Scientists from MIT have developed a technique that helps to fine-tune predictive models for extreme weather events by combining machine learning and dynamical systems theory. Currently, climate models are run decades and even centuries in advance to assess a community’s risk to extreme weather but these generally operate at a rough resolution. As a result, they provide a broad overview for larger regions, for instance, the northeastern U.S., without offering specific predictions for particular locations, such as Boston. To cater to specifics, policymakers usually merge a coarser model’s large-scale predictions with a finer-resolution model, which makes a more accurate estimation of how often Boston is likely to experience damaging floods due to climate change. However, the effectiveness of this risk analysis is dependent on the predictions from that initial, coarser climate model.
Themistoklis Sapsis and his colleagues have come up with a method to “correct” the predictions from coarse climate models, employing a new correction scheme for any global climate model. Here, machine learning and dynamical systems theory are utilized to refine a climate model’s simulations into more realistic patterns over larger scales. When combined with smaller-scale models to predict specified weather events like tropical cyclones or floods, the team’s approach delivered more accurate predictions for how often specific locations will experience these events over the next few decades.
The team’s new approach addresses a model’s output and superimposes an algorithm that nudges the simulation towards something that more closely represents real-world conditions. The algorithm, based on a machine-learning scheme, takes in past data for temperature and humidity worldwide, and discerns associations within the data that represent fundamental dynamics among weather features, correcting a model’s predictions with these learned associations. For example, this method can correct how an extreme weather feature, like wind speeds during a Hurricane Sandy event, will look like in the coarse model versus reality, leading to more accurate statistics for the frequency of rare extreme events.
As a pilot test of this new approach, the team used the machine-learning scheme to correct simulations produced by the Energy Exascale Earth System Model (E3SM), a climate model run by the U.S. Department of Energy, simulating global climate patterns at a resolution of 110 kilometers. Through this, the corrected version yielded climate patterns that closely aligned with real-world observations over the last 36 years. The corrected coarse model then combined with a specific, finer-resolution model of tropical cyclones, accurately reproduced the frequency of extreme storms in specific locations worldwide.
