Discovering and forecasting extreme events via active learning in neural operators

Area of Science:

• Artificial Intelligence
• Data Science
• Complex Systems Analysis

Background:

• Extreme events in nature and society pose significant risks but are challenging to characterize due to their rarity and the complexity of underlying systems.
• Traditional methods struggle with the inherent difficulties in identifying and quantifying rare events from limited or incomplete data.

Purpose of the Study:

• To develop a novel, scalable artificial intelligence (AI)-assisted framework for the efficient characterization and inference of extreme events.
• To overcome the limitations of existing methods in analyzing rare phenomena within complex, potentially infinite-dimensional systems.

Main Methods:

• Utilized a model-agnostic framework combining output-weighted training in Bayesian experimental design (BED) with an ensemble of deep neural operators.
• Employed a BED scheme for active data selection to quantify extreme events and deep neural operators to approximate nonlinear operators.

Main Results:

• The proposed AI framework demonstrated superior performance compared to Gaussian processes.
• Key findings include optimal performance with shallow ensembles, effective extreme event detection irrespective of initial data, elimination of 'double-descent' phenomena, and robust performance with suboptimal sampling.
• Monte Carlo acquisition proved more effective than standard optimizers in high-dimensional settings.

Conclusions:

• The developed AI infrastructure offers a scalable solution for efficiently inferring and pinpointing critical situations in various domains.
• This approach enhances the ability to understand and predict catastrophic events by leveraging advanced AI techniques.