Supervised Extraction of the Thermal Sunyaev$-$Zel'dovich Effect with a Three-Dimensional Convolutional Neural Network

The thermal Sunyaev$-$Zel'dovich (SZ) effect offers a unique probe of the hot and diffuse universe that could help close the missing baryon problem. Traditional extractions of the SZ effect, however, exhibit systematic noise that may lead to unreliable results. In this work, we provide an alternative solution using a three-dimensional Attention Nested U-Net trained end-to-end with supervised learning. Our labeled data consists of simulated SZ signals injected into $\textit{Planck}$ frequency maps, allowing our model to learn how to extract SZ signals in the presence of realistic noise. We implement a curriculum learning scheme that gradually exposed the model to weaker SZ signals. The absence/presence of curriculum learning significantly impacted the amount of bias and variance present in the reconstructed SZ signal. The results from our method were comparable to those from the popular $\textit{needlet internal linear combination}$ (NILC) method when evaluated on simulated data as well as real-world SZ signals. We conclude by discussing future avenues for advancing machine learning extractions of SZ signals.