Understanding the Impact of Semi-numeric Reionization Models when using CNNs

TL;DR

This study examines how CNNs trained on semi-numeric 21cm reionization models perform poorly across different models, emphasizing the need for diverse training data to accurately interpret Epoch of Reionization observations.

Contribution

It demonstrates that CNNs for reionization analysis require training on multiple semi-numeric models to generalize effectively across different data sources.

Findings

CNNs trained on a single model perform poorly on other models.

Including multiple models in training improves CNN prediction accuracy.

Results highlight the importance of diverse training data for EoR analysis.

Abstract

Interpreting 21cm measurements from current and upcoming experiments like HERA and the SKA will provide new scientific insights and exciting implications for astrophysics and cosmology regarding the Epoch of Reionization (EoR). Several recent works have proposed using machine learning methods, such as convolutions neural networks (CNNs), to analyze images of reionization generated by these experiments since they could take full advantage of information contained in the image. Generally, these studies have used only a single semi-numeric method to generate the input 21cm data. In this work, we investigate the extent to which training CNNs for reionization applications depends on the underlying semi-numeric models. Working in the context of predicting CMB optical depth from 21cm images, we compare networks trained on similar datasets from 21cmfast and zreion, two widely used semi-numeric reionization methods. We show that neural networks trained on input data from only one model produce poor predictions on data from the other model. Satisfactory results are only achieved when both models are included in the training data. This finding has important implications for future analyses on observation data, and encourages the use of multiple models to produce images that capture the full complexity of the EoR.