Project Summary

Climate change is a known threat multiplier, but its impacts have remained difficult to capture with high resolution, accuracy, and confidence due to the complexity and dynamics of the interconnected Earth systems involved. Known and unknown climate tipping points will exacerbate environmental change and have catastrophic impacts both regionally and globally. As climate science advances, the ever-increasing complexity of climate models presents growing challenges in computational resources, model disagreement, and explainability. Furthermore, methods for understanding causality and anticipating tipping points are insufficient. To overcome these challenges, we are working on a developing a Physics-informed AI Climate Model Agent Neuro-symbolic Simulator (PACMANS) for Tipping Point Discovery. Our approach includes a neuro-symbolic AI simulated environment where agents explore a family of surrogate climate models with the goal of discovering conditional scenarios that lead to climate tipping points. Explainability will be addressed with a causal representation of the scenarios that are generated as part of this adversarial game. Finally, a new neuro-symbolic language will enable translation between “what if” questions and the simulated environment. This approach pushes climate modeling into the third wave of AI by extending explainability, and by informing climate intervention decisions or adaptation strategies. This first of its kind comprehensive approach to climate and tipping point modeling will be enabled by a partnership between Johns Hopkins University Applied Physics Laboratory (APL) researchers who excel in integrating complex AI systems to achieve scientific and defense-oriented results and world-class experts in dynamical systems and climate systems modeling from Johns Hopkins University (JHU) and expertise in the area of hydrology and machine learning from UC Davis.

Code https://github.com/JHUAPL/PACMANs - Public to all (Up to date with all code from start to MS 3 deliverables)

Documentation https://pacmans.readthedocs.io/en/latest/

Datasets https://data.idies.jhu.edu/PACMANS/