About
A weather model on a Mac mini
A spherical cow is what a physicist reaches for when the real thing is too complicated. A neural weather model is the same instinct at scale: a 711-million-parameter approximation of the most famously chaotic system we live inside.
The model
Forecasts here come from FourCastNet 3, NVIDIA's open-weights weather model. It predicts 72 atmospheric variables — winds, temperature, humidity, geopotential — on 13 pressure levels across a 0.25° global grid, stepping forward six hours at a time. It is probabilistic by construction: noise evolves through the forecast, so running it several times gives a spread of plausible outcomes rather than one false certainty.
Why a Mac
Because it turned out to be possible. The model's spherical convolutions ship as CUDA kernels with no Apple equivalent, so the core operator was reimplemented: a discrete-continuous convolution on the sphere is a circular cross-correlation in longitude, which makes it a narrow banded matrix in Fourier space — something Metal can do natively.
The result runs a full global forecast step in about 30 seconds on a desktop machine that draws less power than a lightbulb. A ten-day global forecast takes roughly twenty minutes, on hardware you could put on a shelf.
Where the data comes from
Initial conditions are NOAA's GDAS analysis, free and public, available about four hours after observation time. Only the fields the model needs are fetched — around 60 MB per cycle instead of half a gigabyte.
Most comparable projects initialise from ECMWF's ERA5 reanalysis, which is excellent and roughly five days behind real time. Four hours versus five days is the difference between a forecast and a hindcast.
Honesty
Every forecast is scored against what actually happened, and the results are published — including the variables this setup handles badly. This is a curiosity project by someone with access to good public data and too much interest in chaotic systems. It is not a weather service.