Modeling complex engineering systems often requires balancing fidelity against computational cost, as the interacting physical processes that govern system behavior cannot be resolved at high resolution continuously. We present a probabilistic graphical model architecture for digital twins, in which model resolution becomes a dynamic decision variable rather than a fixed design choice. Baseline inference and data assimilation are carried out using a low-resolution representation, while the high-resolution model is introduced only intermittently to inject additional physical detail when uncertainty exceeds specified tolerances or higher predictive fidelity is needed. The framework explicitly manages information transfer across resolutions, ensuring that state estimates remain continuous and that high-resolution corrections persist after reverting to low-resolution dynamics. By enabling sparse but targeted use of expensive models, the approach maintains consistency and predictive capability without sustained high-resolution computation. Proof-of-concept studies demonstrate the framework across different multi-resolution strategies, where the single physical asset is represented at varying levels of mechanistic detail.