A new report by the MIT Climate Grand Challenges team reveals that simple, science-based climate prediction models can outperform deep-learning approaches when predicting future temperature changes. Deep learning does have great potential for estimating more complex variables like rainfall. The study was reported by Adam Zewe in the August 26th issue of MIT News. Zewe writes that AI models struggle to predict local temperature and rainfall:
“Environmental scientists are increasingly using enormous artificial intelligence models to make predictions about changes in weather and climate, but a new study by MIT researchers shows that bigger models are not always better…
The researchers developed a more robust way of evaluating…techniques…showing that simple models are more accurate when estimating regional surface temperatures…
They used these results to enhance a simulation tool known a climate emulator, which can rapidly simulate the effect of human activities onto a future climate.”
Is this a cautionary tale about the value of deploying wholesale AI models for climate science? The researchers think so.
Read the report.