Bridging human and artificial intelligence for skillful, trustworthy, and insightful seasonal-to-decadal climate prediction

Abstract

Seasonal-to-decadal climate variability is inherently difficult to predict and is intimately connected to human and natural systems worldwide. Skillful forecasts on two-month to ten-year timescales would enable proactive and informed decision-making for many industries, including fisheries, water management, and agriculture. Understanding the behavior of seasonal-to-decadal climate variability provides context for our changing environment. Neural networks, a class of artificial intelligence tools, are well-suited for exploring teleconnections, precursors, and patterns of variability, since they can identify complex relationships within immense quantities of data. Neural networks have traditionally been used as "black-box" models that produce predictions but are inherently difficult to explain. There has been a recent push to develop "interpretable" models that can be understood by human scientists. In this dissertation, I bridge human and artificial intelligence to leverage interpretable AI for skillful, trustworthy, and insightful prediction of seasonal-to-decadal climate variability. First, I show how interpretable neural networks can be used to optimize a simple forecasting method, analog forecasting. This approach highlights four precursor patterns for one-year forecasts of El Niño Southern Oscillation in the Tropical Pacific, West Pacific, Baja Coast region, and Tropical Atlantic. In addition, when making five-year forecasts of observed sea surface temperature variability in the North Atlantic, this optimized analog forecasting approach rivals the performance of an initialized decadal prediction system. Second, I design neural networks to learn patterns of internal variability and forced change. Using these neural networks, I perform climate change attribution for observed sea surface temperatures. Despite the unprecedented, record-high, global-mean sea surface temperature in 2023, our results suggest that much of this warming can be explained by internal variability, as anomalously cold conditions in 2021 and 2022 shifted to anomalously warm conditions in 2023. Third, I use neural networks to make decadal forecasts of the likelihood that annual-global-mean temperature exceeds 1.5˚C, a critical Paris Agreement temperature threshold. These forecasts predict that it is very likely that annual-global-mean temperature exceeds 1.5˚C in the next decade (2024-2033), serving as a harbinger for future climate change. These forecasts are consistent with dynamical initialized prediction systems, demonstrating that neural networks can provide skillful decadal forecasts at reduced computational expense. Neural networks are powerful tools for prediction, and facilitate deeper discovery of our chaotic, interconnected, predictable Earth.