ECMWF has built its first machine-learning model aimed at weather forecasts two to six weeks out - a range where AI has mostly struggled.
The model, AIFS-SUBS, extends ECMWF's existing medium-range AI system to sub-seasonal timescales by taking smaller 24-hour autoregressive steps, adding stratospheric data and top-of-atmosphere radiation as inputs, and reserving five years of historical data purely for verification. Two variants were tested: one fine-tuned on operational analyses, one trained solely on the ERA5 reanalysis dataset. Both matched the operational Integrated Forecasting System in probabilistic skill across weeks two through six, while reducing systematic biases that tend to compound over long rollouts.
The more striking results involve phenomena that are notoriously hard to predict at this range. For the convective component of the Madden-Julian Oscillation - a tropical weather pattern that drives rainfall and cyclone activity globally - AIFS-SUBS extended skilful forecasts by eight days beyond what the operational IFS achieves. It also reproduced sudden stratospheric warming events with comparable accuracy, and matched IFS on how MJO activity modulates tropical cyclone frequency. These are exactly the signals forecasters need for disaster preparedness at the two-to-four-week window.
The energy figure is worth sitting with: roughly 200 times less compute at inference than the numerical model it competes with. That gap creates room for much larger ensemble runs - more members, more probability spread, better uncertainty quantification - without proportionally higher costs. AI weather forecasting has cleared the medium-range bar convincingly over the past few years; sub-seasonal has been the stubborn next wall. AIFS-SUBS doesn't demolish it, but it's the first credible data-driven attempt to clear it.