Every climate model on the planet makes a quiet compromise. When it looks at a patch of land tens or even hundreds of kilometers wide, it cannot possibly capture the fact that within that patch there may be a soggy hollow, a dry ridge, a forest, a wheat field and a patch of bare rock, all behaving in radically different ways. To cope, land surface models chop each grid cell into fractions, or tiles, each representing a homogeneous piece of the landscape, and compute the physics for each tile separately. The trouble is that in the real world those pieces are not strangers. Water runs downhill from one to another, heat seeps sideways through the soil, and wind drags snow from exposed ground into sheltered depressions. A new scheme called T-REX, developed by Philipp de Vrese and colleagues at the Max Planck Institute for Meteorology and partner institutions, finally teaches the land component of the ICON modeling framework to let its tiles talk to each other.
The scheme, described in the journal Geoscientific Model Development, represents five distinct lateral exchange processes: gravity-driven moisture fluxes at the surface and below ground, the advective heat those moving waters carry with them, diffusive and conductive fluxes of water and heat through the soil, and the redistribution of snow between small-scale topographic features. What makes the approach elegant is its economy. Rather than resolving the landscape spatially, which would be computationally ruinous, T-REX describes the relationship between any two tiles using a set of characteristic connectivities. These are treated as inherent properties of a pair of tiles, invariant in time and independent of location, derived from the internal logic of how the tiles were defined in the first place.
The logic is intuitive. If a grid cell is subdivided into elevations, slopes and lowlands, then runoff generated on the elevated ground will flow toward the slopes, and from the slopes into the lowlands. Those assumptions are encoded in connection matrices, whose elements describe the relative contact length between neighboring patches, the dominant hydrological flow paths between them, and how those paths differ at the surface and below ground. The connectivities then allow the model to calculate the spatio-geometric relationships between tiles, such as the contact length along which diffusive exchange occurs and the center-to-center distance that governs how quickly gradients are felt. These, in turn, define the gradients and time-lag factors that control lateral transport. Because tiles are statistical abstractions rather than mapped objects, the scheme assumes that subgrid patches can be approximated as circles, which lets a single parameter, the characteristic area, describe their geometry.
The hydrological core of the scheme is deliberately consistent with what ICON-Land already does vertically. Runoff generated on one tile is not assumed to reach its neighbor instantaneously. Instead, water is routed through intermediary reservoirs that conceptually represent the flow paths along which it moves downslope, with retention times derived from the model’s existing ARNO rainfall-runoff assumptions for catchment-scale drainage. At the micro scale, where distances of one to one hundred meters matter, lag factors are computed from assumed lateral flow velocities and the distance between tile centers. The authors chose globally uniform velocities of one meter per hour below ground and ten meters per hour at the surface, acknowledging that preferential flow through macropores and pipes makes precise estimates deeply uncertain. Snow redistribution, meanwhile, is parameterized as a bulk effect, shifting freshly deposited snow from higher to lower ground, because the model’s statistical tiles cannot know wind direction.
To test whether any of this matters, the team ran two example applications using ICON-Land standalone simulations driven by climate forcing from the Global Soil Wetness Project Phase 3, covering 1979 to 2018. The first targeted lateral water transport at global scale, subdividing each grid cell into uplands and lowlands, and into local depressions where water can pool and relative elevations where it runs off immediately. Depression characteristics were derived from the 30-meter Copernicus digital elevation model combined with a global topographic index dataset, then aggregated onto a roughly 160-kilometer simulation grid. The results were striking in their geography. Lateral fluxes exceeded 500 millimeters per year in humid regions but stayed small in the arid and semi-arid zones, and even the wet tropics showed modest fluxes where flat topography gives water little reason to move sideways.
The consequences for the simulated land surface were substantial. Lateral transport almost exclusively increased terrestrial water storage, because low-lying parts of each cell could finally retain the runoff generated upslope, and the effect on wetland extent was especially dramatic: the inundated area increased by up to 70 percent across large parts of the temperate, polar and subpolar regions relative to the potentially inundated area. In other words, without lateral exchange, the model simply cannot keep standing water on the landscape. In cold regions, frozen ground forced snowmelt to run off laterally in shallow near-surface layers, and the preferential accumulation of drifted snow in depressions delayed the moment when grid cells became completely snow-free. Yet the grid-cell mean state barely budged. Average surface temperatures cooled by less than 0.2 kelvin, and the turbulent exchange with the atmosphere changed only modestly, with the Bowen ratio shifting by as much as 20 percent in places.
