New research led by the Met Office shows that machine learning-based weather models have the potential to revolutionize global seasonal forecasting.Existing methods to predicting seasonal forecasts, or forecasts for the next three months, involve running physics-based models multiple times to understand the likely scenarios, with probabilities then applied to different conditions.

The Met Office has seen how machine learning can improve skill in short-range forecasting, and are now researching long-range AI models. The Allen Institute for AI’s machine learning model, ACE2, has demonstrated comparable, though lower, performance to existing methods. However, it does require significantly less computing power. Results found that ACE2 has the potential to generate more detailed and accurate forecasts at the seasonal range.

To assess the accuracy of ACE2, the team used an ensemble of global forecasts covering 23 Northern Hemisphere winters and compared them to physics-based predictions for the same period, as well as the actual observed conditions that took place. Varied levels of skill were seen in different locations across the globe, and in general, machine learning-based models did not yet outperform the existing physics-based methodology.

The researchers noted that ACE2 struggled to predict conditions beyond the training data, and therefore had trouble forecasting extreme winter conditions such as those that took place in 2009/2010 that resulted in the eighth coldest winter on record for the United Kingdom.

As for next steps, the team says they will need to stop relying on atmospheric conditions alone in the model, and start including ocean and other components of the climate system to help the model improve in long-range forecasting.