Optimal Experimental Design for Digital Twins Using a Network-Flow Approach
Abstract: Efficient data assimilation is a significant bottleneck toward the development of highly scalable and reliable digital twins. Uncertainty quantification and optimization becomes particularly challenging when either the nature of the system or cost limitations prevent the deployment of a large number of sensors for data collection.
We propose a novel trajectory design approach for the optimal experimental design of contiguous data acquisition based on a network-flow formulation. In contrast to conventional control strategies, we discretize the spatial domain into a set of network nodes and identify the optimal path between nodes to minimize uncertainty. This approach allows for a more thorough exploration of the design space than what multi-start local optimizations can do.
The network-flow approach we propose translates into a challenging discrete optimization problem, whose objective is expensive to evaluate, highly nonlinear, and can be non-smooth. Moreover, the objective function needs to be evaluated over the entire path instead of being summed edge-by-edge. With an increasing number of edges, the number of possible paths rapidly increases and the combinatorial problem becomes intractable. We showcase a branch-and-bound algorithm we developed, which leverages tailored bounds and constraints to tackle the optimization at a tractable computational cost.
See all upcoming talks at: https://www.anl.gov/mcs/lans-seminars.