Project G
From Molecular Lines to Physical Conditions: Teaching Neural Networks to Read the Interstellar Medium
Lukas Neumann (ESO), Caterina Bracci (ESO), Francesco Belfiore (ESO)
A major challenge in understanding star formation is determining the physical conditions of the interstellar medium, particularly the dense molecular gas closely associated with active star-forming regions. The most direct method relies on far-infrared dust emission, which is well mixed with the gas and can be modeled to infer temperature and column density. However, far-infrared observations require space-based facilities across infrared wavelengths that are currently not available and typically offer much lower angular resolution compared to modern radio and optical telescopes.
An alternative approach uses molecular line emission in the radio/millimeter regime, accessible with high-resolution observatories such as the IRAM 30m telescope and ALMA. Molecular lines are sensitive to a wide range of physical parameters -- including density, temperature, chemistry, and radiative transfer --- but this sensitivity creates significant complexity, making traditional forward modeling extremely challenging.
The growing availability of machine-learning techniques presents a powerful new way to tackle this problem. In this project, the student will explore a neural network-based framework -- using the tensorflow package in python -- to infer physical conditions of molecular gas from multi-line observational data. The model will be trained using a suite of molecular line data from an IRAM 30m large program (LEGO), combined with Herschel dust observations that provide benchmark physical parameters. Once developed and validated, this approach can be applied across Galactic star-forming regions and extended to high-resolution extragalactic observations from facilities such as ALMA.
