We describe a new multidisciplinary effort to understand how Artificial Intelligence (AI) could be used to improve climate tipping point discovery using the collapse of the Atlantic Meridional Overturning Circulation (AMOC) as a case study. Our methodology includes an AI simulated environment where a Generative Adversarial Network (GAN) is used to play an adversarial game between two deep networks. One network, the generator, tries to invoke an AMOC collapse tipping point. The other network, the discriminator, tries to predict which model configurations lead to an AMOC collapse, and which do not. The discriminator uses a climate surrogate model to run model configurations suggested by the generator network. In this study we have explored using a well-researched reduced oceanography four-box model (Gnanadesikan et al. 2018) as the surrogate. In addition, we introduce a new type of surrogate model that is used for expanding to larger models, which is based on modeling stochastic differential equations to estimate bifurcation escape times, enabling a deeper understanding of the unstable areas and how to escape them. We describe how we apply this method to larger Global Circulation Models (GCMs) by means of a CESM2-calibrated dataset. Included in this methodology is a neuro-symbolic representation of the model configurations that translates into natural language questions enabling climate modelers to ask questions of the learned adversarial space and to provide explainability. We describe the results of applying this method to study the AMOC, and evaluate our method by comparing how well it learns to predict AMOC collapse, how well it learns the area of uncertainty in state space, and how effective the neuro-symbolic language is in providing a natural language interface into the adversarial game.@en

We propose a hybrid Artificial Intelligence (AI) climate modeling approach that enables climate modelers in scientific discovery using a climate-targeted simulation methodology based on a novel combination of deep neural networks and mathematical methods for modeling dynamical systems. The simulations are grounded by a neuro-symbolic language that both enables question answering of what is learned by the AI methods and provides a means of explainability. We describe how this methodology can be applied to the discovery of climate tipping points and, in particular, the collapse of the Atlantic Meridional Overturning Circulation (AMOC). We show how this methodology is able to predict AMOC collapse with a high degree of accuracy using a surrogate climate model for ocean interaction. We also show preliminary results of neuro-symbolic method performance when translating between natural language questions and symbolically learned representations. Our AI methodology shows promising early results, potentially enabling faster climate tipping point related research that would otherwise be computationally infeasible.@en

In 2018, the IPCC summarized in a special report the potential risks surrounding climate tipping point. In 2019, Lenton et al. highlighted climate tipping points that could contribute to irreversible changes to our world including ice melt, deforestation, and circulation slowing. In this work we describe a study of the Atlantic Meridional Overturning Circulation (AMOC), and the use of an Artificial Intelligence (AI) assisted climate model methodology capable of performing unsupervised tipping point discovery to discover conditions which could lead to AMOC collapse.

We developed a novel AI generative adversarial network (GAN), where a set of deep learning generators attempt to invoke AMOC collapse by perturbing a constrained set of parameters, while another deep learning network, the discriminator, tries to learn how to avoid AMOC collapse. A surrogate model is used to run model configurations to test this adversarial method. We show that our methodology can be used to discover areas in model space that are consistent with fold bifurcations where the system moves from an on state to an off state. We measured the performance of this method by comparing it to an AMOC four box model and experiments described in (Gnanadesikan 2018) which uses the four box model to understand overturning stability.

We have found that the deep learning method can be used to exploit the area of uncertainty that is consistent with the area that separates the two stable states in a fold bifurcation model. When we compared the results of the adversarial network to the Gnanadesikan experiments we observed that when incorporating information regarding the uncertainty in the loss function, increasing the number of AI generators caused the AI agents to become more focused on this area of uncertainty. This area of uncertainty is consistent with what is described as the separatrix. In this study, we show the benefit of using this novel unsupervised approach as part of an AI assisted climate modeling methodology.@en