Infrastructure systems typically require continuous operation with minimal downtime, while subjected to high variability of environmental and operational loads and progres- sive damage and deterioration processes. Consequently, static models trained under fixed operational regimes often fail to generalize to changing conditions, thereby limiting their effectiveness for condition-based maintenance and decision-making in evolving operational contexts. This work attempts to address these challenges by leveraging PHLieNet, a re- cently proposed framework employing Parametric Hypernetworks for Learning Interpo- lated Networks. PHLieNet relies on inferred embeddings that are tasked to modulate a hypernetwork responsible for generating the weights of dynamics propagation models. Via linear combinations of the inferred latent embeddings, the network can smoothly in- terpolate in the weight space and represent a continuous family of dynamics. The main contributions of this work are twofold: First, we attempt to extend the original framework to high-dimensional systems by integrating autoencoders that identify computationally tractable spaces for subsequent model generation. Second, a data-efficient adaptation strategy is explored to extend the validity of the trained models to new parameter regimes representing changing operational conditions. By incorporating additional learnable em- beddings while keeping the core network fixed, the model can capture previously unseen dynamics corresponding to new operational conditions. This approach requires optimizing only a small subset of parameters, making it significantly more efficient than full model retraining. We demonstrate the effectiveness of this adaptive procedure on stochastically excited structural dynamic systems undergoing damage. The extended models maintain accurate and stable long-term predictions, enabling optimal maintenance planning and minimizing operational disruption.