Ask a farmer in Narok County what the weather will do tomorrow, and you are asking a question global weather models are not built to answer. They describe the weather across an area roughly 50 kilometres wide: one number covering highland and valley, forest and grazing land, the windward and the leeward side of the same ridge.

For a continental overview, that is enough. For someone deciding whether to plant this week, where to move livestock, or whether to evacuate a low-lying settlement, it is not. Rain that falls on one side of a hill and not the other is the difference between a harvest and a loss. A flood warning issued for “the region” leaves everyone guessing whether it means them.

Within the SAFE4ALL project, Neuralio AI has developed the HyperLocal Weather Intelligence Engine to close part of that gap, turning coarse global forecasts into detailed local ones.

Why Not Simply Run a Numerical Model?

There is a conventional way to produce a high-resolution forecast: run a physics-based regional weather model over the area you care about. It works, and it is standard practice. But it is expensive. A single run can occupy a computing cluster for hours, and an ensemble — several versions of the same forecast, to capture uncertainty — multiplies that many times over.

For many national meteorological services, this is the binding constraint. The science is not the obstacle. The computing budget is.

The engine takes a different route. Rather than solving the physics again at high resolution, its ORBIT2 model learns the relationship between a coarse forecast and a detailed one, then applies what it has learned.

Teaching an AI Model What Detail Looks Like

ORBIT2 was trained on roughly 17 terabytes of paired data: five years of coarse global forecasts alongside the detailed 2-kilometre simulations produced from them. Given enough examples, it learns how a coarse field becomes a detailed one — where a valley traps cool air overnight, where a slope forces rain out of a passing system.

The model does not work from weather data alone. Information about the land itself, including elevation, land cover and soil type, is fed into the process, so that the output follows real topography rather than inventing detail that merely looks plausible.

The result is a forecast at 2 kilometres, a twenty-five-fold increase in resolution, covering temperature, rainfall, wind and relative humidity out to seven days ahead — produced in minutes on a single GPU, at a small fraction of the cost of conventional high-resolution modelling.

That last point is not a convenience but the thing that makes the approach usable at all. A forecast taking minutes rather than hours can be refreshed as new global data arrives, run as a full ensemble rather than a single best guess, and produced on modest hardware.