The principal goal of the project is to improve the spatiotemporal characterization of El Nino Southern Oscillation (ENSO) and Pacific and Atlantic multidecadal variability (e.g., PDO, AMO) and objectively assess their evolution, links, and regional climate impacts in historical simulations generated by the Energy Exascale Earth System Model (E3SM, Large Ensemble). Mode characterization and lead-lag links will be improved using machine-learning and evolution-centric analysis of the observed records, yielding refined surface-subsurface structures and regional hydroclimate footprints – the simulation targets.
The project's scope is expansive, not by choice but because of inter-basin links in multidecadal variability. It seeks answers to RGMA's Research Questions #1 (Feedback/ interaction between modes) and #4 (Representation of modes and related teleconnections) and will exploit the gained understanding in developing quantitative metrics for assessing the realism of multidecadal variability in historical simulations – metrics, currently unavailable in the PCMDI Metric Package. The proposal is well-aligned with the community-based guidance (AI4ESP's) on machine learning methods for climate analysis.
Project Description: Several spatiotemporal modes constitute ENSO variability, whose simulation and prediction remain challenging. The challenge will be addressed by improving the characterization of its growth and decay phases and its modulation by Pacific and Atlantic decadal-multidecadal variabilities. Our understanding of the Pacific and Atlantic multidecadal variability remains rudimentary, in part, because of the focus on surface fields, shortness of observational record, and deficient representation of this variability in climate simulations. The project will advance evolution characterization, dynamical understanding, and model representations by focusing on the ocean surface-subsurface fields, Pacific-Atlantic basin links, and innovative machine-learning and statistical analysis of observations and historical climate simulations. It will advance rapid assessment of the realism of multidecadal variability in E3SM simulations by developing well-defined structural and impact metrics and their implementation in the PCMDI Metrics Package.
Methods: Unsupervised machine learning (e.g., growing neural gas and autoencoders) will be used to extract meaningful patterns with minimal data preprocessing. When combined with causal inference tools, machine learning (ML) can help quantify causal relationships and shed light on across-scale dynamics. The important evolution information (e.g., the difference in development and decay phases, lead-lag links between different patterns) will be sought by modifying the input to the ML algorithm, as in extended-EOFs, which will also be extensively deployed in analyses of ENSO and multidecadal variability. Results from both methods will be compared and assimilated to identify the robust variability features. Observational records and the E3SMv1 large ensemble will be analyzed to identify deficiencies in ENSO representation; the long simulations will be particularly valuable for analyses of multidecadal variability.
Potential Impacts: The project will advance the characterization of ENSO diversity, evolution, and hydroclimate impacts over the Americas, spurring the development of quantitative metrics for the rapid assessment of E3SM simulations. A similar strategy will provide insights into multidecadal evolution in the Pacific and Atlantic basins, including the role of ocean dynamics (from subsurface analysis) and inter-basin links, all advancing the development of quantitative metrics for evaluating E3SM simulations. The use of ML in interannual and multidecadal variability analysis is in accord with AI4ESP guidance.