Microsoft researchers have a new way to make forecasts two to six weeks out more accurate, and it doesn't need a supercomputer or a rack of GPUs. The method, called Probabilistic Bias Correction (PBC), was published in Nature today and released as open-source software, The Verge reports.
PBC isn't a new weather model. It uses machine learning to spot the systematic errors existing models make and correct them. One example from Microsoft's Lester Mackey: a model trained on historical records can lean colder than today's warming climate. In the study, PBC improved the European Centre for Medium-Range Weather Forecasts' physics-based model on 98% of precipitation targets and 92% of temperature targets, and it doubled the accuracy of ECMWF's new AI forecasting system. The paper's authors come from Microsoft, MIT, ECMWF and other institutions.
That range is where forecasts usually fall apart, for physics-based and AI models alike, and it's the range farmers and emergency planners care about. Because PBC runs on CPUs, it's within reach of weather agencies without much money. Co-author Genevieve Flaspohler now runs the nonprofit Rhiza Research, which is testing PBC with the Kenya Meteorological Service ahead of a rainy season expected to be more extreme during a record-breaking El Niño. ECMWF's Matthew Chantry called it a good example of new ideas for the community, while noting that "turning ideas into reality [in] an operational process is a most challenging step."
— via The Verge