Abstract: We present results from a DARPA-funded project on AI-assisted climate modeling for tipping point discovery. In this work, we use a Generative AI method that is trained to generate reduced model parameters of the Atlantic Meridional Overturning Circulation (AMOC) that will cause the AMOC to collapse. The tipping point generative adversarial network (TIP-GAN) appears to model the state space bifurcations by working in the parameter space of the reduced model. A bidirectional translation is used to learn a mapping between scientific questions and a neuro-symbolic representation used by TIP-GAN, enabling scientists a way to interrogate the learned model using natural language.

We show how TIP-GAN, applied to a reduced model of the AMOC, was able to learn the contours of a fold bifurcation based on perturbing three parameters described in an experiment in a 2018 Gnanadesikan et al. paper. In other experiments, we show that TIP-GAN can be a sufficient replacement for the reduced climate model with F-Measure scores above 90% in classifying models as those which will tip and those which will not tip. We also show that by including the neuro-symbolic bi-directional translation, scientists are able to ask natural language scientific questions that can be answered by TIP-GAN without running the reduced climate model, offering a significant\ speed-up in the scientific process.

Bio: Jennifer Sleeman is a Senior AI Researcher at the Johns Hopkins University Applied Physics Laboratory and a Research Assistant Professor at the University of Maryland, Baltimore County. Jennifer has been an active AI researcher since the early 2000s and holds a Ph.D. in Computer Science from the University of Maryland, Baltimore County (UMBC).