New research could lead to faster, trustworthy AI models for climate prediction

illuminem summarises for you the essential news of the day, reviewed by our editorial team. Read the full piece on MIT News or enjoy below:

🗞️ Driving the news: MIT researchers have developed a new training technique that makes climate AI models faster to produce, more adaptable and potentially more reliable
• The method uses carefully optimized synthetic data to train climate emulators capable of predicting unfamiliar scenarios, addressing a major limitation of existing approaches

🔭 The context: Traditional Earth system models (ESMs) simulate climate processes in considerable detail but can require months of supercomputing time for a single scenario
• Climate emulators replicate these models in seconds, enabling scientists to explore thousands of possible futures, although their training typically relies on only six to eight costly emissions scenarios
• The MIT team found that conventional training data can leave emulators poorly equipped to predict conditions beyond those represented during training

🌍 Why it matters for the planet: More reliable climate emulators could improve assessments of emissions pathways, regional pollution controls and climate policies
• Faster development and updating of these tools could also provide policymakers and communities with more timely climate information for resilience planning and mitigation decisions

⏭️ What’s next: The researchers aim to test their approach using emissions scenarios generated by full Earth system models and apply it to other climate modeling systems. Their longer-term goal is to make updated, robust climate emulators more accessible to researchers and practitioners.

💬 One quote: “In the future, we are definitely going to see climates and scenarios we haven’t seen before. If we can simulate these different scenarios more robustly, we can provide more reliable information to the stakeholders who need it,” — Noelle E. Selin, MIT professor and senior author

📈 One stat: Climate emulators can run in seconds, compared with the months required by full Earth system models to simulate a single future scenario

Subscribe to our free newsletters to never miss a beat in sustainability.

The world needs sustainability knowledge. At illuminem, no interest group or shareholder can influence our work. Thank you for supporting our mission to make high-quality and independent sustainability information free for all. Every contribution helps. Thank you for donating today.