.

Speaker: Dr Sebastian Götschel (TUHH)

11:35 Lunch and networking