A global climate model (GCM) divides the Earth into grid cells and gives every point inside a cell the same averaged elevation, so a whole mountain range can collapse into one flat, blocky plateau. Temperature and rainfall are then computed from that smoothed terrain — real peaks, valleys and coastlines never enter the calculation, which is why coarse global models routinely miss local effects like rain shadows.
T(u,v) = T0 − lapse·max(elev,0) + coastal(u,elev)
oro(u,v) = K·(∇elev · windDir)
P(u,v) = max(0, P0 + coastal_moisture(u) + oro)
Regional downscaling doesn't rerun global physics at high resolution everywhere — it takes the GCM's coarse, large-scale forcing and reapplies it on a fine-resolution elevation map. Temperature recovers almost exactly, since it mainly tracks local elevation via the lapse rate. Rainfall recovers most of its local structure but keeps a trace of the coarse cell's baseline moisture supply, which is why its error never drops to zero the way temperature's does.
- GCM grid resolution — fewer, bigger cells mean more of the terrain's local relief is averaged away before the physics ever sees it.
- Terrain relief — taller, sharper mountains are exactly what a coarse grid is worst at representing, so error grows with relief.
- Wind direction — rotates which slopes are windward (wet, uplift) and leeward (dry, rain shadow); downscaling has to rediscover this from the fine terrain every time.
Real-world relevance: this is exactly the gap that CMIP6-class Earth System Models leave for statistical and dynamical downscaling to fill when producing the regional projections used for local infrastructure and agriculture planning.