**Abstract:**

Prediction is a core element of science and engineering. Sophisticated mathematical models

exist to make predictions in a variety of physical contexts including materials science, fluid

mechanics, and solid mechanics. Most of these models do not have known analytical solutions.

Moreover, they are generally difficult to solve numerically. In order to perform numerically

tractable computations, researchers often try to develop reduced models that account for the

essential physics while doing away with the complexity of the full model. Development of

reduced models necessarily induces model errors and the impact of these errors on scientific

and engineering predictions must be assessed and quantified. Even the most sophisticated

mathematical models contain errors and uncertainties due to the limits of human knowledge.

The field of uncertainty quantification seeks to rigorously quantify uncertainties and assess their

impact on predictions in scientific and engineering applications.

This talk will begin by providing an overview of uncertainty quantification in the context of

scientific and engineering predictions. I will then discuss recent results on the development of in-

adequacy models for chemical kinetics with applications to turbulent combustion. In particular,

a new physics-based inadequacy model is introduced that accounts for model error between a

detailed chemical kinetics model and a reduced model. Limitations and extensions of the model

are discussed. If time permits, I will describe recent attempts to develop physics-aware machine

learning algorithms that may be able to learn model terms in reduced models. Examples from

fluid turbulence and astrophysics will be presented.

Speaker: David Sondak

Affiliation: Departamento: Institute for Applied Computational Science (IACS) Harvard University.

Posted on Jan 5, 2018 in Seminario Aprendizaje de Máquinas, Seminars