Scientists led by Themistoklis Sapsis at MIT’s Department of Mechanical Engineering have developed a strategy to “correct” the predictions of coarse global climate models, enhancing the accuracy of risk analysis for extreme weather events. Global climate models, used by policymakers to assess a community’s risk of severe weather, can predict weather patterns decades or even centuries into the future, but only on a large scale.
Such models might allow predictions for broader regions like the northeastern US, but not specifically for cities like Boston. To estimate the risk of extreme weather like flooding in Boston, a coarse model’s predictions are combined with a fine-resolution model that provides estimates of how often damaging floods might occur. The accuracy of this analysis is dependent on the predictions offered by the coarse climate model.
“If you get those wrong for large-scale environments, then you miss everything in terms of what extreme events will look like at smaller scales, such as over individual cities,” notes Sapsis. MIT’s new “correction” uses machine learning and dynamical systems theory to nudge a climate model’s simulations into patterns that more accurately reflect real-world conditions. These corrected models, when combined with smaller-scale models, can offer more precise predictions for extreme weather events in specific locations in the next decades.
The “correction” is a general method which can be applied across any global climate model. Once corrected, models can provide greater clarity on when and where extreme weather will strike as global temperatures increase. This could have profound significance across a range of fields and aspects of human life, including biodiversity, food security and the economy.
These coarse climate models often simulate weather features like temperature, humidity and precipitation on a global, grid-by-grid basis. Computationally intense, these models account for weather processes every 100 kilometers or so, but can’t capture finer-scale phenomena like storms or clouds. The MIT team’s approach corrects a model’s outputs, rather than its underlying equations that describe the physical interactions in the atmosphere and oceans.
A machine-learning algorithm overlays a model’s output, nudging the simulation to better reflect real-world conditions. This algorithm learns associations within the data that represent key dynamics among weather features and uses these to correct a model’s predictions. The research was published in the Journal of Advances in Modeling Earth Systems, and supported in part by the U.S. Defense Advanced Research Projects Agency.
As a first test, the new approach was used to correct simulations produced by the Energy Exascale Earth System Model (E3SM), a climate model run by the U.S. Department of Energy. The algorithm was trained using eight years of past weather data, and then run forward in time for about 36 years with its outputs corrected to better match real-world observations. When paired with a tropical cyclone model, it was able to accurately reproduce the frequency of these storms in specific locations around the world. The corrected model can assist with understanding the likely future nature of forest fires, flooding events and heatwaves.
