The rapid sizing of mechanical structures is a major challenge in industrial design, especially in aeronautical applications where many configurations must be assessed under strong time constraints. Although high-fidelity numerical simulations provide accurate predictions, their computational cost remains a limitation for fast pre-design iterations. Reduced Order Modelling (ROM) techniques, such as Proper Orthogonal Decomposition (POD) and Proper Generalized Decomposition (PGD), are efficient tools to accelerate the solution of partial differential equations by projecting them onto low-dimensional spaces [1,2]. However, classical ROM approaches usually require fixed or parametrized geometries, often sharing a common discretization. This strongly limits their use for industrial problems involving unstructured meshes, non-parametrized geometries, and significant topological variability. This thesis introduces a hybrid artificial intelligence and ROM framework designed to overcome these limitations. The proposed approach relies on Graph Neural Networks (GNNs) to predict reduced bases directly from unstructured finite element meshes. Graphs naturally encode both geometry and connectivity, making them suitable for transferring information across highly variable mechanical configurations [2,3]. The governing equations are then solved by projection onto the predicted reduced basis, ensuring that the solution remains constrained by the underlying physical equations. To improve robustness, the GNN-generated basis is completed when necessary through an adaptive enrichment strategy inspired by PGD. This boosted PGD enrichment allows the method to correct the initial neural prediction and preserve accuracy for unseen configurations [3]. Complementary neural-network-based developments are also proposed to accelerate nonlinear solvers already used in industrial simulation environments. The complete framework is validated on two original datasets developed and released during this thesis, inspired by Safran industrial use cases and characterized by strong geometric and topological variability. Numerical results show computation time reductions of up to 70% without significant loss of accuracy. This work therefore provides a practical strategy for integrating artificial intelligence into industrial simulation pipelines while preserving the physical consistency required for reliable mechanical analysis.