Enhancing the use of AI in the climate modeling community

The climate modeling community has been exploring ways to speed up the traditional physics based models using AI for many years. The current path includes using "Foundation Models" to emulate the physics based models because they run to completion much faster than the traditional full physics based models. A Foundation Model (FM) is a massively pre-trained Machine Learning (ML) model upon which "fine tuning" can be performed to generate a new model capable of answering specific science questions posed by our scientists. There are now at least two dozen FM available in the public domain that have been generated by large technology companies such as Google, NVIDIA, Microsoft, IBM, etc as well as individual researchers in Academia. While these models represent technological advances, they also often lack significant documentation on their use for specific types of science problems which makes it challenging for scientists to understand what each one can or should be used for. There have been some efforts at intercomparison between the models, notably the Weatherbench2 from Google, but there is still a lack of clarity in terms of Quantitative evaluation of specific science applications.

We have implemented a framework for easily incorporating new FM into a workflow that allows the fine-tuning of each model to the exact same use cases defined by a group of climate modeling scientists to facilitate a true apples to apples comparison. In this presentation we will show the framework, use cases and results for the suite of Foundation Models that were tested in our system. The implication of our results is to provide a knowledge base for the climate scientists to make informed decisions about which, if any, FM are appropriate for them to use to answer the science questions for their respective research. Essentially we are providing an information center to help jumpstart the scientists use of AI driven climate based foundation models.