Jorge Saavedra Bastidas   (Chile)

jorge.saavedra @ uni-heidelberg.de

Machine Learning in Astronomy - Regression Approaches for Efficient Estimation of Astrophysical Quantities

Over the past decade, the exponential growth in computing power and the volume of data available in astronomy has transformed the way astrophysical systems are modeled and analyzed. In this context, machine learning (ML) methods have become widely adopted tools for extracting complex patterns from high-dimensional data, particularly in tasks such as classification, regression, and anomaly detection.

A significant portion of computational astrophysics relies on computationally intensive simulations, such as N-body, hydrodynamical, and magnetohydrodynamical simulations. Processes such as radiative transfer, emission and absorption mechanisms, and cosmic ray propagation are typically evaluated through post-processing steps, which substantially increases the total computational cost and restricts the efficient exploration of parameter spaces.

Although ML methods are well-suited for high-dimensional and complex data, their reliable integration into astrophysical modeling frameworks remains an open challenge, as they are typically data-driven and may lack explicit physical constraints. In this thesis, we systematically study the use of machine learning-based regression models as emulators of different complex astrophysical processes. In particular, under which conditions these models can approximate computationally expensive physical functions with sufficient accuracy, and how their performance is affected by the complexity of the system, the availability of data, and the inclusion (or absence) of physical constraints.

Supervisor:    Ralf Klessen  (ITA)