We're much better at following weather patterns today than a couple hundred years ago, but we're still relatively green when predicting extreme weather. Today's community-specific extreme weather forecasts are based on coarse climate models, which are far more capable of anticipating climate events for vast swaths of land than for particular cities or towns. For storm readiness or other forms of extreme weather preparation, this low-res take on meteorology could have harrowing implications for a community's infrastructure or population. Engineers at MIT have found a way to make modern weather predictions more community-specific.
Rather than trying to improve the stability of existing climate models, the MIT team created a non-intrusive machine learning algorithm that sits on top to add depth, like a magnifying glass. Their solution is described in a paper published Tuesday in the Journal of Advances in Modeling Earth Systems. The algorithm starts by taking in data from past weather conditions and events: temperatures, humidity, precipitation, and so on. Then, it compares that data with predictions made by coarse climate models for those periods. Depending on how the predictions and the after-the-fact weather results match up, the algorithm builds associations that allow it to enhance predictions made by climate models in the present.
In one experiment, the engineers corrected and enhanced simulations from the US Department of Energy's Energy Exascale Earth System Model (E3SM). Like most large-scale climate prediction models, E3SM can only simulate weather patterns down to a resolution of 100 to 110 kilometers (about 62 to 68 miles). But with eight years of confirmed climate data from the past, the team taught their algorithm new associations between E3SM and actual weather. When they reran E3SM and overlaid their algorithm, the enhanced results matched real-world weather occurrences far more closely than E3SM's original predictions.
In an MIT statement, the team clarified that a "correction" could mean just a 10-degree Fahrenheit difference between a climate model's original output and their algorithm's enhanced results. But to humans, 10 degrees makes a pretty significant difference: sweating or not sweating, for example, or getting frostbite versus being able to save our toes. This margin of error also significantly impacts severe weather events like blizzards, tornadoes, tropical cyclones, heatwaves, and fire seasons—events that can make or break entire communities or ecosystems. As the larger phenomenon of climate change exacerbates these events or weather patterns, accuracy will be vital to preparing for what's to come.