Microsoft researchers and their collaborators published a machine-learning method for improving weather forecasts two to six weeks ahead in Nature on 14 October, MIT News reported. Called Probabilistic Bias Correction, it learns to adjust errors in existing forecasting systems. The researchers say their forecasts beat climatologic averages across a period in which predictions have usually struggled to do so.
Key points
- The method learns recurring errors in a weather model and adjusts its forecasts for the following two to six weeks.
- Researchers reported improvements for 98 percent of precipitation targets and 92 percent of temperature targets when applying it to an ECMWF physics-based model.
- A competition entry using the correction won first place in ECMWF’s AI Weather Quest.
- The software is open source, and researchers say the resulting model does not require a supercomputer to run.
Probabilistic Bias Correction works on existing forecasts
Conventional weather models calculate how the atmosphere evolves, while AI models learn patterns from past weather data. Both can develop recurring errors where observations are sparse or the processes being modelled are difficult to represent. Those errors can build as a forecast looks further ahead. Microsoft’s approach does not replace either kind of model: its machine-learning layer studies earlier forecasts and their errors, then adjusts the model’s output, The Verge reported.
Forecasts at this range assign likelihoods to possible outcomes rather than relying on one predicted result. Probabilistic Bias Correction shifts those likelihoods using patterns it has learned from historical data, according to MIT researcher and co-author Judah Cohen. The idea is akin to recalibrating a scale that tends to read low, except that the correction adjusts a spread of possible weather outcomes. Microsoft researcher Lester Mackey gave The Verge an example: a model drawing on older records could favour temperatures that are too cold for a warming climate.
Planning when to fertilise around rain could benefit from an earlier indication of likely conditions if the reported improvement carried over to a particular place and season. It would still involve choosing between possible outcomes, because the method adjusts the likelihoods in a forecast rather than supplying a certain date for rainfall.
Cohen describes the two-to-six-week interval as a “predictability desert”. Shorter-range forecasts can draw heavily on the atmosphere’s starting conditions, while longer-range predictions can make use of slower-changing influences. According to Cohen, the interval between them makes less use of either advantage. The researchers trained their correction model using MIT Engaging, a supercomputing cluster, although MIT News says the resulting model does not need a supercomputer to run.
ECMWF targets and the competition result
In the researchers’ analysis, applying the correction to the European Centre for Medium-Range Weather Forecasts’ physics-based model improved accuracy for 98 percent of precipitation targets and 92 percent of temperature targets, The Verge reported. The researchers also found that the framework doubled the accuracy of ECMWF’s AI forecasting system in the study’s analysis. Those percentages describe the share of targets that improved, rather than the size of the improvement for each target.
The competition tested a related combination. The entry that took first place in ECMWF’s AI Weather Quest joined an ECMWF forecasting model with another model before applying the correction, according to MIT News. The researchers reported that this entry doubled the original model’s predictive power over the subseasonal range and beat forecasts based on climatologic averages. That comparison matters because a forecast weeks ahead must offer more than the weather typically expected for the time of year to be useful.
The researchers also reported a statistically significant improvement in predicting extreme weather. Cohen says ordinary error correction can pull forecasts towards average conditions, making unusual events harder to predict. He pointed to ECMWF’s non-AI correction, which he said worsened its model’s predictions of extremes. Matthew Chantry, ECMWF’s strategic lead for machine learning, called turning research ideas into an operational process “a most challenging step” in comments to The Verge.
CPU use and testing in Kenya
For organisations building forecasts, the method offers a way to work with an existing system rather than construct a new one. Mackey says the correction can run on CPUs without GPUs, and the software has been released as open source. That running requirement differs from the computing used to develop it: the team used MIT’s supercomputing cluster to build the model. Cohen says the researchers also applied the correction to other forecasting models and improved their output.
Genevieve Flaspohler, a paper co-author and executive director of the nonprofit Rhiza Research, told The Verge that a lightweight method could let scientists with fewer computing resources tune forecasts for particular localities. The claim is being put to a practical test. Rhiza Research is working with the Kenya Meteorological Service Authority, with Gates Foundation funding, to test whether methods including Probabilistic Bias Correction can help deliver more accurate early warnings.
