Urgent.News

What's breaking now, across hundreds of outlets.

AI

Unrealistic emissions scenarios help AI climate models prepare for unfamiliar futures

Accurate, up-to-date climate projections are essential for planning decisions that enhance the resilience of power grids, protect public health, maximize agricultural yields and more. Now, MIT researchers have developed a technique that streamlines the production of fast, trustworthy and adaptable artificial intelligence models for climate prediction.

Unrealistic emissions scenarios help AI climate models prepare for unfamiliar futures

Accurate climate projections are crucial for planning decisions that enhance grid resilience, protect public health, increase agricultural yields and more. MIT researchers have created efficient artificial intelligence models called climate emulators to simulate the effects of human activities on future climates, predicting impacts on temperatures or sea levels quickly. These emulators run in seconds, allowing scientists to test numerous scenarios and identify unexpected outcomes of regional interventions.

MIT researchers developed a training technique that enhances climate emulator performance and adaptability while using considerably less data, significantly reducing production time. They discovered that an emulator trained with smaller, carefully selected synthetic data accurately predicts outcomes of scenarios not seen during training, while traditional emulators fail to adapt to new situations.

Climate emulators excel in modeling policy-relevant scenarios. By generating optimized training data for an emulator using the method, the production of reliable emulators can be accelerated, boosting trust in their outputs for policy decisions. Christopher Womack, lead author and postdoc at ETH Zurich, works with Shahine Bouabid, Andrei Sokolov, Popat Salunke, Glenn Flierl, Sebastian D. Eastham, and senior author Noelle E. Selin on this technique.

Traditionally, Earth system models (ESMs) project long-term global temperature or precipitation changes in response to varying greenhouse gas emissions. However, these models are resource-intensive, taking months to run on supercomputers for a single scenario. Selin and her team developed climate emulators, AI surrogate models that mimic full-scale ESM behavior, helping scientists test many scenarios, make accurate forecasts, and explore specific pollutants and regional effects.

Climate emulators run in a few seconds, providing scientists with expanded capabilities to explore future climate scenarios. When designing climate emulators, researchers typically train the AI model with realistic future scenarios based on emissions. However, the MIT researchers found that using such scenarios as training data can result in emulators that underperform, failing to predict unfamiliar or novel scenarios.

The researchers developed an algorithm that generates synthetic data sets of optimized climate scenarios tailored to boost one emulator's predictive performance. These synthetic scenarios, containing rapid variations, enable emulators to extract more information about the overall system. The training algorithm iteratively updates training data to increase the emulator's predictive skill, generating high-frequency wiggles and oscillations in emissions scenarios.

In the future, Womack anticipates encountering climates and scenarios beyond current predictions, making robust simulation more critical to provide reliable information to stakeholders.

Rewritten by urgent.news; may contain errors — check the original.

Source: phys.org

Read the original at phys.org →

Report a problem with this story

More in AI

More from Thursday 12 November →