A knowledge-guided daily multi-layer soil freeze-thaw dataset for the Northern Hemisphere during 1950-2025

Abstract. Soil freeze-thaw (FT) dynamics regulate key thermal, hydrological, and ecological processes in cold regions, yet long-term, spatially continuous, and vertically resolved daily FT records remain scarce, particularly before the satellite era. Here, we present a soil Freeze-Thaw dataset generated using Knowledge-Guided Machine Learning (FT-KGML), providing daily FT states at 0.1° spatial resolution across Northern Hemisphere frozen-ground regions from 1950 to 2025 at depths of 10, 30, and 50 cm. FT-KGML was generated using a knowledge-guided neural network model that learns process-based FT relationships from soil-temperature simulations and is subsequently constrained by in situ observations, enabling a continuous 76-year reconstruction of subsurface FT dynamics. Independent station evaluation yielded overall daily classification accuracies of 86.12 %, 89.67 %, and 90.39 % at 10, 30, and 50 cm, respectively. Seasonal FT phenology was also well reproduced, with generally stronger agreement for the onset of soil freezing than for the onset of soil thawing. Matched comparisons with passive-microwave FT products and ERA5-Land showed broadly consistent daily FT dynamics and seasonal transition timing across datasets, providing additional support for the reliability of FT-KGML despite differences in their spatial scales and FT representations. The long-term record reveals relatively weak and spatially heterogeneous FT phenology changes during 1950-1978, followed by widespread earlier spring thaw and later autumn freezing after 1979. During 1979-2025, spring thaw advanced by 2.93-3.12 d decade^-1, whereas autumn freezing was delayed by 2.22-3.24 d decade^-1 across the three soil depths. By combining long temporal coverage, daily resolution, and multi-depth representation, FT-KGML provides a new resource for cryospheric, hydrological, ecological, and land-surface modelling communities investigating frozen-ground variability, subsurface seasonality, and long-term environmental change. The dataset is available at https://doi.org/10.5281/zenodo.22019944 (Yu and Wu, 2026).