YouTube: https://www.youtube.com/watch?v=BVpiR3izxJo

Abstract:

Multi-fidelity stochastic modeling entails the use of variable fidelity methods and models both in physical and probability space. In this new paradigm, we target learning from variable information sources as even under-resolved simulations or simplified mathematical models or empirical correlations can be employed to construct an accurate stochastic response surface. The goal of this talk is to provide an introduction to multi-fidelity modeling using probabilistic machine learning and Gaussian processes. In particular, we will present a general, yet flexible data-driven framework, that allows us to simultaneously track both parametric and modeling uncertainties by synergistically combining models of variable fidelity and exploiting their cross-correlation structure. A collection of benchmark problems will demonstrate the robustness of the proposed algorithms with respect to model misspecification (e.g. inaccurate low-fidelity models or noisy measurements), as well as their accuracy and efficiency for a wide range of target applications including uncertainty quantification, data-assimilation, inverse problems, design optimization, and beyond.

Presenter biography:

Paris Perdikaris is a post-doctoral associate at the Department of Mechanical Engineering at MIT, bringing expertise on scalable statistical learning algorithms, uncertainty quantification, computational fluid dynamics and parallel computing. His current research focus includes the development of multi-fidelity information fusion algorithms for data assimilation, inverse problems, engineering design optimization under uncertainty, and beyond. This work has received support by AFOSR, DOE, and DARPA (2010-present).

 This webinar is hosted by the MSM Theoretical and Computational Methods WG

*Please note that webex may not be compatible with Windows 10