Friday, May 13 2022
8:30am - 9:30am
PhD Thesis Presentation
Thesis Proposal Presentation Taylor Roper

Title: A Deep-Learning Enhanced Pipeline for Learning Uncertain Quantities for Data Consistent Inversion

Data consistent inversion (DCI) is a measure-theoretic framework for constructing pullback measures on model inputs (called parameters) from an observed measure on model outputs (called quantities of interest, or QoI for short). Most of the prior work on DCI focused on a priori specification of QoI for which data are collected and analyzed to define an observed distribution that quantifies the uncertainty in such QoI due to the uncertainty in parameters. A recent framework for Learning Uncertain Quantities (LUQ) encoded within an open-source Python module is designed to transform noisy temporal data into low-dimensional QoI samples whose uncertainties are predominantly due to the underlying uncertainties in parameters. The LUQ framework is a machine-learning based data-to-distribution pipeline that proceeds in three steps to (1) filter measurement noise via the formulation and solution of optimization problems, (2) label and classify dynamics via unsupervised learning, and (3) learn QoI through kernel-based feature extraction techniques. A recent work demonstrates the successful coupling of LUQ and DCI across models of various dynamical systems arising in the life and physical sciences including a model of glycolysis exhibiting Hopf bifurcations and a model of storm surge for a simulated extreme weather event near the Shinnecock Inlet located in the Outer Barrier of Long Island, NY, USA. However, the existing LUQ framework is limited in many ways, in particular with its current inability to perform the first step to non-temporal data. The proposed research focuses on the investigation and removal of these limitations primarily through the integration of a deep-learning assisted filtration step. Additional proposed research topics involve the integration of optimal experimental design techniques within the LUQ framework to determine optimal data collection strategies in space and time and the further development of the LUQ software module.
Speaker:Taylor Roper
Location:Zoom link in email

Download as iCalendar