Physically informed TabPFN mapping of glacier collapse susceptibility in the southeastern Qinghai–Xizang Plateau

Regional assessment of glacier-collapse susceptibility is constrained by sparse event inventories and spatial dependence among samples from individual failures. We developed a physically informed framework that connects multi-source glacier characterization, two-stage predictor refinement and spatially constrained model comparison, anchored to four documented release-area failures in the Sedongpu valley. The labelled archive contains 19,460 records at 30 m grid-cell support, including 3632 collapse-associated cells from one connected event system rather than independent regional failures. Twenty predictors represent terrain, glacier geometry and kinematics, archive-conditioned hydrothermal conditions, structural and seismic context, and mechanical proxies. Hydrothermal predictors use observations preceding the 22 March 2021 archive cutoff, whereas annual velocity mosaics are retrospective calendar-year descriptors. Correlation control and performance-weighted Shapley additive explanation refinement reduced redundancy. We compared TabPFN, four tree-based classifiers, soft voting, and stacking under local spatial-block separation. TabPFN achieved the highest threshold-independent discrimination (area under the receiver operating characteristic curve, 0.9513; average precision, 0.8690) and the lowest Brier score (0.0874). Nine alternative local partitions supported TabPFN as the aggregate model choice, although average-precision ranking favoured XGBoost in four realizations. In retrospective July 2022 susceptibility mapping across the southeastern Qinghai–Xizang Plateau, high and very high classes occupied 731.9 km 2 , or 10.03% of the valid glacier prediction domain, as spatially selective clusters. These classes rank glacierized areas by their similarity to the Sedongpu environmental signature and can prioritize follow-up inspection; they do not estimate event frequency or absolute failure probability. The study establishes a regional benchmark linking physical predictor design with pretrained tabular learning and offers a transferable evaluation workflow for data-scarce glacierized regions worldwide, with cross-basin performance to be tested through independent event-resolved validation.