Sprecher/in: Tobias Buck (Heidelberg)

Titel: Machine Learning in Astrophysics: From Black-Box Magic to New Scientific Methods

Zusammenfassung: Machine learning has moved from the margins of astrophysics into the core of how we model, infer, and explore the Universe. Yet to many it still looks like black magic: powerful, opaque, and hard to trust. In this talk I argue that, used carefully, ML is becoming a genuine scientific method, one that extends rather than replaces the classical interplay of theory, simulation, and observation. I organise the discussion around three questions we ask of any physical model.
Can we predict? Neural emulators replace expensive forward models, from galactic chemical evolution to stiff astrochemical networks compressed into low-dimensional latent dynamics with neural ODEs. GPU-native, end-to-end differentiable simulations in JAX, spanning N-body dynamics and multiphase galactic winds, turn the simulator itself into an object we can optimise and learn.
Can we infer? Simulation-based inference provides principled inverse modelling, for example recovering the accretion history of the Milky Way from stellar dynamics and chemical abundances. Differentiable simulators open the door to gradient-based inference directly through the physics.
Can we see structure? Clustering and representation learning reveal substructure in high-dimensional data, from accreted debris in phase space to galaxies and their haloes.
Throughout, I emphasise what separates method from magic: calibrated uncertainties, validation against physics, and open, reproducible software.

• Datum:
  19. November 2026, 14:30 – 15:30
  
  
• Ort:

  AIP; Kirch-Haus, Konferenzraum

  
  
• Kontakt:
  Ewald Puchwein
  epuchwein@aip.de