Predicting Critical Transitions in Multiscale Data
PI: Rogene Eichler West
Objective
- Systematically evaluate factors impacting the performance of reservoir computing (RC) regarding the prediction of critical transitions in fast-slow systems.
- Develop methods to separate out metastable (slow) from driving (fast) components of the input signal.
- Develop methods to approximate the conditional distributions over futures using Wiener projections of the Koopman operator (Mori-Zwanzig formalism).
- Investigate the mathematical foundations of transitions via sign persistence.
Overview

Predicting the dynamics of complex nonlinear systems remains a challenging problem both in dynamical systems theory as well as real-world science and engineering applications. Data-driven methods utilizing the latest advances in machine learning (ML) provide a promising new paradigm for this task. Our approach is based on reservoir computing (RC), which is capable of skillfully predicting chaotic dynamics in multiscale systems. The overarching goal of this project is to explore the capability of a proposed framework built using RC for anticipating critical transitions in systems, assuming the underlying slow dynamics are known, for predicting transitions solely from time-series data.
To achieve our objective, we must gauge the sensitivity of the technique to changes in forcing terms, its robustness to local modifications in a slower model, and its applicability in higher dimensions. Determining the limits of its predictive capabilities will better inform the techniques this project will develop for inferring governing equations and help improve the approach’s effectiveness.
Impact
- Current approaches require knowledge of underlying equations; they hand-tune the RC architecture to the dynamics of the driving signal.
- The impact of a successful solution is the generalization of a data-driven approach to the prediction of critical transitions in problem domains where the underlying physics is unknown.
- The transition of this approach to neuromorphic technologies open new capabilities in sensor intelligence and anomaly detection.