Machine learning correction doubles the skill of weather forecasts two to six weeks out
In a Nature paper, a team including MIT and Microsoft researchers, with a Harvard College undergraduate among the lead authors, described a machine learning method that learns from a forecast model's past errors and corrects its output two to six weeks ahead. Applied to a gold-standard ECMWF model, it doubled predictive power in that range, placed first in a forecasting competition and improved extreme weather predictions.
This is applied math in action: chaos theory explains why forecasts break down after about two weeks, and probabilistic modeling plus statistical learning pushed past that limit. The method is an open-source add-on that can wrap any model and does not need a supercomputer. Dynamical systems, probability and numerical methods are the foundation.
Which interests you more: the physics models or the data-driven corrections on top?

0 Comments
Sign in to join the discussion
No comments yet. Be the first to comment!