In February 2021, three severe winter storms crippled Texas' energy infrastructure, killing 246 people and highlighting the unreliability of forecasts beyond two weeks. A MIT‑led team created a machine‑learning post‑processing model that learns to correct systematic errors of existing numerical weather prediction (NWP) systems, dramatically improving accuracy for the 2‑to‑6‑week (subseasonal) horizon.
When applied to a gold‑standard NWP model, the approach doubled forecast skill and won first place in a prominent ECMWF seasonal‑forecast competition, marking the first time a model outperformed climatological averages in this range.
The key idea is to treat AI as a “wrapper”. Using historical observations, the team quantifies the error distribution of the base model, then trains a neural network to adjust each new forecast accordingly. As more forecasts and observations become available, the wrapper continuously refines its corrections.
Training was performed on MIT Engaging, a super‑computing cluster, but the final model runs on ordinary hardware. The researchers submitted three variants that incorporated the post‑processing layer to ECMWF; the winning entry combined ECMWF’s forecast with another model and then applied the wrapper, achieving a two‑fold skill increase in the subseasonal window.
Beyond overall skill, the method yielded a statistically significant boost in extreme‑event prediction (floods, cold spells, etc.). Traditional bias correction pushes forecasts toward the mean and often suppresses tail probabilities, whereas the AI wrapper preserves and amplifies signals associated with rare events.
Because the technique is an add‑on, it can be attached to any NWP or AI‑based system, even those already bias‑corrected by other algorithms, and it remains open‑source and low‑cost. Kenya’s meteorological service has already expressed interest in using it to improve monsoon forecasts for agriculture, and energy and emergency managers note that a two‑week warning can be sufficient for preparation.
The work was supported by the U.S. National Science Foundation, the Carlos Chagas Filho Foundation, and the European Union.
Review