Abstract:
Many objective optimization problems (MaOPs), defined here as problems with four or more objectives, inherently arise when designing and operating sustainable process and energy systems. Nonetheless, MaOPs remain a particularly difficult and understudied problem class owing to challenges in both scalably generating solutions (in the form of a Pareto frontier) and interpreting the solutions we do get. This talk begins by providing a tutorial overview of methods commonly used to generate Pareto frontiers for multi-objective (defined as two or more objectives) optimization problems, alongside a discussion of why these approaches tend not to scale well when increasing the number of objectives. Then, I present the Objective Reduction Community Algorithm (ORCA), our group’s novel approach for reducing the objective dimensionality of MaOP’s to a manageable two or three a priori to solving the problem. At a high level, this approach identifies which objectives are likely to be pointing to similar decisions, embeds this information into a graph, and uses community detection to partition this graph to give groupings of objectives which are correlated within groups, but competing between groups. Three case studies are presented which demonstrate the power and efficacy of this approach for sustainable decision making. In the first, the correlation of planetary boundary objectives within a large scale sustainable fuel supply chain is assessed. The second case study considers changes in objective correlation due to intermittency in objective parameters, and assesses the correlation between cost and emissions driven industrial demand response for two different electrified chemical processes. The third case study utilizes an extension of ORCA to nonlinear MaOP’s, showcasing its utility on a commonly used set of benchmark problems with known objective correlations, the DZLT5 problems, as well as on a design problem for a carbon capture, utilization, and storage network. I conclude by providing some perspective on future applications of our approach to problems of distributed model predictive control and aligned agentic AI systems.

Biography:
Andrew Allman is currently an assistant professor in the Department of Chemical Engineering at the University of Michigan, and has been in this position since fall 2020. Andrew is an alumnus of the University of Minnesota Department of Chemical Engineering and Materials Science, where he obtained his Ph.D. with Prodromos Daoutidis in 2018 and subsequently worked in a post-doctoral position with Qi Zhang from graduation until joining Michigan. His awards include receiving the NSF CAREER Award in 2023, and the CAST Director’s Student Presentation Award in 2018. His recent research at Michigan has focused on solving many-objective optimization problems, assessing the operation and control of modular chemical production systems, embedding machine-learned classifier models to enhance moving horizon decision making, and solving structure detection problems for networks with constraints or uncertainty for distributed optimization.

For the most up to date information on the location, please check the related link the week of this event