Predicting Timescales in Dynamical Systems with Explainable Machine Learning Models

Presenter

Demetri Liousas

Campus

UMass Amherst

Sponsor

Steve de Bruyn Kops, Department of Mechanical and Industrial Engineering, UMass Amherst

Schedule

Session 2, 11:30 AM - 12:15 PM [Schedule by Time][Poster Grid for Time/Location]

Location

Poster Board A71, Campus Center Auditorium, Row 4 (A61-A80) [Poster Location Map]

Abstract

Fluid turbulence affects almost every aspect of human existence, from weather and climate to the cooling of our smartphones to processes within our own bodies.  Turbulence, though, is not well understood and poorly modeled. Machine learning (ML) offers the promise of taking data, either from physical measurements or from very large simulations, totaling petabytes of data, and reducing them down to models for fast-turnaround predictive simulations.  An open question with this approach is how to explain an ML model’s predictions, which is necessary for developing trustworthy models. We take a step toward answering that question by developing a framework to extract timescales governing the time evolution of a dynamical system from an ML model. We use this framework to develop physical explanations for time-series ML models.

Turbulence is complicated, and it is not the only dynamical system that might be modeled with the help of ML. We consider a simpler dynamical system, namely the damped pendulum, to demonstrate this framework for explainable ML models because we can analytically derive its two timescales. We conducted a parametric study on the data sequencing hyperparameters for time-series ML models: sequence length and sampling frequency. Our analysis reveals that by discovering the minimum information required to train such a model with respect to the data sequencing hyperparameters, we can identify the physical timescales of the underlying dynamical system. By doing so, we also develop physics-based explanations for those hyperparameters. These are important steps towards creating trustworthy ML models for predicting dynamical systems such as fluid turbulence.

Keywords

machine learning, timescales, explainability, dynamical systems, pendulum

Research Area

Engineering

SIMILAR ABSTRACTS (BY KEYWORD)