Has this been tried for weather forecasting?

[meglamaniac]

So in one of my random brain spasms, I wondered about bayesian learning.
Many spam filters use it - it starts off crap, but as you teach it (by clicking "this is not spam" or "this is spam" when it gets it wrong), it "learns" the characteristics of your email, and after a couple of days it can filter with a very low false positive rate.

Could this be applied to weather? Or has it already been done?
So far as I'm aware from sort of general knowledge, current weather forecasting relies on networks of sensors and satelite data, which is then plugged into mathematical models, squeezed through the Met's latest supercomputer, and prettified for our viewing pleasure. The obvious downside to that is that you only really know the conditions where the sensors have sampled; anything else is extrapolation (guesswork), and small differences scaled up can mean the forecast for 3 days down the line is totally screwed.

Would it be possible to somehow combine the models with the bayesian techniques, so that the system could "learn" to spot when the weather is likely to do something unexpected that the rigid maths won't pick up?

Just a bit of a musing really.