The existence of a large population of Compton thick (CT; NH > 1024 cm-2) active galactic nuclei (AGN) is a key ingredient of most cosmic X-ray background synthesis models. However, direct identification of these sources, especially at high redshift, is difficult due to flux suppression and complex spectral shape produced by CT obscuration. We explored the Chandra Cosmological Evolution Survey (COSMOS) Legacy point source catalogue, comprising 1855 sources to select, via X-ray spectroscopy, a large sample of CT candidates at high redshift. Adopting a physical model to reproduce the toroidal absorber and a Monte-Carlo sampling method, we selected 67 individual sources with > 5 per cent probability of being CT, in the redshift range 0.04 ≲ z ≲ 3.5. The sum of the probabilities above NH > 1024 cm-2 gives a total of 41.9 effective CT, corrected for classification bias. We derive number counts in the 2-10 keV band in three redshift bins. The observed log N-log S is consistent with an increase of the intrinsic CT fraction (fCT) from ˜0.30 to ˜0.55 from low to high redshift. When rescaled to a common luminosity (log(LX/ erg s-1) = 44.5), we find an increase from fCT = 0.19_{-0.06}^{ 0.07} to 0.30_{-0.08}^{ 0.10} and fCT = 0.49_{-0.11}^{ 0.12} from low to high z. This evolution can be parametrized as fCT = 0.11_{-0.04}^{ 0.05}(1 z)^{1.11± 0.13}. Thanks to Hubble Space Telescope-Advanced Camera for Surveys deep imaging, we find that the fraction of CT AGN in mergers/interacting systems increases with luminosity and redshift and is significantly higher than for non-CT AGN hosts.
