Machine Learning Techniques for Astrophysics and Cosmology: Simulation-Based Inference

TL;DR

This paper reviews simulation-based inference (SBI) techniques for astrophysics and cosmology, emphasizing their applications, diagnostics, and current challenges like limited training data.

Contribution

It provides a comprehensive overview of SBI methods, practical guidance, diagnostics, and discusses key challenges in applying SBI to astrophysics and cosmology.

Findings

SBI enables inference with intractable likelihoods in astrophysics.

Diagnostics are essential to validate SBI results.

Limited simulation budgets are a major challenge for SBI applications.

Abstract

Simulation-based inference (SBI) enables parameter inference by training neural networks on forward simulations. It is being applied both for intractable likelihoods as well as under time constraints on the posterior sampling. After motivating situations in which SBI is useful, we give a pedagogical description of the basic techniques. These are posterior, likelihood, and ratio estimation. Alternatives, sequential versions, and learned summaries are discussed briefly. We provide a brief guide to choosing among the techniques in practical scenarios. SBI needs to be verified through diagnostics since failures can be subtle but would invalidate the inference result. We explain the most common diagnostic techniques. We briefly list some recent SBI applications in the cosmology and astrophysics literature. Before concluding, we discuss current methodological challenges. We identify training with limited simulation budgets as the critical problem for applications to cosmology and astrophysics.