Reliable AI-Driven Climate Modeling for more Trustworthy Climate Projection (Invited Presentation)

Remarkably, AI-based weather forecasts based on high-tech versions of analogue forecasting now outperform leading physics-based weather forecast models, and are rapidly being operationalized to make official forecasts affecting lives and property. Similar physics-based weather models form the atmospheric component of climate models, and analogous AI can be applied to climate modeling, already demonstrating considerable potential including 100-1000x speedups and superior performance at coarse horizontal grid resolutions. However several new challenges emerge that don’t fit into standard explainable/interpretable AI methodologies.

First, future climates are unseen: training the AI on historical observations is insufficient, while training on climate model simulations across a broad variety of climates inherits all their biases. Emulators should be easily trained so that most flagship climate models (and nonstandard models like km-scale climate models) have an emulator version, so that projection uncertainty can be estimated.

Second, climate is a coupled earth system process driven by feedbacks between atmosphere, land, ocean, cryosphere, biosphere and human societies. At a minimum, one must accurately emulate the land, ocean and ice/snow as well as the atmosphere, over a broad range of space and time scales, to make useful centennial climate projections. This is a complex AI work in progress.

Third, AI does not necessarily guarantee conservation principles for energy, water and dry air mass that help maintain climate equilibrium - it is nontrivial to build in key physical priors into AI architectures while retaining sufficient expressiveness to enable skillful model training.

Fourth, since AI for climate modeling must be relied on in as-yet unobserved climates, new protocols for robustly testing out-of-sample generalization are essential - the CMIP process provides a template for this.

I believe that all these challenges will be surmounted within the next five years, and that compelling advantages of AI for climate modeling will make it an accepted part of the standard repertoire of climate model development, complementing physics-based climate models.