On the Development of a Deep Learning-Based Surrogate Model for Fleet-Wide Probabilistic Modeling

High-fidelity numerical models are widely used to study the behavior of complex structures in Structural Health Monitoring (SHM), but their high computational cost limits their use in stochastic settings such as fleet-level applications. In practice, fleets of engineering assets show natural variability due to differences in loading, materials, and manufacturing, making them inherently stochastic. To address these challenges, this work develops a probabilistic surrogate model based on conditional variational autoencoders (CVAEs). The CVAE is trained to reconstruct the high-dimensional boundary response field of a critical structural region while explicitly conditioning on operational and structural parameters. By learning a latent probabilistic representation, the model explains the behavior of all individual members of a homogeneous population. Synthetic training and testing data are generated using a finite element model together with an aerodynamic panel model of a UAV. Results show that the CVAE can efficiently reproduce the spatial and stochastic features of the system response, providing accurate approximations at a fraction of the computational cost of high-fidelity simulations.