In recent years, global kilometer-scale convection-permitting models have shown promising results in producing realistic convection and precipitation. Cold pools, which can be represented by km-scale models, are identified as one of the significant mesoscale processes responsible for modulating the life cycle of mesoscale organized convection. However, there is still a lack of understanding about cold pool properties across the spatio-temporal scales, as well as their representation in models. In this study, a 2.5 km global Icosahedral Nonhydrostatic (ICON) model simulation run for 40 days (06 UTC 01 Aug - 23 UTC 10 Sep 2016) from the Dynamics of the Atmospheric general circulation Modeled On Non-hydrostatic Domains (DYAMOND) initiative is used to identify thermal cold pools (using T_v) over the tropical oceans. The diurnal cycle of simulated thermal cold pools is compared against NASA’s RapidScat-observed gradient feature (GF) frequency and IMERG precipitation. ICON and IMERG exhibit morning peaks in cold pool activity similar to RapidScat GF frequency but miss the afternoon peak. Advanced Scatterometer (ASCAT) and RapidScat GF spatial climatology is also compared to ICON cold pools, where ICON shows more cold pools over the Indo-Pacific and western Atlantic basins. Random forest regression is applied to identify critical environmental properties with column water vapor found to be most important for controlling cold pool properties. Regional differences between cold pool properties are explored, where easterly waves dominate the eastern Pacific and Atlantic cold pool activity. The western Pacific and the Indian Ocean cold pools are controlled by local mesoscale forcing and intraseasonal oscillations.