LANS Informal Seminar: Zheng Zhang

Seminar Title: Uncertainty quantification techniques for data-intensive engineering design
Speaker: Zheng Zhang, Postdoctoral Appointee, MCS, Argonne National Laboratory

Date/Time: 2016-01-20 15:00
Location: Building 240 rm 4301

Description:
The design and analysis of many real-world engineering problems are subject to unavoidable uncertainties. Representative examples include (but are not limited to) nano-electronic devices and circuits that are sensitive to process variations, energy systems integrated with weather-dependent renewable sources, and biomedical computational tomography that infers tissue properties from noisy signals. These seemingly irrelevant engineering problems share a common data-intensive feature — a huge number of stochastic or parameterized simulation samples are required to support reliability/yield analysis, stochastic optimization and control, predictive modeling and data inference, and so forth. As an efficient non-Monte-Carlo uncertainty quantification tool, stochastic spectral methods can reduce significantly the cost of stochastic simulation and provide computationally efficient surrogate models to accelerate subsequent computation and design tasks.

In this talk, I will present some recent advancement of stochastic spectral methods with diverse engineering applications. In the first part, I will present an efficient method to simulate nonlinear dynamic systems influenced with uncertain parameters. Basic formulations for general cases and advanced algorithms for stochastic oscillatory systems will both be covered. A major concern of stochastic spectral methods is the “curse of dimensionality” — the computational cost can grow very fast as the number of random parameters increases. Thus, in the second part of this talk I will show how to efficiently handle high-dimensional uncertainty quantification problems from a tensor (which is a multidimensional generalization of matrix) and hierarchical perspective. On the various integrated circuit (IC), MEMS and power system benchmarks we have tested, our algorithms significantly outperform standard stochastic spectral methods in terms of efficiency, and their speedup factor over Monte Carlo can be up to 1000 times. For high-dimensional problems, the tensor approach can reduce the exponential complexity of some algorithms to a linear one.

I will also talk about some future topics in this field. Specifically, I will cover the following issues: theoretical challenges of uncertainty quantification, interface of uncertainty quantification with other research topics such as machine learning and high-dim statistics, as well as some promising engineering applications.