Advisor: Dr. Dimitri Mavris

Milestone: PhD Thesis Final Examination (Defense)

Degree Program: Aerospace Engineering

Title: A Scalable and Efficient Framework for Graph-Based Analysis of the Digital Thread

Abstract: The increasingly connected and rapidly evolving aerospace operational environment, coupled with tight budgets and aggressive timelines, poses significant challenges to the current document-centric analytical paradigm. Addressing these challenges requires Digital Transformation and the application of Digital Engineering principles, with the Digital Thread serving as a key enabler. However, the Digital Thread’s scale and complexity can limit the efficiency of advanced analytics. Digital Thread graphs can support these analyses through graph representation learning, particularly Graph Neural Networks (GNNs). Processing entire graphs introduces scalability challenges, while partitioning them into smaller isolated subgraphs restricts information flow. Furthermore, learning representations for prediction does not, by itself, provide an effective strategy for design-space exploration. This thesis addresses these challenges through hybrid graph representations and gradient-based exploration of regression embeddings. Hybrid graphs preserve detailed local information while incorporating aggregated global context from external partitions. Embedding space exploration searches learned representations and decodes them into admissible input states. The AC Optimal Power Flow (AC-OPF) problem in electricity transmission serves as the primary proxy for Digital Thread analytics. The methods are also applied to flight delay prediction and schedule optimization to assess whether the observed trends hold in a different application domain. Across both applications, hybrid graphs recover predictive accuracy lost through isolation while reducing peak memory requirements relative to processing entire graphs. This follows the expected balance between information retention and computational efficiency. Training time benefits are more conditional: additional training batches limit savings in the electricity transmission experiments, whereas hybrid models reduce training time in the flight delay demonstration without subsampling. Embedding space exploration also reduces search time and memory compared with optimizing the original system inputs directly. The improvement achieved in the learned representation must be translated/decoded back into feasible changes to the engineering system. Decoding preserves a substantial portion of the improvement obtained through direct input optimization in the electricity transmission experiments, with a greater loss of improvement in the flight schedule application. These findings show the potential and limitations of hybrid graphs and learned representations for computationally efficient graph-based analysis and optimization of engineered systems, providing a foundation for their application to Digital Thread graphs.

Date and time: 2026-11-16, 8:30 AM

Location: Collaborative Visual Environment (CoVE); Weber Space Science and Technology Building (SSTIII)

Committee:
Dr. Dimitri Mavris (advisor), School of Aerospace Engineering
Dr. Kyriakos G. Vamvoudakis, School of Aerospace Engineering
Dr. Kai A. James, School of Aerospace Engineering
Dr. Woong Je Sung, School of Aerospace Engineering
Dr. Arun Ramamurthy, Siemens Technology (Princeton