Where the scheme truly transformed the simulation was in spatial variability. Because moisture converged into small depressions, local evaporative cooling far exceeded the grid-cell average, and the temperature difference between the warmest and coldest tile within a cell increased by as much as 3 kelvin, an order-of-magnitude rise across most of North America and Eastern Siberia. Comparing simulations with only micro-scale or only macro-scale transport revealed that the global patterns were similar, controlled mainly by climate and soil rather than local topography, but the micro-scale fluxes were on average about 40 percent larger and their effect on water storage about 25 percent stronger. Interestingly, the macro-scale fluxes boosted the latent heat flux more, because they moistened a large lowland area rather than concentrating water in a small depression, and evapotranspiration responds non-linearly to how widely moisture is spread.
The second application zoomed into the Arctic tundra, where patterned ground such as non-sorted circles and ice-wedge polygons creates dramatic sub-meter contrasts in soil organic matter, and therefore in thermal and hydraulic properties. The team represented these structures as circles divided into an organic-rich center, an intermediate rim and a mineral outer section, and compared ICON-Land results with DynSoM, a true two-dimensional pedon-scale soil model developed for permafrost research. Both models agreed on the essential physics: without lateral exchange, the insulating organic-rich center stays cooler in summer than its surroundings, and with lateral heat transport switched on, those temperature differences are strongly reduced, leaving notable variability only in the uppermost meter and during the snow-free season. The comparison also revealed subtleties, such as how differences in soil saturation between circle sections modulate heat conductivity and therefore the strength of lateral fluxes.
The tundra experiments also exposed the scale dependence of the physics. With a circle radius of one meter, lateral heat fluxes largely equilibrated soil temperatures. At ten meters they could no longer balance the seasonal swings at the surface, though annual mean temperatures still converged at depth. At one hundred meters the fluxes became essentially negligible, and temperature contrasts persisted below four meters. Microtopography mattered too: in low-centered and high-centered structures, a mere ten-centimeter vertical offset between sections, combined with vertical soil temperature gradients of up to 20 kelvin per meter, produced temperature differences of several kelvin, sometimes even reversing the sign of the contrast seen without lateral exchange.
The authors are candid that T-REX is less a finished set of parameterizations than a flexible framework, one designed to integrate with ICON-Land’s existing vertical physics while leaving room for improvement. The circular patch geometry can misestimate fluxes by up to an order of magnitude for elongated features, and the snow scheme ignores wind direction, a limitation that matters at hillslope scales. But the broader message is hard to ignore. As computing power pushes model resolutions toward the kilometer scale, refining the grid a hundredfold multiplies the number of cells ten-thousandfold, while adding tiles scales roughly linearly. For the micro-scale structures that shape soil temperatures, wetlands and methane-producing saturated soils, statistical tiling with honest lateral coupling is likely to remain indispensable, and T-REX shows precisely what climate models have been missing when their landscape fragments lived in isolation.
Subject of Research: Tile-based representation of lateral water and heat exchange processes in the ICON-Land land surface model
Article Title: T-REX: the tile-based representation of lateral exchange processes in ICON-Land
Article References: de Vrese, P., Stacke, T., Gayler, V., Bergstedt, H., von Baeckmann, C., Thurner, M., Beer, C., & Brovkin, V. (2026). T-REX: the tile-based representation of lateral exchange processes in ICON-Land. Geoscientific Model Development, 19(18), 9203-9234. https://doi.org/10.5194/gmd-19-9203-2026
Image Credits: AI Generated
Keywords: land surface modeling, ICON-Land, T-REX, lateral exchange processes, subgrid heterogeneity, soil hydrology, soil temperature, snow redistribution, permafrost, patterned ground, Earth system modeling, Geoscientific Model Development
News Source: Violet Maxwell. (October 9, 2026). T-REX teaches climate models how landscapes secretly share water and heat. Scienmag.
