This dissertation develops a generalizable architecture for real-time data-driven decision support in remote edge deployments where model fidelity depends on computationally intensive, high-latency simulation. Cyber-physical applications increasingly require low-latency spatial inference while continuously adapting models to evolving environmental conditions. High-fidelity physics simulations and retraining workflows, however, are computationally expensive and typically executed on remote high-performance computing (HPC) systems under batch scheduling, creating a fundamental mismatch between the responsiveness required at the edge and the cost and availability of simulation-driven model updates.
We ground this architecture in a concrete application example: a real-world digital agriculture deployment. We first develop and validate a computational fluid dynamics (CFD) model for predicting spatial airflow within a Citrus Under Protective Screens (CUPS) deployment --- a novel agricultural structure that modifies the microclimate within a field of citrus trees and requires modified farm management strategies as a result. The model and its validation methodology form the science driver for all subsequent system work.
We then present a closed-loop hybrid modeling system that couples edge-based surrogate inference with continuous simulation-driven model updating across dedicated cluster and shared HPC resources. The system deploys lightweight surrogate models at the edge while incorporating improved models asynchronously as they become available from HPC. We instantiate this architecture in the CUPS deployment over a private 5G wireless fabric that connects field sensors, edge nodes, and remote computing facilities across the continuum. Results demonstrate edge inference in about a second or less and a 2.7x improvement in model publication cadence during an active allocation, when opportunistic HPC execution augments the dedicated pipeline.
Finally, we characterize model accuracy decay as a function of staleness and introduce decision-time heuristics for opportunistic HPC execution. Each heuristic estimates the percentage improvement a queued job would deliver over the currently deployed model from observations already available when the job reaches the head of the batch queue, and gates release primarily on the agreement among three independently trained surrogates rather than on the confidence of any one of them. The No-Regret heuristic eliminates all degrading releases across the 102 decisions of the study while avoiding roughly 89% of the HPC allocation that unconditional release would consume---some 560,000 core-hours---and the complementary High-Yield heuristic captures more of the improvement that was available than No-Regret does, 26% of what perfect foresight would collect, while more than doubling the number of opportunistic model updates delivered to the edge and still avoiding 72%, demonstrating that selective prediction-guided opportunistic execution substantially outperforms unconditional release.
Although grounded in digital agriculture, the architecture, methodology, and analytical framework are domain-agnostic. Together, these contributions form a reusable foundation for real-time data-driven decision support across cyber-physical systems where edge responsiveness must be reconciled with the latency and cost of high-fidelity simulation.