MIT researchers develop faster and more reliable AI models for climate prediction
An MIT-led team has developed a new technique that accelerates the training of AI models that simulate climate scenarios. By achieving high accuracy with less synthetic data, the method reduces the development time of climate policy tools from months or years to a much shorter duration.
MIT researchers have developed a new training technique that simplifies the production of AI models known as climate emulators, which can simulate the effects of human activities on the climate in seconds.
The method, developed by a team led by Christopher Womack and Noelle E. Selin, uses an algorithm that generates optimized synthetic training data. This approach enables models to more accurately predict scenarios not seen during training, outperforming traditional models.
The study, published in the journal Proceedings of the National Academy of Sciences, brings together researchers from MIT, ETH Zurich, and Imperial College London. The new technique aims to provide more reliable tools for policymakers by reducing the development time of climate emulators from months or years to a much shorter duration.
Sources
- news.mit.edu · Link