(1) This study downloaded five variables (temperature/precipitation/snow depth/solar radiation/leaf area index) from the 6th International Coupled Model Intercomparison Project Phase 6 (CMIP6) organized by the World Climate Research Program (WCRP), and 22 climate model data were used to simulate the maximum freezing depth of seasonally frozen soil（ https://esgf-node.llnl.gov/projects/cmip6/ ）. These mode outputs use different spatial resolutions and time ranges from 1850 to 2100. This article uses the outputs of five CMIP6 tests: including one historical simulation test from 1850 to 2014; Five future climate prediction experiments from 2015 to 2100. The historical climate simulation experiment is based on actual observations and is driven by external forces to study the historical climate evolution from 1850 to 2014. For the estimation of future experiments, the CMIP6 model uses Shared Socio economic Pathways (SSPs) and Representative Concentration Pathways (RCPs) to describe the possible future development of society without the impact of climate change or climate policies, emphasizing the consistency between future radiative forcing scenarios and shared socio-economic scenarios. The selected future data includes four different scenarios: SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5. SSP1, SSP2, SSP3, and SSP5 represent four paths of sustainable development, moderate development, local development, and conventional development, respectively. Use bilinear interpolation method to unify the resolution of CMIP6 mode data to 0.5 °× 0.5 °.

    (2) Soil Texture Data: A comprehensive, gridded global soil dataset (GSDE) utilizing Earth system models was developed by the Land Atmosphere Interaction Research Group at Sun Y at sen University, http://globalchange.bnu.edu.cn/research/soilw#download ）The soil data provided by GSDE includes soil particle size distribution, organic carbon, and nutrient content. GSDE is based on world soil maps and various regional and national soil databases, including soil attribute data and soil maps. The resolution is 20 arc seconds. This article uses four soil attributes, namely sand content, mud content, clay content, and gravel content.

 Integrating station data with maximum freezing depth and predictive factors such as temperature, precipitation, freezing index, melting index, snow depth, solar radiation, leaf area index, soil texture, etc., using station data with maximum freezing depth as training data, XGBoost algorithm is used to construct a soil freezing depth model for the Northern Hemisphere. During the model training process, ten fold cross validation was used to run 300 times and output the optimal model.