Dr. Laura Mansfield’s research into machine learning and climate modeling revealed promising insight into the sources and quantifiable influence of uncertainty in long-term predictions, potentially paving the way for AI to create more reliable and more precise models of our future climate.

Machine learning is revolutionizing climate modeling, but it also boosts the creation of a byproduct: uncertainty. As a new generation of models run exponentially faster and handle complex data to generate precise predictions, Dr. Laura Mansfield is drawn to the role that uncertainty plays. She studies the integration of AI into climate models through her Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship, exploring how a deeper understanding of uncertainty can bring Earth’s future into clearer focus.

Mansfield explains climate models in simple terms: They predict long-term conditions, including temperature and rainfall in particular regions, by estimating the current state of the climate and then applying known physics equations to step forward in time. As she thought about ways to improve the models, especially through machine learning, Mansfield says, “I became interested not just in looking at predictions, but in trying to predict uncertainties.”

Predecessors who developed conventional, fully physics-based models paid close attention to uncertainty, Mansfield says, but she didn’t see the same focus on quantifying uncertainty in the development of AI-driven climate models. To do that, scientists need to examine the source—generally, the data or the model. Data uncertainty typically stems from the training of a machine learning model to build a dataset, while model uncertainty stems from the way the model handles data, often complicated by combining multiple models to generate predictions.

While cautioning that her research hasn’t yet been applied to full climate models, Mansfield’s investigations have yielded promising findings. “It looks like data uncertainty is more important on shorter time scales. Model uncertainty comes in more on longer time scales. If that holds true in a full, complex climate model, it would tell us where to focus our efforts.”

Mansfield also sees opportunities to apply modeling approaches in new ways, inspired by working with a model that applies physics- and AI-based methods to different tasks. “We use physics for large-scale processes, but then we use machine learning for small-scale processes. But we could use physics for the small-scale processes, AI for the large-scale.”

Support from Schmidt Sciences includes access to the computational resources necessary to train and run models, but Mansfield says she’s found as much value in her fellowship’s connections to other researchers. Organized meetups and even lunches have become regular forums to discuss AI with peers working on different problems across disciplines. Within her own group, Mansfield has noticed more people thinking about uncertainties in climate models that combine machine learning and physics. 

Mansfield has formally presented her research at conferences in Cambridge, Vienna, and Washington, D.C. She expects to submit a paper on her work for peer review later this year, which will also be published on ArXiv. She views these as preliminary steps, though. Ultimately she hopes her findings will influence others in her field to leverage the probabilistic capacities in AI-based models, in concert with physics-based models, to better understand uncertainty and provide a more reliable forecast of how local communities will be impacted by a changing climate.