Advancing Decadal Predictions by Optimally Detecting Differences in Causal Relations

Project Summary: The goal of this project is to improve climate predictions and advance our process-level understanding of the ocean and atmosphere on interannual to multi-decadal time scales. To accomplish this goal, we propose to test the consistency of models and observations using a comprehensive, multivariate framework, and then construct a multi-model prediction system based on the subset of models whose internal predictability and climate change signals are consistent with observations. By selecting only models whose variabilities, and presumably physics, are consistent with observations, the resulting multi-model predictions are expected to perform better than predictions based on models with inconsistent variabilities. Also, the nature of the model inconsistencies will be diagnosed in detail. The climate models analyzed in this project will come primarily from the CMIP5/CMIP6 archive. We will select variables that can be validated observationally, such as sea surface temperature, salinity, height, and sea-level pressure, and will focus on the Atlantic and Pacific Oceans, areas where there are outstanding questions regarding the correct physics. Although numerous inconsistencies may be found in any model, inconsistencies in variables that have strong causal links to internally predictable components are the most problematic for prediction. Therefore, variables with the strongest causal relations with predictable components will be identified. Optimization techniques will be used to find the combination of variables and spatial and temporal information that (1) maximizes predictability, (2) maximizes the causal relation to that predictability, or (3) maximizes the ability to discriminate between models. Inconsistencies revealed by this analysis will elucidate how differences in process-level mechanisms between models impact internal variability and predictability. Simple or dynamically-meaningful metrics of these inconsistencies will provide model developers with new tools for model evaluation that will be of immediate relevance to improving predictability. To compare model internal variability to observations, the model’s climate change signal will be removed from observations using optimal fingerprinting techniques. If the internal variabilities are consistent, then they can be pooled in a multivariate test for the consistency of climate change signals between models and observations. If inconsistencies in climate change signals are found, then these will be diagnosed in simple or dynamically-meaningful ways. Models that are found to be consistent in both their internal predictability and climate change signals will then be combined to construct a multi-model prediction system. Empirical prediction models for multi-year prediction will be derived and used to make multi-year predictions of Atlantic and Pacific sea surface temperatures.Relevance to Competition: This proposal responds directly to all priorities in the funding call. Specifically, the proposed research will rigorously quantify “how well our models perform at simulating” decadal climate variability. The proposed research will “improve climate models and predictions” by explicitly constructing a multi-model prediction based on a subset of models whose predictability and causal relations are consistent with observations. The optimized discriminant functions derived in this research will provide a new tool to “enhance our process-level understanding of the climate system” that will be of immediate relevance to improving predictability. We will specifically examine predictability of the Pacific and Atlantic Oceans, and analyze both observations and the CMIP5/CMIP6 data set, consistent with the funding announcement.