Post-processing Subglacial Physics Model with Per-site JEPA Specialist Models Enable Natural-language Glaciological Intelligence

Glaciers slide over their beds, and how fast they slide depends on the pressure of meltwater trapped beneath them. This subglacial environment is nearly impossible to observe directly and is one of the largest uncertainties in how fast glaciers lose ice and how much water they deliver to communities downstream. Ice-sheet models represent it with empirical rules that are hard to test, and although machine learning and queryable artificial-intelligence systems are now reshaping much of earth science, no comparable interpretable-learning analysis or embedding-model system yet exists for subglacial processes. We address this gap in two ways. Treating coupled physics-based subglacial hydrology and ice velocity simulations of four contrasting glaciers (Himalaya, Antarctica, southeast Greenland, and northwest Greenland) as controls, we use interpretable machine learning, from linear models to recurrent networks, to understand what controls sliding. We find that bed water pressure relates to ice surface flow speed once the sliding component is isolated from the deformation component($r=-0.50$). The predictive memory of the system mostly lives in the present state of the bed, regardless of meltwater history, with state-aware models reaching $R^{2}=0.84$ against $0.24$ without it. Across the four test glaciers two consistent regimes emerge - one dominated by water pressure and the other by bed shape, while instantaneous meltwater input is essentially never the direct control, indicating where common sliding rules may not apply. Leveraging these findings, we build a natural-language glaciological intelligence system for subglacial processes: a fleet of self-supervised graph foundation models, one joint-embedding predictive architecture (JEPA) per glacier, that learn rich, non-degenerate representations. These are queried through an agentic large-language-model (Claude Sonnet/Opus) router that retrieves from the appropriate specialist and improves answer quality compared to baseline, with the value of added context scaling with the consuming model's reasoning ability.