Digital Twins (DTs) are increasingly used in Structural Health Monitoring (SHM) to support safety assessment and life-cycle planning. While most applications focus on individual assets, greater value is obtained at the fleet level, where data from similar systems can be pooled to inform population-wide maintenance and operational strategies. Fleet-level DTs must represent the evolution of assets under uncertainty, which requires probabilistic life-cycle modelling rather than single deterministic simulations. In practice, this leads to repeated execution of high-fidelity models to characterise fleet-wide variability, which quickly becomes computationally prohibitive. The resulting cost limits the use of such models in time-critical decision-making. This work addresses this limitation by developing a surrogate modelling framework for life-cycle analysis that enables efficient probabilistic predictions at the fleet level. More specifically, the authors introduce a conditional Variational Autoencoder (cVAE) as a surrogate model for predicting distributions of Key Performance Indicators (KPIs) conditioned on management policies and operational constraints. The cVAE encodes fleet life-cycle behaviour in a low-dimensional latent space, and uncertainty in the KPI predictions is generated through sampling the latent space while remaining consistent with the imposed operating context. Training data are generated using a high-fidelity Discrete Event Simulation (DES) model, which was selected for its ability to simulate complex logistics and maintenance events. We apply the framework to the lifecycle of an aircraft fleet, assessing whether the employed generative model can effectively replace the fleet DES model in terms of computational time and accuracy. In this context, the proposed framework is examined as a means of supporting maintenance optimisation under the practical constraints of industrial DT applications.