Jun 10 2013

Selen Cremaschi, University of Tulsa

June 10, 2013

11:00 AM - 12:00 PM


CEB 218


810 South Clinton Street, Chicago, IL 60612

On Tactical and Strategic Decision Making under Uncertainty

Most tactical and strategic decisions in process industries, such as research & development (R&D), capital, and infrastructure investments, are made in highly uncertain environments. The explosion of the alternatives with increasing uncertainty prohibits the generation of a comprehensive search space for such problems, and limits the applicability of existing tools and methods. This talk will give an overview of our research activities on developing novel methods to optimize the operation and design of a wide range of applications under uncertainty: (1) surrogate model generation and derivative-free optimization for process synthesis and design of multi-scale processes, (2) systematic procedures for uncertainty propagation and reduction for robust design of chemical processes, (3) efficient solution algorithms for optimization under endogenous uncertainty for strategic decision making. Superstructure optimization is, in theory, a very powerful approach to address tactical problems such as process synthesis for energy applications. However, the resulting mathematical program is difficult, and in some cases impossible, to solve. Surrogate models, simpler functional representations of the underlying complex system, can be used to simplify the problem. We developed three sequential design algorithms to construct accurate surrogate models to be used in optimization problems. These algorithms scale well to problems in high dimensions, and investigate the trade-off between space-filling and adaptive nature of sampling methods. Our results reveal that an adaptive sampling approach is required in order to accurately model strong nonlinearities. Whether we are able to use the fundamental models or surrogate models to define our systems, there are many sources of uncertainty in these models, which in turn leads to uncertainties in the outputs such as plant throughput, and environmental impacts. Our group successfully demonstrated that the unique combination of a novel data clustering approach to identify the relevant experimental data with model fine-tuning and evaluation techniques reduces the prediction uncertainty from up to five orders of magnitude to the same order of magnitude. Although this approach is powerful for systems with uncertainties that are independent of decisions, optimization problems with decision-dependent, i.e., endogenous, uncertainty are commonly observed in process industry, e.g., synthesis of process networks with uncertain process yields, and biomass-to-commodity chemicals investment planning. Despite this, mostly due their challenging nature, the research community strayed away from these problems. Our analysis using simulation-based optimization (SIMOPT) approach revealed the importance of endogenous uncertainties on the overall cost and the resulting decision tree for these problems along with the significant computational cost of SIMOPT. As a new direction to solve these problems, we have recently developed a novel heuristic approach that decomposes the original multi-period multi-stage stochastic program into a series of two-stage stochastic programs, which are then solved rolling through the stages and periods. This approach is highly-parallelizable, and our preliminary results suggest that the solutions obtained are close to the true solution.


UIC Chemical Engineering

Date posted

Jun 17, 2019

Date updated

Jun 17, 2019