Rather than building a new standalone weather model, PBC uses machine learning to identify and correct systematic bias in existing ones. One example: a model trained on historical records can lean towards colder temperatures than the world actually experiences today as greenhouse gas emissions warm the planet. Such approximations for gaps in data, whether from sparse weather stations or poorly understood atmospheric phenomena, compound into errors the further ahead a forecast looks.

In the study's analysis, PBC improved the accuracy of a leading physics-based model from the European Centre for Medium-Range Weather Forecasts (ECMWF) for 98 percent of precipitation and 92 percent of temperature targets, and doubled the accuracy of ECMWF's new AI forecasting system. ECMWF had already recognised the research in its AI Weather Quest competition before publication, and its strategic lead for machine learning, Matthew Chantry, called the paper a good example of new ideas developed for the community's benefit, while cautioning that turning ideas into operational processes remains the most challenging step.

A practical selling point is efficiency. Where physics-based modelling needs supercomputers and generative AI leans on powerful GPUs, senior principal researcher Lester Mackey says PBC can run on CPUs alone, which the authors hope makes it accessible to weather agencies in economically developing countries. Co-author Genevieve Flaspohler, now executive director of nonprofit Rhiza Research, says the team is testing PBC with the Kenya Meteorological Service Authority, funded by the Gates Foundation, ahead of an expected more extreme rainy season during a record-breaking El Niño.