Conveners
Machine Learning: From physics-Informed machine learning to machine-learning-informed physics, and back again.
- Frederic De Ceuster
Machine Learning: Improving theoretical models with ML
- Camille Landri
Description
Machine learning (ML) revolves, for a large part, about data. Yet in astronomy and astrophysics we often possess an additional, extremely valuable source of information: the governing physical equations. In this lecture, we describe how physics can inform machine learning, and conversely, how we machine learning can inform our physics simulations.
First, we introduce physics-informed approaches to model complex observations, such as ALMA spectral line images, and describe the successes and pitfalls of this approach.
Next, we reverse the perspective and show how machine learning methods, such as Gaussian Process inference, can be used, for instance, to improve solvers for differential equations.
This is known as probabilistic numerics, and we will demonstrate it with examples from radiation transport and numerical relativity.
Finally, we briefly discuss how this seemingly newfound marriage of theory and observations is not new, but provides an interesting formal framework to describe how we have been doing science in the past and will be doing it in the future.
Machine learning emulators are emerging as powerful tools for accelerating and enhancing astrophysical simulations. They are particularly powerful in the context of astrochemistry, where traditional chemical networks often involve hundreds of species and thousands of reactions and make simulations computationally expensive even under severely restrained spatial, temporal, and physical complexity. By providing fast and accurate approximations to these detailed chemical solvers, ML emulators offer a practical way to explore richer chemical–physical parameter spaces and to incorporate more realistic chemistry into simulations of star and planet formation, the ISM, and other astrophysical environments. In this lecture, I will introduce the principles behind ML-based emulators for chemistry, explaining how they are trained, validated, and deployed to replicate full chemical solvers at a fraction of the computational cost while retaining high accuracy. I will discuss current methodologies, their advantages and challenges, and show how they integrate into state-of-the-art astrophysical simulations.
