CAREER: LEarning to Search with Structure (LESS), a Unifying Algorithmic Framework for Gray Box Optimization of Biomanufacturing Systems

点击翻译按钮获取中文摘要

This Faculty Early Career Development Program (CAREER) grant will contribute to the advancement of national prosperity and economic welfare by studying efficient operations of single use scalable individualized manufacturing systems. Developing and manufacturing a new drug now often faces very tight deadlines, and a myriad of individual variants may be required to be produced. In such scenarios traditional large batch-production is generally poorly suited because of low flexibility in the quantity and type of drug being manufactured, large set up times between production runs and inability to distribute manufacturing capacity across locations and product types. This award supports better understanding of single use manufacturing as the fundamental enabler for renewed production flexibility in terms of both type variety and volume. This research will serve the biopharmaceutical manufacturing environment as well as the manufacturing community at large, promoting a new way to scale out instead of scaling up manufacturing systems. The accompanying educational plan aims to broaden STEM interest in simulation and, particularly, simulation based optimization aiming at the development of new tools for teaching and research, with a particular focus on creating a diverse research and educational ecosystem.This research will focus on the advancement of simulation based optimization methods to support decision making for the operation of individualized manufacturing systems. The project will result in new methods for the acceleration of black box optimization. The framework will consider the specific challenges of operating a large number of manufacturing processes at small scale, allowing to use the process simulation not only as a means to evaluate the performance, but also to provide structural properties of the process being operated. This research fills an important gap in the black box optimization literature, which reportedly suffers from poor finite time performance, only exploit output of the simulation model and ignores sample path information, and, finally, faces important hurdles in scaling to solve high dimensional problems. The analytical infrastructure leverages and extends state-of-the-art techniques from Bayesian optimization, high dimensional statistics, and model predictive control. The project will devise methods to efficiently achieve satisfactory solutions to simulation based optimization in presence of discontinuities resulting from the dynamics of the system. Finite time performance of the algorithms will be studied, and high dimensional problems will be central to the development of the techniques. The performance of the techniques will be evaluated not only using synthetic state of the art black box optimization problems, but using large scale production of a large variety of bio-products in an individualized manufacturing set-up. With the idea to promote this new idea of manufacturing and the concepts of simulation-supported decision making, a game will be implemented to attract students, as well as, potentially, practitioners, to the area of manufacturing and operations research.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.