Japan Plans 30-Year Supercomputer Forecasts

BaltikaTroika writes "According to a ministry representative, 'Japan is planning ultra long-range 30-year weather forecasts that will predict typhoons, storms, blizzards, droughts and other inclement weather.' Maybe they should tell their secret to my local weatherman, who usually can't even get tomorrow's weather right. Whatever happened to chaos?"

Comments

RTFA, submitter

[Anonymous Coward]

The results will help establish predictable routes for typhoons and identify areas that are recurring targets for heavy rains, abundant snow, high waves, heavy winds, scorching heat or crop-threatening droughts.

They're not trying to forecast weather 30 years in the future, they're just looking for statistical trends in locations where hurricanes and such are more likely to occur, based on predictions of the overall global climate. Things like "there are probably going to be 20-40% more typhoons off the east coast of Japan in 10 years", not "watch out, Tokyo is going to be hit by a tsunami on August 12, 2032".

A Few Things

[Hoplite3]

1) The computer will be doing CLIMATE modeling, not weather prediction. That's a different bird. It's like the difference between the average score on a test and your score on the test. Or like describing the flow of heat, but not knowing the underlying collisions that result in the transfer of energy.

2) Higher precision does help you model chaotic systems longer, but... If you run your model until the difference between your prediction and the actual system is larger than a tolerance, the time when this happens is called the horizon time. If you improve your accuracy (let's say your computer system is perfect and errors only occur in getting the initial state right), you only improve the horizon time as the LOG of your improvement. In an age where quadratic methods are just adequate in scientific computing, this is unbearable.

3) Another weather (not climate) prediction option is to use a statistical cohort model. Such a model just takes in data and tries to predict what will happen next based on past trends. It doesn't know any physics, and can take a while to train. This means that the cohort you train in London is useless in Paris. Such "models" often beat physical models in predictive ability, but don't give any insight into why. If you want to fly a plane, they're fine. If you want to do science, see (1) or (2).

Also, this computer is way, way cooler than the one predicting nuclear bomb blasts. But that's, just like, my opinion, man.