Policymakers rely on global climate models to assess a community’s risk of extreme weather. These models, run decades and even centuries forward, gauge future climate conditions over large areas but have a coarse resolution and are not definitive at the city level. To remedy this overlap, they may combine predictions from a coarse model with a finer-resolution model to estimate risk of extreme weather such as flooding in specific places, like Boston. The final risk analysis, however, is as accurate as the initial more general climate model.
Themistoklis Sapsis, a professor and director at MIT, and his colleagues have developed a method to refine the predictions from broad climate models. They combined machine learning with dynamical systems theory to develop a model whose predictions closely resembled real patterns. This was found to give more accurate predictions for the frequency of specific events like cyclones or floods over the next decades in specific places than predictions without corrections.
Sapsis stated that climate change would have an effect on all aspects of life, so if we had the capabilities to predict accurately how extreme weather would change at specific locations, it would make a significant difference in preparation and providing solutions. This method opens the avenue to do so.
Today’s large-scale climate models simulate weather features such as temperature, humidity, and precipitation worldwide on a grid-by-grid basis. Simulations of these models require significant computing power and average features every 100km or so to estimate how weather features will interact over the years. To enhance these models, scientists have attempted to correct the models’ underlying dynamical equations describing how atmospheric phenomena should interact.
Sapsis and his team took a different approach, correcting the model’s output rather than the equations. They used an algorithm that matches the output to real-world conditions. This algorithm, a machine-learning scheme, takes in data like temperature and humidity and learns the fundamental dynamics of weather features. These learned dynamics are then used to correct a model’s predictions.
The team tested their method using data from the Energy Exascale Earth System Model, a model from the U.S. Department of Energy, which simulates climate patterns globally at 110km resolution. They trained the model using eight years of data and applied it to simulations run forward in time for about 36 years. The corrected versions provided climate patterns that closely resembled real-world observations from the last 36 years not used for training. When the corrected model was then paired with a finer-resolution model of tropical cyclones, it accurately reproduced the frequency of extreme storms in specific global locations.
This correction, Sapsis says, can be employed for understanding how forest fires, flooding events, and heatwaves will look under future climate situations. The work was partially supported by the U.S. Defense Advanced Research Projects Agency.
