Researchers at the Massachusetts Institute of Technology (MIT) have developed a new method to improve the accuracy of large-scale climate models. These models, used by policymakers to understand the future risk of extreme weather like flooding, often lack precise data for smaller scales without considerable computational power. By combining machine learning with dynamical systems theory, the team’s approach can refine these models, enhancing predictions for certain weather events. The approach offers more accurate forecasts for the frequency of such events at specific locations over coming decades.
Themistoklis Sapsis, director of the Center for Ocean Engineering in MIT’s Department of Mechanical Engineering, explains that their technique involves ‘correcting’ the models, rather than trying to fix their underlying dynamical equations. The correction scheme is general, and can be applied to any global climate model to ascertain where and how often extreme weather will occur as global temperatures rise.
The research team developed an algorithm, based on a machine-learning scheme, which adjusts the simulation to better resemble actual global conditions. The method uses past data for temperature, humidity, and other weather features, and learns the associations within that data which represent the dynamics among features. From this, the algorithm corrects a model’s predictions, improving the accuracy of dynamical features such as wind speeds during a hurricane event. The team says correct dynamics leads to correct statistics and better frequencies for rare extreme events.
Researchers first applied the machine-learning scheme to the Energy Exascale Earth System Model (E3SM), a climate model run by the U.S. Department of Energy. They trained the algorithm with eight years of historical data for temperature, humidity, and wind speed. They then applied the trained algorithm to E3SM’s future simulations and found that the corrections produced climate patterns that more accurately reflected the actual weather in the past 36 years. Following further tests that paired the corrected large-scale model with a smaller-scale model of tropical cyclones, the approach accurately replicated the frequency of extreme storms in specific locations around the world.
The refined model effectively predicts the frequency of events for the current climate, and will be useful for predicting the impacts of forest fires, floods, and heat waves in a future climate. Future work is now focused on analyzing future climate scenarios. Given the promising results, Pedram Hassanzadeh, an associate professor who leads the Climate Extremes Theory and Data group at the University of Chicago, is keen to see the frameworks projections once future greenhouse gas emissions are incorporated. The study is published in the Journal of Advances in Modeling Earth Systems. Funding support came in part from the U.S. Defense Advanced Research Projects Agency.
