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]
A team of researchers at the Allen Institute for AI and UC San Diego have introduced a climate modeling approach boasting 25× speed-ups and associated energy savings over the emulated physics-based model FV3GFS. Specifically, the method takes about 2 hours and 56 minutes to run a 10-year simulation The physics-based FV3GFS model takes about 78 hours and 4 minutes to run the same simulation, making the novel approach a 26.6 times faster than modeling with FV3GFS.
The model doesn’t combine generative AI with physics in the sense of a hybrid model. Instead, it’s a generative AI model trained to emulate FV3GFS, learning from that model’s simulation data. The physics comes in through the training data, not through direct incorporation of physical equations into the model architecture. This architectural choice enables significant computational efficiency while maintaining physical consistency.
The model adapts diffusion-based architectures for climate modeling through what the researchers term “Spherical DYffusion.” This architecture implements a conditional diffusion process—similar to those powering Stable Diffusion’s and DALL-E 2’s image generation and AlphaFold 3’s protein structure prediction—but operating directly on spherical geometries rather than traditional rectangular grids. The diffusion process begins with a noise distribution and progressively refines it through learned denoising steps, guided by physics-informed conditioning signals from FV3GFS training data. Like RFdiffusion’s approach to protein design, the model achieves its efficiency through a optimized spherical convolution operator that preserves rotational equivariance while processing global patterns. This architectural enables the model to capture complex spatial dependencies across the Earth’s surface without the computational overhead typically associated with managing coordinate system discontinuities at poles.
for climate modeling through what the researchers term “Spherical DYffusion.” This architecture implements a conditional diffusion process operating directly on spherical geometries, departing from traditional rectangular grid representations. The diffusion process begins with a noise distribution and progressively refines it through learned denoising steps, guided by physics-informed conditioning signals from FV3GFS training data. Notably, the model achieves its efficiency through a carefully optimized spherical convolution operator that preserves rotational equivariance while processing global atmospheric patterns. This architectural choice enables the model to capture complex spatial dependencies across the Earth’s surface without the computational overhead typically associated with managing coordinate system discontinuities at poles. When compared to standard diffusion models, the spherical adaptation demonstrates superior computational efficiency, requiring approximately 25 times fewer resources while maintaining physically consistent predictions across global climate patterns.
The researchers note in the paper summarizing the research, published on OpenReview, that the promise of deep learning in climate modeling has been uncertain given the inherent data complexity and long inference involved. They describe their 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.”
The model’s performance metrics highlight trade-offs. While optimizing for computational efficiency gains, the model exhibits average climate biases approximately 50% higher than the reference physics-based model across all 34 predicted fields. This accuracy gap widens for certain atmospheric variables, especially at higher altitudes and in polar regions where complex atmospheric dynamics pose greater challenges. While 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.
These limitations stem partly from the model’s purely data-driven architecture. It lacks explicit physical constraints that traditional models incorporate through fundamental equations. Additionally, the current implementation relies on annually repeating climatological sea surface temperatures and fixed greenhouse gas concentrations, restricting its ability to model climate change scenarios.
