Global climate models predict future weather conditions, but these models are limited in their ability to provide detailed forecasts for specific locations. Policymakers often need to supplement these coarse-scale models with high-resolution ones to predict local extreme weather events. However, the accuracy of these predictions heavily depends on the initial coarse model’s accuracy. Themistoklis Sapsis, from MIT’s Department of Mechanical Engineering, along with his team, has developed a method to refine the predictions of coarse climate models.
The researchers combined machine learning with dynamical systems theory to guide the climate model’s simulations towards more realistic, large-scale patterns. The modified scheme was then applied to smaller scale models to generate more accurate predictions of specific events like floods and cyclones in defined locations over coming decades. This method can be applied to any global climate model. The refined models can be more effective at identifying regions at risk of extreme weather resulting from global warming.
Sapsis suggests that understanding how extreme weather will change due to climate change, especially in specific locations, would greatly affect preparations and engineering solutions to deal with these events. The team’s new approach involves adjusting the model’s output rather than its dynamical equations, the latter traditionally being a cumbersome process.
The scheme uses machine learning to supplement simulation data with past information like world temperature and humidity levels. The algorithm recognizes associations within the data which provide crucial insights about weather feature dynamics, which are then used for refining the model’s forecasts. The method corrects the model’s dynamics, which helps in achieving accurate predictions of rare extreme events’ frequency.
The MIT team tested the new approach on the Energy Exascale Earth System Model (E3SM), which models global climate patterns at a resolution of 110 kilometers. The researchers trained their algorithm using eight years of past data in temperature, humidity, and wind speed, which then corrected the patterns simulated by E3SM. When superimposed with a high-resolution tropical cyclone-predicting model, they found that the refined model produced frequency predictions of extreme storms in particular locations that closely matched real-world patterns from the previous 36 years.
Sapsis acknowledges that while the differences in absolute terms between predictions from uncorrected and corrected simulations aren’t significant, these might have significant implications for humans facing such situations. He further states that understanding the dynamics of extreme weather event, in a present or future climate, becomes more improved once the corrections are applied. Even phenomena like forest fires, heatwaves, and floods can be better understood and anticipated.
The study received praise from Pedram Hassanzadeh, an associate professor at the University of Chicago. He highlights the importance of incorporating future greenhouse emissions scenarios into this framework. The U.S. Defense Advanced Research Projects Agency has provided part funding for the work.