CDS&E/Collaborative Research: Local Gaussian Process Approaches for Predicting Jump Behaviors of Engineering Systems

点击翻译按钮获取中文摘要

This award will contribute to national prosperity and economic welfare by developing tools to support the efficient and effective design of modern engineering systems such as smart factories and smart autonomous systems for material handling. A major challenge in designing such systems is that their performance can change abruptly with small changes in design variables, creating discontinuous design responses. This grant will develop efficient surrogate modeling methods to predict design performance in the presence of such discontinuities which can then be exploited in design optimization. This work will facilitate the solution of complex engineering design problems and will be evaluated in the design of a smart manufacturing system for carbon nanotubes, and the design of automated material handling systems. The award will also contribute to the development of a data science-capable workforce by providing multidisciplinary research, training, and international collaboration opportunities for K-12, undergraduate, and graduate students. The research team will broadly disseminate their research findings and share data and the resulting software packages to the data science and systems engineering community. This research will make substantial contributions to the areas of surrogate modeling, sequential design, active learning, system design, and advanced manufacturing. System performance is modeled as a piece-wise continuous function of design variables, motivating local Gaussian process (GP) surrogate modeling. The approach accommodates regime changes around a prediction location, segmenting local data based on the estimated partition(s). Only the local data belonging to the same regime as a prediction location affects the model prediction. Research activities will explore two ideas: (1) local GP modeling with local data selection; and (2) smoother alternatives that augment design variables with probabilistic regime estimates. A sequential design approach to optimize data acquisition plans for training the new surrogate models will also be investigated. The resulting new meta-models and sequential design scheme will be validated using design problems in carbon nanotube synthesis and smart material handling systems.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.