Using AI to refine climate projections

For the new study, Diffenbaugh and Barnes trained an AI system to predict how high global temperatures could climb, depending on the pace of decarbonization.

When training the AI, the researchers used temperature and greenhouse gas data from vast archives of climate model simulations. To predict future warming, however, they gave the AI the actual historical temperatures as input, along with several widely used scenarios for future greenhouse gas emissions.

“AI is emerging as an incredibly powerful tool for reducing uncertainty in future projections. It learns from the many climate model simulations that already exist, but its predictions are then further refined by real-world observations,” said Barnes, who is a professor of atmospheric science at Colorado State.

The study adds to a growing body of research indicating that the world has almost certainly missed its chance to achieve the more ambitious goal of the 2015 Paris Climate Agreement, in which nearly 200 nations pledged to keep long-term warming “well below” 2 degrees while pursuing efforts to avoid 1.5 degrees.

A second new paper from Barnes and Diffenbaugh, published Dec. 10 in Environmental Research Letters with co-author Sonia Seneviratne of ETH-Zurich, suggests many regions including South Asia, the Mediterranean, Central Europe, and parts of sub-Saharan Africa will surpass 3 degrees Celsius of warming by 2060 in a scenario in which emissions continue to increase – sooner than anticipated in earlier studies.

Extremes matter

Both new studies build on 2023 research in which Diffenbaugh and Barnes predicted the years remaining until the 1.5 and 2 degrees Celsius goals are breached. But because these thresholds are based on average conditions over many years, they don’t tell the full story of how extreme the climate could become.

“As we watched these severe impacts year after year, we became more and more interested in predicting how extreme the climate could get even if the world is fully successful at rapidly reducing emissions,” said Diffenbaugh, the Kara J Foundation Professor and Kimmelman Family Senior Fellow at Stanford.

For a scenario in which emissions reach net zero in the 2050s – the most optimistic scenario widely used in climate modeling – the researchers found a nine-in-ten chance that the hottest year this century will be at least 1.8 degrees Celsius hotter globally than the pre-industrial baseline, with a two-in-three chance for at least 2.1 degrees Celsius.

For a scenario in which emissions decline too slowly to reach net zero by 2100, Diffenbaugh and Barnes found a nine-in-10 chance that the hottest year will be 3 degrees Celsius hotter globally than the pre-industrial baseline. In this scenario, many regions could experience temperature anomalies at least triple what occurred in 2023.

Investing in adaptation

The new predictions underline the importance of investing not only in decarbonization but also in measures to make human and natural systems more resilient to severe heat, intensified drought, heavy precipitation, and other consequences of continued warming. Historically, those efforts have taken a back seat to reducing carbon emissions, with decarbonization investments outstripping adaptation spending in global climate finance and policies such as the 2022 Inflation Reduction Act.

“Our results suggest that even if all the effort and investment in decarbonization is as successful as possible, there is a real risk that, without commensurate investments in adaptation, people and ecosystems will be exposed to climate conditions that are much more extreme than what they are currently prepared for,” Diffenbaugh said.