Publication

This is the final technical report of the Systems Engineering Research Center (SERC) research task WRT-1025. This research investigated digital twin design architectures that support Artificial Intelligence (AI) and Machine Learning (ML) formalisms working side-by-side as a team. The unique aspects of this research are the use semantic technologies that can leverage AI and ML providing complementary and supportive roles in the collection, formalizing representations and processing of data, identification and correlation of events, in evolving spatial contexts and automated decision making throughout the system lifecycle. The research developed graph embedding procedures with ML tasks, which together can enhance digital twin design and decision making to factor in evolving temporal and spatial information, such as those encountered in urban settings.

This research builds upon our previous work on teaching machines to understand urban networks with graph analytics techniques. These advances are due in part, to advances in computer, communications and sensing technologies that evolved over the past three decades in large-scale urban systems and are now far more heterogeneous and automated than their predecessors. They may, in fact, be connected to other types of systems in completely new ways. These characteristics create challenges now that we have opportunities to better integrate mission engineering, system design, analysis and integration of multi-disciplinary concerns. We have made progress against these challenges by teaching machines to understand graphs that represent urban networks. This report discusses research efforts and accomplishments for using a recently developed graph autoencoding approach to encode the structure and associated network attributes as low-dimensional vectors. We successfully demonstrated the approach on a problem involving identification of leaks in urban water distribution systems.

We have made unique progress and have been able to share work in SERC events such as the workshop on AI4SE and SE4AI. However, there is still more research needed. We would propose to investigate ways for automating cluster identification, and test the scalability of our approach (i.e., test larger graphs). In addition, several objectives from previous reports are still pending such as exploring composition (i.e., learning graph topology in parts), decoder architectures, and linking models of graph topology to models of system behavior and identification of events, among others. Therefore, needed investigations should consider the following set of basic questions, such as does graph composition help alleviate the challenges posed by larger graph sizes and complexity?