Machine learning improves the accuracy of climate models – especially for complex extreme events

Researchers have devised new machine learning methods to improve predictions for large-scale climate models, demonstrating that new tools make models more accurate at both the global and regional levels. This advancement should provide policymakers with improved climate forecasts that can be used to inform policy and planning decisions.

The paper, “Complete density correction using CMIP6 GCMS normalization flow (CDC-NF)” is published in the journal Science data.

“While global climate models are essential to policy planning, these models often wrestlew with “combined extreme events.” This is when extreme events occur in a short period of time. This can be when extreme rain is quickly followed by periods of extreme heat.”

“Specifically, these models struggle to accurately capture observed patterns regarding the combined events of the data used to train the models,” says Fang. “This creates two additional issues: it is difficult to provide accurate predictions for complex events at a global scale, and it is difficult to provide accurate predictions for complex events at a local scale. The work we did here addresses all three of these challenges.”

“All models are incomplete,” says Sankar Arumugam, author of the paper and professor of civil, construction and environmental engineering in North Carolina. “Models can underestimate rainfall or overestimate temperatures. Model developers have a set of tools they can use to correct these so-called biases, which will improve the accuracy of the model.

“However, there are important limitations to existing tool suites. While it's very good at fixing defects in a single parameter (such as rainfall), it's not very good at fixing defects in multiple parameters (such as rainfall and temperature),” says Arumugam. “This is important because complex events can pose serious threats and, by definition, can absorb the social impact of two physical variables: temperature and humidity. This is where new methods come into play.”

This new method takes a new approach to the problem and uses machine learning techniques to change the output of the climate model by moving the projection of the model into patterns that can be observed in real-world data.

Researchers tested a new method (fully density correction using normalization flow (CDC-NF)) using five most widely used global climate models. The test was conducted both worldwide and nationwide in the continental United States.

“The accuracy of all five models was improved when used in conjunction with the CDC-NF method,” Fang says. “And these improvements were particularly pronounced in terms of accuracy for both isolated and complex extreme events.”

“We have published our published code and data, allowing other researchers to use our methods in conjunction with modeling efforts and further modify the way we meet our needs,” says Arumugam. “We are optimistic that this can improve the accuracy of the forecasts used to inform climate adaptation strategies.”

This paper was co-authored by Emily Hector, an assistant professor of statistics in NC. Brian Reich and Gertrude M. Cox distinguished between NC professors of statistics. Reetam Majumder, assistant professor of statistics at the University of Arkansas.