In climate science, advanced computer models can now simulate thousands of possible climate futures: What happens if CO2 emissions stay high? What if environmental policies change — or don’t? How will California’s rainfall patterns shift over the next 80 years?  

But each scenario generates massive datasets, and traditional methods for analyzing them, such as averaging and statistical summaries, may miss critical patterns. 

ClimateSOM, a new AI-powered tool developed by Yuya Kawakami, a Ph.D. student in computer science at the University of California, Davis, combines interactive visualization with machine learning to help researchers navigate the overwhelming flood of data and produce more accurate analyses and better predictions. A paper on ClimateSOM was published in IEEE Transactions on Visualization and Computer Graphics.

“There is a gap between standard statistical techniques that exist in climate science and the vast amount of data being produced,” said Kawakami. “I think visual analytics and workflows can be designed to help that, and with climate change imminent, it’s imperative we do this today.” 

From Billions of Data Points to Insights

Climate models generate so much data because they run different scenarios over 40 to 100 years, adjusting variables such as CO2 emissions, deforestation, urbanization, ocean circulation patterns and potential tipping points. This creates thousands of possible futures that need to be analyzed and compared. 

Kawakami’s study used 14 models across four scenarios, yielding 56 projections totaling over 1 billion data points. That’s just a fraction of what climate scientists face globally. 

Traditional methods reduce decades of climate data to a single average. It’s like saying that California will get 10% more rain by 2100, but that average could play out in vastly different ways, from steady increases each year to dramatic swings between drought and downpour.

“Averaging data may not be the most accurate way to analyze it,” Kawakami said. “I started thinking, how can I visually encode the distributional nature of things?”

How ClimateSOM Works

A peer of Kawakami’s suggested a self-organizing map, or SOM, a machine learning tool that organizes complex data into a 2D grid. 

For his paper on ClimateSOM, Kawakami’s SOM sifted through thousands of precipitation snapshots — maps showing where it’s raining across California at different points in time — and intuitively arranged them onto a 2D screen. 

This resulted in patterns such as “heavy rain in Northern California, dry in the south” clustering in one region of the grid, and “statewide drought” clustering in another. Scientists can see at a glance which precipitation patterns are most often produced by different climate models.

Kawakami combined this mapping technique with interactive visualization, creating ClimateSOM — a framework that lets scientists explore the full range of climate possibilities rather than collapsing everything into a single average. 

The workflow begins by training a self-organizing map on climate data, creating a grid of nodes representing distinct spatial patterns. Users then organize and label regions of interest, aided by AI that can answer natural-language queries, before analyzing individual climate models, comparing scenarios side-by-side and identifying clusters of similar behavior. 

In practice, a climate scientist studying California precipitation might anchor wetter climate patterns to the left side of the screen and drier patterns to the right, with a north-south gradient from top to bottom. They might then annotate a region in the upper left as "Wet Sierra Nevada winters." When they run the analysis, they can see which emission scenarios push models into that region and which models behave differently from the rest.

Finding Patterns Traditional Methods Miss

Working with climate scientist Daniel Cayan from Scripps Institution of Oceanography, Kawakami tested ClimateSOM on precipitation projections for California and the Pacific Northwest. The tool revealed patterns that would be nearly impossible to spot using traditional methods.

For instance, ClimateSOM identified two distinct clusters of climate models predicting California's January precipitation. Models forecasting a wetter January were linked to a specific pattern the previous November: overall dryness with particular aridity in Northern California. Kawakami presented this work to several climate scientists, and their response was enthusiastic. 

As AI-powered climate models become more sophisticated, they’ll produce exponentially more data. Tools like ClimateSOM may become essential for extracting meaningful findings from the flood of information. 

“There’s no problem more pressing or more important to think about than climate change,” Kawakami said. “I would encourage anyone who’s interested to pursue this direction, to explore ways to bolster our analysis tech capabilities, especially within climate science, because it’s needed.”