MIT develops training technique to improve AI climate models

MIT develops training technique to improve AI climate models


Researchers at the Massachusetts Institute of Technology (MIT) have developed a training method that could make artificial intelligence (AI)-based climate prediction models faster.

According to MIT researchers, the technique is designed to improve climate emulators – specialised AI models that simulate how human activities, including greenhouse gas emissions and pollution, affect future climate conditions such as temperature and sea levels.

Climate scientists typically rely on Earth system models (ESMs), which use complex physical simulations to project changes in global temperatures and precipitation under different emissions scenarios. However, running these models can take months on supercomputers, even when examining a single possible future.

On the other hand, climate emulators offer a faster alternative, producing simulations within seconds and allowing researchers to examine thousands of possible scenarios. However, conventional emulators can struggle to predict outcomes that differ significantly from the scenarios used to train them.

To address this limitation, the MIT team developed an algorithm that generates optimised synthetic climate scenarios for training AI models. The method improves an emulator’s ability to predict unfamiliar situations while requiring substantially less training data.

Rather than running a full-scale climate model six to eight times to generate training datasets, researchers using the technique would need to run it only once. The algorithm then adjusts the synthetic data to improve the emulator’s predictive performance.

The researchers also said that emulators trained using the method could predict scenarios they had not encountered during training. 

These included situations in which regional policies reduce emissions of a particular pollutant while other emissions continue to rise. Traditional emulators evaluated in the study failed to capture these effects.

“Climate emulators are especially well-suited for modelling policy-relevant scenarios,” said Christopher Womack, the study’s lead author and a postdoctoral researcher at ETH Zurich. The research was published in the Proceedings of the National Academy of Sciences journal on November 9.



Source link

Leave a Reply