Accurately representing permafrost in Earth System Models is a grand challenge that creates major uncertainty. A promising path forward is to create hybrid models that synergize process-based physics with deep learning, but this is fundamentally hindered by the non-differentiable nature of traditional land surface models (LSMs), which are incompatible with modern AI workflows. To overcome this limitation, we present NoahPy, a differentiable LSM developed by reconstructing the Noah LSM's governing partial differential equations into a process-encapsulated Recurrent Neural Network (RNN), with the heat–moisture solver forming the computational core. We first demonstrate that NoahPy very closely replicates the numerical behaviour of the modified Noah LSM, achieving Nash-Sutcliffe Efficiency (NSE) coefficients above 0.99 for both soil temperature and liquid water. We then show that at a permafrost site, the calibrated NoahPy achieves robust simulation performance for soil temperature (NSE > 0.9) and liquid water (NSE > 0.8). Critically, the differentiable workflow, when combined with the Adam optimizer, is significantly faster, more stable, and yields simulations with lower uncertainty compared to traditional Shuffled Complex Evolution (SCE-UA) calibration algorithm. NoahPy thus provides a foundational, “glass-box” framework that closes a key technical gap, enabling the development of the next generation of hybrid AI-physics models needed to more reliably predict the future of the cryosphere.
Based on ascending and descending orbit SAR data from 2017–2025, this study analyzes the long time-series deformation monitoring and slip pattern of an active-layer detachment thaw slump, a typical active-layer detachment thaw slump in the permafrost zone of the Qinghai-Tibetan Plateau, by using the small baseline subset InSAR (SBAS-InSAR) technique. In addition, a three-dimensional displacement deformation field was constructed with the help of ascending and descending orbit data fusion technology to reveal the transportation characteristics of the thaw slump. The results show that the thaw slump shows an overall trend of “south to north” movement, and that the cumulative surface deformation is mainly characterized by subsidence, with deformation ranging from −199.5 mm to 55.9 mm. The deformation shows significant spatial heterogeneity, with its magnitudes generally decreasing from the headwall area (southern part) towards the depositional toe (northern part). In addition, the multifactorial driving mechanism of the thaw slump was further explored by combining geological investigation and geotechnical tests. The analysis reveals that the thaw slump’s evolution is primarily driven by temperature, with precipitation acting as a conditional co-factor, its influence being modulated by the slump’s developmental stage and local soil properties. The active layer thickness constitutes the basic geological condition of instability, and its spatial heterogeneity contributes to differential settlement patterns. Freeze–thaw cycles affect the shear strength of soils in the permafrost zone through multiple pathways, and thus trigger the occurrence of thaw slumps. Unlike single sudden landslides in non-permafrost zones, thaw slump is a continuous development process that occurs until the ice content is obviously reduced or disappears in the lower part. This study systematically elucidates the spatiotemporal deformation patterns and driving mechanisms of an active-layer detachment thaw slump by integrating multi-temporal InSAR remote sensing with geological and geotechnical data, offering valuable insights for understanding and monitoring thaw-induced hazards in permafrost regions.
Accurately representing permafrost in Earth System Models is a grand challenge that creates major uncertainty. A promising path forward is to create hybrid models that synergize process-based physics with deep learning, but this is fundamentally hindered by the non-differentiable nature of traditional land surface models (LSMs), which are incompatible with modern AI workflows. To overcome this limitation, we present NoahPy, a fully differentiable LSM developed by reconstructing the Noah LSM’s governing partial differential equations into a process-encapsulated Recurrent Neural Network (RNN). We first demonstrate that NoahPy perfectly replicates the numerical behaviour of the modified Noah LSM, achieving Nash-Sutcliffe Efficiency (NSE) coefficients above 0.99 for both soil temperature and liquid water. We then show that at a permafrost site, the calibrated NoahPy achieves robust simulation performance for for soil temperature (NSE > 0.9) and liquid water (NSE > 0.8). Critically, the differentiable workflow, when combined with the Adam optimizer, is significantly faster, more stable, and yields simulations with lower uncertainty compared to traditional SCE-UA calibration algorithm. NoahPy thus provides a foundational, "glass-box" framework that closes a key technical gap, enabling the development of the next generation of hybrid AI-physics models needed to more reliably predict the future of the cryosphere.
Accurate, spatially explicit data on soil thickness is a critical missing link for Earth system modeling on the Qinghai-Tibet Plateau (QTP), where complex terrain and a lack of subsurface observations limit the performance of traditional empirical mapping approaches. This study presents a new high-resolution (1 km) soil thickness dataset for the QTP, specifically defined as the solum thickness extending from the surface to the C horizon interface. To overcome the inherent biases of sparse observation networks in high-altitude regions, we developed a revised mass balance model that ensures physical consistency across the plateau’s diverse landscapes. Unlike existing products derived from statistical extrapolation or machine learning, our dataset is grounded in the mechanistic coupling of climate-driven weathering with a multi-processes erosion framework that partitions sediment transport into wind, hydraulic, and gravitational components. The resulting dataset reveals that solum thickness on the QTP ranges from 0.39 m to 2.04 m, with a spatially averaged mean of 0.89 m, exhibiting a pronounced decreasing gradient from the warm, humid southeast to the cold, arid northwest. Validation against 552 soil profile observations demonstrates that this physically-based approach achieves a Root Mean Square Error of 0.34 m and a Mean Relative Error of 0.78, outperforming existing national-scale soil maps (e.g., the Shangguan and Liu maps) by approximately 10–17 %. Crucially, the model effectively reproduces realistic geomorphological patterns, such as the sharp differentiation between depositional basins and erosional ridges, which purely data-driven models often fail to capture in undersampled high-altitude regions. This dataset provides essential model parameter for hydrological, ecological, and cryosphere models, facilitating more reliable assessments of the “Third Pole” under a changing climate. The dataset is available at figshare via https://doi.org/10.6084/m9.figshare.30925358.
Widespread thermal degradation in permafrost regions is accelerating the development of retrogressive thaw slumps (RTS), severely threatening ecological stability and critical infrastructure. Current RTS modeling efforts, however, are largely confined to static susceptibility mapping, lacking the capacity to predict their spatiotemporal evolution. To bridge this gap, we developed RTSEvo, a novel dynamic RTS evolution model that couples three modules: (1) a time-series forecast of regional RTS area demand, (2) a machine-learning module for pixel-level probability mapping, and (3) a physically constrained spatial allocation module that simulates RTS expansion by integrating neighborhood effects, stochasticity, and a novel retrogressive erosion factor. Validated using manually interpreted RTS maps for 2021 and 2022 in the Beiluhe Basin on the Qinghai-Tibet Plateau, the model successfully simulated dynamic RTS growth, with the Logistic Regression-based model showing superior stability and accuracy. Furthermore, cross-regional validation confirmed the framework's structural generalizability. An interesting finding is that predictive skill is significantly enhanced by integrating process-based rules with statistical probability. Specifically, the inclusion of the retrogressive erosion factor, which mechanistically simulates upslope headwall retreat, proved critical, improving model performance by up to 29.3 % as measured by the Figure of Merit. The primary innovation of this study is the successful realization of a regional-scale dynamic simulation and prediction of RTS expansion. RTSEvo offers a highly robust scientific tool for informing proactive RTS-related risk mitigation strategies.