AI Goes Global: Can This New Climate Model Predict the Future of Earth Faster and Cheaper?
Imagine a world where forecasting Earth’s climate could occur 26 times faster and use a fraction of the energy. This futuristic possibility is getting closer thanks to a groundbreaking new climate model developed by researchers at the Allen Institute for AI and UC San Diego.
This innovative model utilizes the power of artificial intelligence to emulate a widely used physics-based climate model, FV3GFS, currently adopted by the United States for its global weather forecasts. While FV3GFS requires nearly 80 hours to simulate a decade of climate data, this new AI-driven approach can achieve the same outcome in just over two and a half hours.
The secret to this remarkable speed lies in a technique called "Spherical DYffusion." This method adapts generative AI architectures, similar to those powering image-generating tools like Stable Diffusion and DALL-E 2, to the unique challenges of climate modeling. Instead of relying on traditional rectangular grids, the model operates on a spherical representation of the Earth, enabling it to efficiently capture complex atmospheric patterns without getting bogged down by coordinate system complexities that exist at the poles.
"The promise of deep learning in climate modeling has been uncertain given the inherent data complexity and long inference involved," the researchers explain in their paper published on OpenReview. "We describe our model as the first ‘conditional generative model that produces accurate and physically consistent global climate ensemble simulations by emulating a coarse version of the United States’ primary operational global forecast model, FV3GFS."
While acknowledging the significant speed and efficiency gains, the researchers highlight that the model isn’t perfect. In its quest for efficiency, the model exhibits slightly higher climate biases compared to the traditional FV3GFS model.
Rose Yu, a faculty member in the UC San Diego Department of Computer Science and Engineering, and Ph.D. student Salva Ruhling Cachay examine data. [David Baillot/UC San Diego Jacobs School of Engineering]
"The model’s ensemble-mean predictions reduce these biases by 29.28% through averaging, the model still falls short of matching FV3GFS’s theoretical minimum uncertainty threshold," they explained.
This limitation stems from the model’s reliance on data alone, without directly incorporating the physical constraints found in traditional models.
Despite these challenges, this novel approach represents a significant leap forward in climate modeling. As researchers continue to refine this AI-powered tool, it holds immense potential for revolutionizing our understanding of Earth’s climate and informing critical decisions about our planet’s future.
/cdn.vox-cdn.com/uploads/chorus_asset/file/25770485/15ae9897b2ae3c7edda1e072ee5d1376d8b4a211ccedda6afa501ec3690f3294.jpg "GOG’s preservation program lets you keep playing games after they’re delisted")
