Daniel Walter (Germany)
daniel.walter @ stud.uni-heidelberg.de
A common problem in astrophysics is to infer system parameters from observations. Usually, a physics-based model (SED modeling codes, numerical simulations, etc.) connects inference parameters with observables. In these circumstances, frequentist and Bayesian inference typically require advanced computational methods (e.g., MCMC). However, likelihood-based methods can perform poorly when the likelihood function is hard to compute or completely intractable, which happens frequently when the forward model contains stochastic elements. In these cases, so-called likelihood-free methods can provide better solutions. In recent years, modern machine learning techniques have vastly extended the scope of likelihood-free approaches. Generative models, like normalizing flows, can learn complicated high-dimensional posterior distributions and can outperform more traditional approaches in computational efficiency.
My research focuses on the development and application of machine learning-based inference methods, with a particular emphasis on Bayesian inference with normalizing flows. In addition, I try to find new ways to estimate information-theoretic quantities, like mutual information, to better understand the information that observables contain about inference parameters. These methods can be applied to many astronomical inference problems. As a starting point, I am working on modeling posterior distributions of star cluster parameters (e.g., age, mass, and metallicity) from unresolved broadband photometry of the PHANGS surveys of nearby galaxies.
Supervisor: Ralf Klessen (ITA)