VBFF Generative active learning along unstable dynamics for modeling unprecedented extreme events

Abstract

This project aims to advance fundamental knowledge on the algorithms and theory for the efficient computation of extreme event probabilities, precursors and design of mitigation strategies in complex dynamical systems. For a plethora of systems relevant to the DoD we have stochastic dynamics characterized by strongly transient features that rise without obvious warning. These transient extremes, although rare, can be catastrophic and therefore it is essential to forecast them, mitigate them if possible, and quantify the probability of their occurrence in the long term. Typical examples include risk quantification for extreme events in human-made systems (e.g. aeroelastic instabilities), extreme climate events such heat waves and cascade effects to interconnected systems, but also mission-critical algorithmicproblems such as search and path planning operations for extreme anomalies. Existing mathematical methods have fundamental barriers on their applicability for the analysis of extreme events, due to their strongly transient, multi-scale and nonlinear nature. We propose a transformative framework for the prediction, characterization and mitigationof extreme rare events even when those are not contained in the training data. The framework consists of three key components: a) generative modeling along unstable directions of the dynamics, b) learning algorithms that encode awareness for extreme events through a few observables, c) dynamical selection criteria for data and models that promote healthier parametric or non-parametric learning for extreme events. Specifically, we plan to formulate and apply ideas rooted in generative algorithms in the context of extreme events by constraining the generation of scenarios consistently with the instabilities of the dynamical system. The motivation for this direction is the aim to generatenew, potentially unseen, extreme events, which will provide unprecedented capabilities for prediction, statistical modeling and quantification. The extreme event generators will be complemented with an extreme-event-aware learning method that will result in not only dynamics-consistent scenarios of extreme events but also probabilistically feasible, i.e. nonexotic. The new generative learning approach will be combined with a new class of dataand model- selection criteria for extreme events that will dynamically optimize the training criteria, so that errors associated with rare events are automatically detected and penalized accordingly. The proposed framework is motivated by real-world problems in fluids and mechanics, modeled by continuous and discrete systems, on which we plan to validate and assess ourmethods. We expect that the proposed effort will provide educational opportunities for students and postdocs who can later be part of the defense workforce, but also lead to new areas for future research aligned with the DOD#scurrent and projected future challenges.Approved for Public Release.

Document Details

Document Type
DoD Grant Award
Publication Date
Dec 14, 2024
Source ID
N000142512059

Entities

People

  • Themistoklis Sapsis

Organizations

  • Massachusetts Institute of Technology
  • Office of Naval Research
  • United States Navy

Tags

Readers

  • Distributed Systems and Data Platform Development
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Bayesian Inference
  • AI & ML - DoD AI Strategy