Globally convergent optimization for data-dependent systems enabled through a novel data-driven branch-and-bound framework

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

Decision-making for complex engineering systems depends on the development of algorithms for data-driven optimization based on data generated either by high fidelity simulations and/or experiments. Despite the high potential of data-driven optimization, there is currently a lack of efficient and scalable methods that can provide high quality solutions for a general class of data-dependent problems. The proposed work is motivated by the increasing number of applications that can benefit from data-driven optimization and control, including chemical process synthesis, enhanced oil recovery, carbon dioxide sequestration, energy efficiency of buildings, and many more. The proposed research is focused on the integration of traditional process systems engineering with machine learning and uncertainty quantification concepts to overcome key challenges of data-dependent optimization which currently hinder their efficiency and scalability in applications with a high number of dimensions and constraints. The objectives of the proposed research are (a) the identification of efficient space and variable decomposition strategies for creating tractable optimization sub-problems, (b) the formulation of theoretically overestimating and underestimating approximating functions for data-dependent correlations by leveraging data and model uncertainty, and (c) the study of convergence rates and optimality bounds of data-driven branch-and bound optimization for a large set of challenging benchmark problems, as well as challenging case studies for oil-field operations, enhanced oil recovery and building design and efficiency. The central idea of this work is the formulation of novel under/over-estimating approximations, which will be incorporated within a novel customized branch & bound search to systematically identify optimal solutions with a tractable number of samples and improved convergence rates. Scalable data-driven optimization tools and a benchmarking library will be created and made publicly available with examples drawn from prominent fields, such as mechanical and structural design, chemical flowsheet design, oilfield control, parameter estimation, and protein folding. There is also a plan to incorporate data-science concepts into the chemical engineering education in the form of teaching modules that will be made available to the academic community at large. A Vertically Integrated Projects (VIP) program is also proposed aimed at attracting undergraduate students from science and engineering disciplines to work as members of interdisciplinary teams towards solving challenging data-driven optimization problems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.