Title: Constraining the reionization history using deep learning from 21cm tomography with the Square Kilometre Array

Abstract: Upcoming 21cm surveys with the SKA1-LOW telescope will enable imaging of the neutral hydrogen distribution on cosmological scales in the early Universe. These surveys are expected to generate huge imaging datasets that will encode more information than the power spectrum. This provides an alternative unique way to constrain the reionization history, which might break the degeneracy in the power spectral analysis. Using Convolutional Neural Networks (CNN), we create a fast estimator of the neutral fraction from the 21cm maps that are produced by our large semi-numerical simulation. Our estimator is able to efficiently recover the neutral fraction ($x_{\rm HI}$) at several redshifts with a high accuracy of 99\% as quantified by the coefficient of determination $R^{2}$. Adding the instrumental effects from the SKA design slightly increases the loss function, but nevertheless we are still able to recover the neutral fraction with a similar high accuracy of 98\%, which is only 1 per cent less. While a weak dependence on redshift is observed, the accuracy increases rapidly with decreasing neutral fraction. This is due to the fact that the instrumental noise increases towards high redshift where the Universe is highly neutral. Our results show the promise of directly using 21cm-tomography to constrain the reionization history in a model independent way, complementing similar efforts, such as those of the optical depth measurements from the Cosmic Microwave Background (CMB) observations by {\it Planck}.