AbstractThe maximum annual freeze depth (MFD) is a primary indicator of the thermal state of frozen ground, affecting ecosystems, hydrological processes, vegetation growth, infrastructure, and human activities in cold regions. It is thus important to quantify the past, present, and future spatial and temporal variability of MFD at the hemispheric scale. We develop a data‐driven MFD simulation method within a machine learning framework, integrating MFD observations from meteorological stations and several environmental predictors, to analyze past and future scenarios in the Northern Hemisphere (NH). Based on ERA5 reanalysis estimates and historical to future CMIP6 scenarios, the NH MFD averaged 133 cm (ERA5) and 131 cm (CMIP6) during 1981–2010, and will vary 81–112 cm during 2015–2100 depending on the emission scenario. During 1950–2013, MFD decreased by 0.37 cm/a (ERA5) versus 0.22 cm/a (CMIP6), and is projected to decrease 0.16–0.69 cm/a by 2100. During 1981–2010, MFD decreased by an average of 19.1% (ERA5) and 13.9% (CMIP6), with a net change of −17 cm (ERA5) and −13 cm (CMIP6). Depending on the emission scenario, MFD will decrease 11% (−12 cm) to 42% (−19 cm) between 2015 and 2099 relative to the 1981–2010. Warming, increased moisture, warmer cold seasons, warmer warm seasons, shallower snow depths, and increased vegetation cover all lead to a reduction in MFD. The results from this novel machine learning approach provide useful insights regarding the fate of future frozen ground changes.