models/NeuralGCM
NeuralGCM
NeuralGCM is a way of simulating the atmosphere in which a learned piece of software fills in the physics that is too small to see, inside an otherwise conventional weather and climate model.
What it is
A general circulation model simulates the atmosphere by dividing the world into a grid and stepping forward in time, solving the equations of air flow in each cell. But many processes are smaller than a cell: cloud droplets forming, rain falling, sunlight being absorbed, swirls of turbulence. Traditional models handle these with hand-written approximations. NeuralGCM replaces those approximations with a neural network, a function whose behaviour is set by numbers adjusted through training rather than written out by hand. The network looks at the state of each vertical column of air and outputs how fast each quantity should be changing because of the unresolved processes.
The twist is that the solver itself is differentiable: the model can be asked how a small change in the network's numbers would change the simulated weather hours ahead. So the network is not trained in isolation on a pile of examples. It is trained while running inside the simulation, against what the atmosphere actually did, with the errors traced back through the solver. The equations of motion stay in place and keep the result physically anchored; only the uncertain small-scale part is learned.
Why scientists use it
Two things scientists want from an atmospheric model pull in opposite directions. Physical equations generalise: they hold in climates that have never been observed, and they conserve mass and energy. Learned components fit the data: they can reproduce the messy statistics of clouds and rain better than a short formula written by a person. Putting the network inside the solver is an attempt to get both, rather than choosing.
Training inside the simulation also matters for a reason beyond accuracy. A component fitted on single snapshots can look good for one step and then drift, because its small errors are never tested against what they do over hours and days. Letting the errors flow back through the solver means the learned part is judged on the trajectory it produces, which is what forecasters and climate scientists actually use.
In the record
Limits
The learned part is only as good as the record it was trained against, so where observations are sparse or the atmosphere moves outside the conditions seen in training, its corrections are extrapolation and can quietly go wrong. Keeping the equations of motion helps but does not guarantee that learned tendencies conserve anything. The approach also inherits the grid: processes finer than a column are represented, not resolved, and a plausible-looking simulation can still be wrong in ways that only show up in long runs.
Written by AIxSci: the background from general knowledge, the examples from checked records.