CSE DSI Machine Learning Seminar with Anabel del Val (AEM, UMN)
Data-driven stochastic expansions for inverse uncertainty estimation
Uncertainty estimation is essential for ensuring the reliability of predictions. Forward uncertainty estimation relies on prior knowledge about the uncertainties affecting a given system. The prior knowledge can differ from expert to expert and suffers from potential subjectivity. Whenever experimental data is available, inverse uncertainty propagation or Bayesian inference recovers an objective measure of uncertainty by learning model parameters from experimental data. In this talk, I will discuss the use of data-driven stochastic expansions in the context of Bayesian inference for functional outputs. Empirical Karhunen-Loève expansions are capable of approximating functional outputs (measured quantities that change in space or in time) with just a few selected simulations. Further, these expansions can be generalized for any set of inputs by approximating the eigenvalues with Gaussian processes, linking physical model parameters to the expansions’ modes. I will illustrate this methodology through its use in a problem of learning material properties from experiments in plasma flows at the NASA Ames arc-jet.
Anabel del Val is an Assistant Professor in the Department of Aerospace Engineering and Mechanics at the University of Minnesota. Her research is focused on the development and application of stochastic methods to achieve robust predictive modeling of hypersonic and reacting flows. These stochastic methods belong to the broad areas of sensitivity analysis, surrogate modeling, forward uncertainty propagation, Bayesian inverse problems, optimization under uncertainty, and data-driven approaches.
Keller 3-180 or via Zoom.