Assessing Decadal Variability of Subseasonal Predictability Using Explainable Machine Learning

The climate system is chaotic and noisy. Extracting useful information from this noise enhances the ability to make accurate weather and climate predictions. In this seminar, I will discuss innovative machine learning approaches used to explore sources of predictability within the climate system that provide forecasts of opportunity – climate states that result in higher prediction skill– which ultimately lead to more confident climate forecasts on subseasonal (2 week-3 month) timescales. Artificial neural networks, a subset of machine learning, are employed to investigate low frequency variability of subseasonal predictive skill in CESM-2 Large Ensemble climate model data. Further, eXplainable Artificial Intelligence (XAI) methods are used to understand the machine learning model’s decision making strategy to not only gauge trust in our machine learning models, but to identify new sources of predictability. Analysis of the subseasonal prediction skill of the neural networks reveals fluctuations on decadal timescales, particularly for predictions made with high confidence. The analysis is extended to observational data (ERA5), indicating the machine learning models have learned a real and detectable signal in the climate system. The drivers of this low-frequency variability are investigated to understand why certain time periods have higher subseasonal predictability to identify forecasts of opportunity. Analysis reveals the Pacific Decadal Oscillation is modulating when the midlatitudes are teleconnected to the tropics, highlighting low frequency variability of subseasonal predictable states of the climate system.