Machine learning-based weather models could revolutionize world seasonal forecasts, according to a New Met Office-led study.
Peer-reviewed research published in NPJ Climate and Atmospheric Science assesses the feasibility of applying machine learning (ML) weather models to global seasonal forecasts, essential to understanding global weather patterns.
Seasonal forecasts looking at possible conditions for the next three months can provide valuable insight into long-term planning and decision-making, including agriculture, water resources, and even public health.
Existing methods of providing seasonal predictions involve running physics-based models multiple times to understand the most likely scenarios and applying probabilities to various conditions. MET Office's seasonal forecasts are currently one of the best in the world, clarifying potential scenarios for upcoming weather.
New HCCP-funded research shows that the ML model of AI (AI2) known as ACE2 shows comparable (lower) performance to existing methods, indicating significantly less computing power.
Chris Kent, senior scientist and lead author of Met Office, said:
“We found that ACE2 exhibits skills comparable to existing physics-based methods, which could open up new opportunities to generate more detailed and accurate predictions over the seasonal range.”
“The Minister of AI and Digital Government is Feryal Clark,” said:
“Improved weather forecasting capabilities a few months in advance will help businesses and public services provide the information they need to plan with greater confidence. From supporting crop decisions to helping winter preparations, this task shows that new technologies like AI can promote growth, strengthen resilience and improve people's daily lives as part of planning for change.”
Compare accuracy
To assess the accuracy of ACE2 seasonal predictions, we compared a global prediction ensemble covering 23 Northern Hemisphere winters with physics-based predictions over the same period with the reality of observed conditions we know.
The spread of the ensemble was similar to existing methods, with skillful predictions of the North Atlantic Oscillation, which often affects the weather conditions in Europe and North America. When predicting winter North Atlantic oscillations, correlation scores between 0.3 and 0.6 are usually seen across physics-based models, whereas ACE2 achieves scores below 0.5. In this methodology, a score of 1 shows perfect accuracy for the observed conditions.
However, different levels of skill have been demonstrated in different locations around the world, and in general, ML-based models have not yet surpassed existing physics-based methodologies.
Figure: The above map shows correlations of the ML-based ACE2 (left) and physics-based glossea (right) models with observed conditions beyond the 23 Northern Hemisphere winters from 1993/94 to 2015/16. Deep red indicates a higher level of accuracy, while blues indicates a lower level of accuracy.
Extreme prediction
Despite demonstrating the potential winter predictions, this study also highlights the potential limitations that rely on ML-based seasonal predictions.
The winters of 2009/10 were particularly cold, with the record North Atlantic Oscillation and the eighth coldest winter being recorded in the UK in a series dating back to 1884.
Paper co-author and director of the Met Office of Long Distance Prediction, explained: “We found that ML models struggle to predict conditions beyond training data. 2009/10 is a key test case for prediction, and there is no physics of traditional models.
“While it is based on an understanding of atmospheric physics, it will be important to harness the benefits of these fast ML-based models for accelerated improvements in seasonal predictions.”
Next Steps
This proof-of-concept study is another step on the path to next-generation weather forecasting, highlighting Met Office's continued innovative approach to increasing the power of weather intelligence.
The lead research scientist in climate modeling at AI2 Oliver Watt-Meyer said: “It is an exciting moment for geoscience machine learning as we move beyond medium range weather to seasonal and longer timescales.
Collaboration is key to realizing the complete benefits of AI modeling for weather forecasting. Met Office aims to promote both short- and long-term prediction improvements in a safe and reliable way by combining physics and AI/ML-based predictions, with further research needed to better understand opportunities and limitations.
Learn more about AI and Met Office.
