DeepMet has extended the forecast horizon for temperature and humidity to 45 days

A system combining physics and machine learning reduces forecasting errors by between 20 and 60 per cent compared with the European reference centre, over the challenging two- to six-week forecast window.

A team from Wuhan University and the University of Tennessee, led by Jia Xing, has unveiled DeepMet, a weather forecasting system that tackles a persistent blind spot: the sub-seasonal to seasonal timescale – between two and six weeks – where conventional models lose most of their accuracy beyond ten to fourteen days.

Over this timeframe, DeepMet claims a 20–60 per cent reduction in forecast errors compared with the European Centre for Medium-Range Weather Forecasts’ system, and an improvement of up to 138 per cent in the representation of large-scale structures. The detection of extreme heat and cold spells is around 40 per cent more accurate, with forecasts for temperature, humidity and wind extending up to 45 days.

Two design choices set the system apart from conventional AI weather models: a physics-driven approach, and the simultaneous prediction of the entire period rather than a step-by-step sequence that accumulates errors. The entire training process can be completed on a single graphics processing unit within twenty-four hours. Jia Xing emphasises that the combination of physics and machine learning has made it possible to identify signals that traditional forecasting systems usually miss.