EAGER: Collaborative Research: Real-time Heterogeneous Transfer Active Learning to Bridge Knowledge Gaps in System Integration under Environmental Uncertainty

In-space production application is a national initiative to ensure US leadership of in-space manufacturing in low Earth orbit by demonstrating the production of advanced materials and products for the terrestrial market. However, the zero-gravity environment impairs the product quality of many manufacturing systems that perform efficiently and reliably on-ground. Moreover, the data collection cost of in-space manufacturing is so expensive that commonly used quality control and uncertainty quantification strategies fail for such data-scarce systems. This EArly-concept Grant for Exploratory Research (EAGER) project addresses these fundamental issues by establishing a real-time heterogeneous transfer active learning framework. This framework leverages knowledge from well-studied, data-rich on-ground manufacturing systems to enhance experiment design, uncertainty quantification, and quality control in in-space manufacturing systems. Specifically, this project focuses on the in-space electrohydrodynamic inkjet printing and collaborates with the National Aeronautics and Space Administration to collect both on-ground and parabolic flight test data, develop transfer learning models, and validate their performance. The project also contributes to workforce training by promoting the interdisciplinary research of manufacturing, sensing, and data analytics and integrating the research as project topics into undergraduate/graduate courses and various outreach activities. This project leverages the state-of-the-art transfer learning strategy to resolve the urgent need for reliable in-space manufacturing. While transfer learning is effective for dealing with data scarcity, it faces unique challenges when adapted for integrating on-ground and in-space manufacturing systems: 1) The input values, dimensions, or even data types between the on-ground and in-space systems are different. Such heterogeneity requires not only the adaptation of inputs among systems, but also the identification of useful source systems. 2) The in-situ computational resource is limited. This limitation hampers most active learning methods, where the estimated or predicted uncertainty from training data must be recalculated from scratch whenever new experimental data (identified by active learning) is added. This project facilitates a real-time heterogeneous transfer active learning to conduct the heterogeneous transfer learning batch-by-batch within the context of active learning. This project features 1) a flexible and interpretable transfer learning framework to deal with heterogeneous inputs; 2) a Bayesian mechanism to update experiment design and predictions in real-time; and 3) a tailored experiment validation plan for on-ground and in-space manufacturing systems. The successful implementation of the project fills in the knowledge gaps and challenges when translating a manufacturing system into a different environment where there are unforeseen uncertainties. 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.