The variability of the Earth’s weather system on timescales spanning months to several years is dominated by fundamental oscillations such as the Madden–Julian Oscillation (MJO), the El Niño Southern Oscillation (ENSO), and the Pacific Decadal Oscillation (PDO). These phenomena are oftentimes not well-represented in conventional first-principles numerical models, leading to gaps in monitoring and forecasting capabilities in the 1-to-10-year forecasting window. By leveraging dynamical systems theory and scientific machine learning, the main objective of this project is to demonstrate and transition a data-driven computational framework for analysis and prediction of natural hazards dynamics on the 1-to-10-year forecasting window that will help bridge these gaps. This project will (i) translate a previously developed academic research code to a scalable, well-documented Python toolkit with an accessible frontend to Department of Defense (DoD) users; and (ii) develop specifications of training data and predictor variables enabling high-fidelity identification and prediction of the MJO, ENSO, and Pacific Decadal Variability (PDV). The project team will demonstrate this framework through retrospective analyses and real-time forecasts of these phenomena and associated conditional statistics of oceanic and atmospheric variables that are important for operational planning and resilience of DoD installations. The culminating objective of the project is to transition this technology to DoD users working on historical record analyses, interannual prediction, and nonstationary trend detection.

The technology is based on a novel combination of ideas from Koopman operator theory for dynamical systems with machine learning methods for kernel operator approximation. This framework leads to computational algorithms that can efficiently process information from multimodal, unstructured observations over geometrically complex domains to build data-driven dynamical operators that encapsulate complex nonlinear phenomena without biases due to physics parameterizations and/or coarse mesh resolutions. Spectral decomposition of the dynamical operators yields indices (eigenfunctions) adept at capturing coherent oscillatory phenomena such as MJO, ENSO, and the PDO, and objectively delineating them from nonstationary trends. Furthermore, the dynamical operators enable probabilistic prediction models that can assimilate high-dimensional initial conditions in a computationally efficient manner. Using remote sensing, in situ observations and reanalysis data for training, the project team will demonstrate this framework by (i) creating reconstructions of Sea Surface Temperature, Surface Air Temperature, oceanic circulation, atmospheric circulation, and precipitation fields associated with the MJO, ENSO, PDV, and 1-to-10-year forecasting trends; and (ii) performing a suite of real-time prediction experiments of these phenomena over a boreal winter in the second year of the project targeting regions within the US Indo-Pacific Command and US Northern Command areas of responsibility with high concentrations of DoD installations. The success of these experiments will be judged (i) qualitatively on the basis of their ability to provide physically-meaningful, actionable information on key phenomena on the 1-to-10-year forecasting window; and (ii) quantitatively on the basis of forecast skill metrics for individual MJO and ENSO events and physical quantities of interest that depend upon them.

The technology will provide an effective toolkit in the arsenal of prediction and projection frameworks available to the DoD, with a solid focus on phenomena in the challenging 1-to-10-year forecasting window. Due to its flexible, data-driven nature, the technology is able to yield results using modest computational and human resources. In particular, the technology can be readily customized to specific use-cases by analysts working individually or in small teams. With regards to specific phenomena in the 1-to-10-year forecasting window gap, the technology is expected to improve monitoring, data assimilation, and probabilistic forecasting capabilities of the MJO, ENSO, PDV, and the 1-to-10-year forecasting window. Given the importance of improved understanding and forecasting of these phenomena to operational planning and natural hazard resilience in the 1–10 year horizon, it is expected that the technology will be widely applicable and yield high return on investment in Air Force, Army, Navy, and Joint use-cases. In terms of broader impacts, the technology will help popularize and increase the adoption of cutting-edge scientific machine learning methodologies to DoD users. (Anticipated Project Completion - 2026)