Machines Learning to Do Research – Artificial Intelligence as a Tool to Understand the Universe

Abstract

The article explores how artificial intelligence can significantly accelerate cosmological research by emulating computationally expensive theoretical models. In modern cosmology, observations such as the cosmic microwave background and Type Ia supernovae are combined with theoretical frameworks based on general relativity, statistical physics, and particle physics to infer the composition and evolution of the Universe. Parameter estimation is commonly performed using Markov chain Monte Carlo (MCMC) methods, which may require hundreds of thousands to millions of model evaluations, leading to weeks of CPU time.

The author demonstrates how neural networks can be trained to emulate the CLASS code, reducing computation times by several orders of magnitude. Through an iterative active learning approach, training points are selected strategically in high-likelihood regions of parameter space, allowing the emulator to converge toward the results of full numerical calculations. This enables rapid and accurate estimation of Bayesian posterior distributions for cosmological parameters such as neutrino mass and the Hubble constant.

Furthermore, the article highlights new methodological opportunities enabled by fast differentiable emulators, including gradient-based sampling techniques and real-time visualization tools such as the CosmoSlider app. Overall, the work illustrates how artificial intelligence not only enhances efficiency but also expands the methodological toolbox available for exploring the structure and evolution of the Universe.