Microsoft researchers have open-sourced a machine-learning method that corrects systematic errors in existing weather forecasts rather than replacing the underlying model. The approach targets the difficult two-to-six-week range, where both physics simulations and AI systems lose precision as small biases accumulate.
Probabilistic Bias Correction learns recurring errors and adjusts a model’s output. One example is a forecast trained on historical temperatures that underestimates today’s warmer climate. Because the correction layer can run on CPUs, researchers say it may be more accessible to weather agencies that cannot afford the supercomputers or high-end GPUs used by larger forecasting systems.
In a study published in Nature, the method improved a leading European Centre for Medium-Range Weather Forecasts physics model for 98 percent of precipitation targets and 92 percent of temperature targets. The researchers also report that it doubled the accuracy of the center’s newer AI forecasting system under the study’s analysis.
Those percentages describe tested targets, not a guarantee that every local forecast will become twice as accurate. ECMWF’s machine-learning lead noted that moving research into an operational process remains difficult. A nonprofit is now testing the technique with Kenya’s meteorological service before an expected intense rainy season, which should provide evidence about performance in a resource-constrained, real-world warning system.