Scientists found a surprising use for wrong AI hypotheses in climate research
Testing a proposed seasonal shift led International Potato Center scientists to a part of the historical record with weak weather-station coverage. Read the full story on CryptoSlateUnderstand the crypto stories that matter and what to watch next. Get The Catalyst in your inbox
Researchers filling a gap in Peru’s climate record had already spotted implausible dew-point estimates when an AI tool suggested a seasonal explanation. The researchers ruled it out for the affected area. Testing it led them to check how the historical data estimated conditions far from weather stations. Ai2 described the case on Nov. 19, when it launched Theorizer on its Asta research platform and released the tool’s source code.
Piero Palacios and Henry Juarez of the International Potato Center were working with PISCO, a dataset that uses weather-station observations to estimate climate conditions across a grid. Its record of dew point, an indicator of moisture in the air, ended in 2016. Filling the gap would give agricultural models an input they use to estimate how much water crops need.
PISCO’s daily minimum temperature record continued through 2020. The team used the years covered by both records to work out how minimum temperature and dew point varied together, then used that relationship to estimate dew points for 2017 through 2020. Ai2 reported that some reconstructed dew points topped 35°C, implausibly high for the affected areas, and that the extremes clustered in November.
Related ReadingTheorizer searches scientific papers and links its proposed explanations to the studies it draws on. It suggested that humidity rose as the Amazon shifted from its dry season to its wet season. The researchers tested that explanation and ruled it out for the affected area.
That test prompted a second question. Would the extremes be more common in parts of the grid with less support from weather stations? The researchers found the extreme reconstructed values outside the area where station observations constrain the historical estimates. Values above the same threshold were already more common in the historical estimates over the periods the team compared. The cause of the original November spike remains open.