← All news
Global · Education & Research
Global·UK community memberUK·USA community memberUSA·Education & Research

MIT develops synthetic-data training method for climate AI models

MIT researchers developed a technique that trains climate emulators with less data while improving predictions for unfamiliar scenarios. The approach could reduce the time needed to build AI tools for exploring climate policy outcomes.

AI Hubs United Editorial ·

MIT researchers have developed a training method designed to make AI climate emulators more accurate and adaptable while reducing the data needed to build them, according to MIT News. The research, published in Proceedings of the National Academy of Sciences, uses an algorithm to generate synthetic emissions scenarios optimized for an emulator’s predictive performance.

Climate emulators approximate the behaviour of detailed, physics-based earth system models. While those larger models can require months of supercomputer time for a single scenario, emulators run in seconds. Building their training datasets remains costly: scientists commonly use six to eight future scenarios, each requiring a full-scale model run. According to MIT News, the new approach requires just one such run instead.

The algorithm repeatedly adjusts synthetic training scenarios to reduce prediction errors, producing rapid emissions variations rather than smooth, realistic trajectories. The researchers report that an emulator trained on a smaller, carefully selected synthetic dataset accurately predicted unfamiliar scenarios where traditional emulators struggled. For scientists assessing emissions policies or regional interventions, the practical benefit could be faster production of models that support broader scenario testing and more reliable planning.

AI Hubs United original short summary of the source linked above. It is not the original article; read the full story at the source.