MIT researchers develop AI tool to improve subseasonal weather forecasting
Researchers at MIT have developed a machine-learning model, known as Probabilistic Bias Correction (PBC), designed to improve weather forecasting accuracy in the two-to-six-week time frame. This subseasonal range, often called the predictability desert, has historically been difficult to forecast with reliability. The new AI-based postprocessing layer functions as an add-on to existing dynamical models, learning from historical errors to refine future predictions. In testing, the tool doubled the predictive power of the European Center for Medium-Range Weather Forecasts model and demonstrated a statistically significant improvement in predicting extreme weather events, such as cold spells and flooding.
Key points
- MIT researchers developed an AI-based tool called Probabilistic Bias Correction (PBC) to improve subseasonal weather forecasts.
- The tool functions as a postprocessing layer that corrects systemic errors in existing forecasting models.
- PBC doubled the predictive power of the European Center for Medium-Range Weather Forecasts (ECMWF) model in the two-to-six-week range.
- The model shows improved accuracy for extreme weather events, which traditional bias correction methods often struggle to predict.
- The tool is open source and can be applied to various existing forecasting models.
What changed
A team of researchers, including those from MIT, Microsoft, and the European Center for Medium-Range Weather Forecasts (ECMWF), has introduced a new machine-learning technique to enhance subseasonal weather predictions. The tool, detailed in the journal Nature, addresses the two-to-six-week forecasting window, a period previously characterized by low reliability.
The Probabilistic Bias Correction (PBC) tool acts as a wrapper for existing forecasting systems. By analyzing historical data to identify and remove systematic biases, the model adaptively improves its corrections. Unlike traditional bias correction methods that often regress toward the mean, the PBC tool maintains accuracy for extreme weather events.
Availability
The Probabilistic Bias Correction tool is open source and designed to be low-cost. While the researchers utilized a supercomputing cluster at MIT to build the model, the resulting system does not require a supercomputer to operate. The meteorological office for the Republic of Kenya has reportedly expressed interest in implementing the model to assist with agricultural planning.
Why it matters
Improved subseasonal forecasting could allow energy and agricultural sectors to prepare for extreme weather events weeks in advance, potentially mitigating infrastructure failures and protecting crop yields.
What we know
- MIT researchers developed the Probabilistic Bias Correction (PBC) machine-learning model.
- The PBC model improves weather forecasting accuracy in the two-to-six-week subseasonal range.
- The model doubled the predictive power of the ECMWF forecasting model in testing.
- The tool is open source and can be applied to various existing forecasting models.
- The research was published in the journal Nature.
What remains unclear
- The extent to which the model will be adopted by global meteorological agencies remains to be seen.
- The long-term performance of the model across all global climate regions is subject to ongoing evaluation.
How the story developed
MIT researchers publish findings on AI weather forecasting tool
Researchers announce the development of the Probabilistic Bias Correction tool in the journal Nature, demonstrating improved subseasonal forecasting accuracy.
Original source
